People are reverse engineering Adobe products, decompiling GTA V to run it on different devices… AI is “coming for everything and everyone.”

So the question you should be asking is:

That's today's deep dive. But first, a word from our sponsor + the week's top news.

In Partnership with

Amplitude Custom Agents

When I ran growth teams at thredUP, we were obsessed at checking our funnel when something broke. What if you had a custom agent monitor your funnels for you 24/7?

That’s been the promise of AI, but implementing it has been hard. Amplitude’s Custom Agents change that. You point your custom agent at a job, like checkout drop-off, and connect your stack. Agent Connectors then pull in all the relevant context - PRDs, tickets, and conversations about the funnel. When it finds something, it creates a Linear ticket all on its own.

You can run your Custom Agent on a schedule, or whenever you want. You can grab a template, or build your own. I’d start with the checkout drop-off investigator already in the agent library.

Check them out. They’re free in open beta for all customers →

Try Custom Agents

This Week's AI News

The Week's Top News: Meta released its first frontier model

Meta has burst back into the AI race with a frontier model, Muse Wonder. Put together with its new Muse agent, Meta has burst back into the AI race in a real way.

The Artificial Analysis Intelligence Index is striking. Anthropic owns all the top models, Meta’s Muse Wonder Full is next, then OpenAI’s GPT 6 Astra, then Gemini’s 4 Argon, then Meta’s Muse Wonder:

Meta isn’t even selling Muse Wonder Full tokens. Those are only available inside Muse Code. Meta is only selling the smaller model, Muse Wonder, priced at $4 in and $20/M out. So it is priced like Opus-5.5. But it doesn’t deliver that performance, as you can see from the chart above: 54 vs 52.

There’s a reason Meta doesn’t care about selling those tokens. Meta has something the others can't buy: distribution. WhatsApp alone puts Muse one tap away from its 3 billion monthly users.

And Meta isn’t begging big. We learned this week it is up to $200B. Why? As Mark Zuckerberg and Alexandr Wang explained on two good podcasts this week, Meta is making a major play for consumer AI.

If I were OpenAI, I would be worried.

The Other News That Mattered

• Anthropic has become so fast at writing code that Github had become its biggest development bottleneck. So they pulled it out and replaced it with their own system, Scion, built in 5 months by 7 engineers.

• SpaceX is now bundling Grok, Cursor and X Premium into one subscription as low as $8 a month. We’ll see if the bundle helps them gain traction in a crowded race. Grok Bot continues to position them well.

• An OpenAI employee hid a recording device inside toilet paper behind a TV in an exec boardroom. OpenAI security found it, left it there, waited for perpetrator to come check on it, and then caught them. The FBI arrested them.

Feature/Model Releases

• Meta also released the best open weight American model yet, Muse Spark Open. It’s an 816B model that you can run locally on two M5 Ultra Mac Studios.

• Gemini 4 Argon is now open to every developer at $2/$10 per million tokens. It’s an insane deal for the level of intelligence. The catch is the price doubles once the intro period ends. In my own cost test the same job cost $0.28 on Argon and $0.91 on Opus 5.5.

• SpaceX's Grok Voice 3.0 now answers support calls for 8 cents a minute, about $4.80 an hour. Its score on simulated support calls went from 56.5% to 71.6% at the same price.

• Fish Audio Drama 3 is an open-weight voice model you direct in plain English. It can even fix a single word.

• Microsoft-Decision-1 answers in 85ms. That's the fifth big-company decision model in a month.

• Claude Code desktop got rebuilt for running many sessions at once, with a sidebar sorted by what needs you. Projects is now open to every Pro and Max user on the waitlist.

• With ChatGPT Slides, you and your team can edit the same deck live, with shared comments. It exports to PowerPoint, PDF or Google Slides and is rolling out to Pro, Business and Enterprise.

• DeepSeek V4.1 Flash is an open-weight model with an MIT license and a 1M-token context window.

Resources

• An eng lead on Claude wrote a tutorial on being more ambitious with models, and it's worth your read.

Fundraising

• TypeSafe AI raised $870M at a $7.5B valuation, led by Andreessen Horowitz. Its model, Jev, returns decisions and probabilities instead of text.

• Modal raised $750M at $15.75B after crossing a $1B run rate. Accel led the round.

• Character.AI filed for Chapter 11 and agreed to sell Disney a tech license for $20M.

Deep Dive

So can you really reverse engineer anything now?

This GitHub repo, Reverse Engineer Anything, is going absolutely bonkers.

It's sitting at 310K stars right now. It's blowing up.

So I wanted to look into it and understand where we are today in October 2026.

Today’s Post:

1. The state of reverse engineering software today

2. Which moats still work

3. What you should do now

4. My take

Subscribe now

1. The state of reverse engineering software today

We all know AI can code so fast now.

And with repos like REA, which gained ~208K GitHub stars in a single week and is now at 310K, it feels like a lot is possible.

REA is free and open source. You plug it into a coding agent like Codex, point it at an app you've downloaded and ask questions in plain English.

I know you don't have time to test it out, so I did.

I ran it on two tasks, because there are two things we need to know: reading and rebuilding.

Test 1: Reading an app

I pointed REA at Signal's desktop app and asked what a competitor would want to know.

I picked Signal because it's open source, so we can check every answer against the real code.

Here were the results. This is the pin question in my terminal:

The results were pretty wild.

REA pretty much nailed it. It got all 5 questions right, taking 3 to 7 minutes each, and we could check every answer against the open source code.

One run hit an integrity error and timed out. But Codex finished the job by reading the app's files directly, and the answer still matched Signal's real code.

Test 2: Rebuilding Linear

Now the rebuild test. I had Codex rebuild the core of Linear, the issue tracker most software teams live in. Here's my deep dive on how Linear grows.

Linear is an interesting choice because its terms ban reverse engineering, and they've been a sponsor of mine before.

So I did a clean-room build, which means I made a copy without ever seeing the original's code.

I started by writing a one-page spec, using only Linear's public docs and what I already knew about Linear. If you want to write specs like this, here's my AI PRD guide.

Then I handed it to Codex:

Codex never saw Linear's code or app. After 3 minutes and 28 seconds, it had a working issue tracker running, with statuses, priorities and Cmd+K. It all worked:

Codex called the app Fieldwork. Here is the board it built, and the create-issue modal that opens when you press C:

Overall, the rebuild test was great too.

There was one small flaw. Its first self-check failed, but it fixed the check and reran it itself. Then the sandbox blocked a browser, so I had to render the clone myself.

In the end I had a pretty functioning board for this part of Linear. It didn't replicate the whole product, but it did the part I needed.

It's valued at roughly $2.5 billion and powers 40K+ organizations. Back in February, someone cloned its core in "a couple of evenings" with ~20 prompts.

So the big question is what this means for software and software valuations.

Look at Adobe. Its stock is down about 26% over the last 12 months, but its ARR still grew 11% to $27.5B.

And an open-source Photoshop rebuild, PhotoCraft, went viral on X.

2. Which moats still work

If you can reverse engineer anything, how do you even create a moat in software?

How do we as builders, product people and founders think about this?

Let's go back to the classic 7 Powers framework from Hamilton Helmer, which I've written about in the past.

That's the classic list. But which of these hold up when an agent can copy your software?

I boiled it down to the 7 moats that come up most for software, and checked each one against my two tests.

Features broke this year. Codex rebuilt Linear's core screens in 3 minutes 28 seconds.

Habit is wearing down too, because an agent doesn't care where your buttons are. As a16z put it, "agents may kill muscle memory as a moat."

Now look at what my clone doesn't have. Four moats held up, and my Linear clone has none of them:

• Take workflow first. When a PR opens on GitHub, Linear moves the linked issue to In Progress, and to Done when it merges, and your team's whole history of issues lives there too. My clone has neither.

• Then there's data. My clone just starts with 16 fake issues, while Coinbase shows what real data is worth: it trained a small free model on its own fraud data for under $100, and on its own fraud test it beat a top model from a big lab.

• Network: your whole team is already in Linear. Nobody is in my clone, and whoever switches first ends up with a tracker their teammates can't see.

• My clone plugs into nothing. Linear plugs into GitHub, Slack, Figma and Sentry from its integrations directory, and that's distribution.

The fifth moat is the one I'd watch closest: how fast you learn from your own users. The 7 Powers book calls this process power.

A copy can add the same settings and switches. Signal, for example, can change 80 settings from its server without shipping an update.

But a copy can't get real users to test new values on, or the results you collect every time you do.

I think most builders underuse this moat. Network effects and distribution are bets you make as a company. How you learn from users is something you control every week.

So how do you use this on your own roadmap?

3. What you should do now

I think there are two things you need to do.

Step 1: Decide how much each feature deserves

AI has made it much cheaper to add a feature. So the bigger decision now is how much of your best time you give it.

Start with the customer: do they need it? If it's a basic expectation, you build it either way.

Then score the feature against the 5 moats you can actually influence: data, workflow, network, distribution and learning. Give it 0 to 2 points for how much it strengthens each one, up to 10:

Then see how easy it is to copy, like I did with Linear. Write a 1-page spec from your own public docs, give it to Codex and see how far it gets in an afternoon.

The score tells you what you're building up over time. The copy test tells you how long someone else would need to match the feature:

Here's how I'd read the 4 boxes:

• Table stakes: most features end up here. Customers expect them, so build them cheap and stop calling them differentiators.

• Ship fast when the screen is easy to copy but the data or users it brings in aren't. Get it out first.

• Double down: put your best hours here.

• Rethink: if it doesn't strengthen a moat, you could spend months on it and the next model may still copy it.

So where does Linear land? I gave the board view 1 point, for learning. My clone matched it in under 4 min, so I'd call it table stakes.

I gave the GitHub sync 6: issues move on their own as code ships, so 2 for workflow. It records what shipped against what was planned, so 2 for data. It sits in the tool engineers already use, so 2 for distribution.

A competitor could write the integration code in a day. They'd still be missing the history your team has built up in it. So I'd put it in ship fast.

Step 2: Add a moat line to every feature spec

Add 1 line next to the problem and success metric in your spec, one-pager or ticket: which moat does this feature strengthen? Data, workflow, network, distribution or learning. If the answer is none, it's a parity feature, so build it cheap.

If the answer is data, my loops guide shows how to feed what users do back into the product.

If it's distribution, my agent distribution guide shows how to make your product usable from AI agents like Claude and ChatGPT.

4. My take

So is SaaS dying? No, and you can see it in the earnings above. SaaS that sells a feature list is dying, while SaaS that owns its customers' data, workflow or network is doing fine.

I think most products are full of table stakes that their makers call differentiators. That was fine when copying took a quarter. It's a problem now that a first pass takes minutes.

If I were building anything this quarter, I'd do 3 things:

1. Run the rubric on everything you're building or about to build. If it's table stakes, give it to 1 person working with AI instead of a full squad.

2. Spend your best hours on the highest-scoring feature. Run the copy test before every launch.

3. Pick 1 number you can change without shipping an update, like a limit, trial length or price. Tune it every week and watch it like any other funnel. Salesforce priced its AI agents 3 ways in 18 months, and my pricing guide covers the 6 pricing patterns across 50 top AI startups.

That's also how you answer the question founders ask these days:

If your roadmap scores high, it gets better every time a new model drops. If it's full of zeros, most of the gains go to whoever's copying you.

See you next week,

Aakash

Welcome to another AI by Aakash. I follow the AI news, and test all the latest AI tools, so you don’t have to.

Personal agents are THE trend of 2026. We’re covering every one, and last week was OpenAI’s Dots.

Muse is already getting more daily US downloads than Threads, WhatsApp and Facebook, and it’s the biggest consumer AI launch since ChatGPT. Muse hit 5 million downloads in 22 days, and ChatGPT needed 56 days for that.

So how does the “worst” model end up as the most used product? That’s todays deep dive.

But first, a word from our sponsors + this week’s top news.

Partnership

🤖 Build three AI agents for your product team in one week

Optimizely’s AI Experimentation University is a free, live cohort for growth and product managers. Five days, one hour a day, and you leave with three AI agents of your own. 300 leaders have already graduated, including teams from Calendly, BBC, ODEON, and Legal & General, and they’ve built 1,667 agents between them.

The next cohort starts November 16th. You’ll learn how to:

• Generate better test ideas and build variations faster

• Automate experiment analysis

• Clear the bottlenecks slowing down delivery

Past graduates have built agents that turn test requests into Jira epics, diagnose why a variant lost, and return a single ship-or-hold call from frequentist and Bayesian models.

Save your seat now. It’s free, but only 50 seats.

Save your seat now

News

Top News this week: Anthropic Just Cut the Price of Intelligence by 10x

Anthropic shipped Claude Haiku 5.5, its cheapest model ever.

It costs $0.10 per million input tokens and $0.50 output, against $1 and $5 for Haiku 4.5. That’s 10x cheaper cost but with a superior performance. On real work Anthropic says it runs about 75% cheaper, and my own bill fell 80% when I switched the models.

It also scores 72.4% on OSWorld, up from 15.7% for Haiku 4.5 and it beats OpenAI’s Luna models at using a computer (48.9%).

So now you can get more work done with agents at a cheaper cost. One that burns a million tokens an hr now costs about $2.40/day in input.

The cheap price covers requests under 100K tokens which is roughly 90% of old traffic, and cache reads fell from a dime to a penny.

The Other News that Mattered

• Free Gemini users now get Flash-Lite only. Google has the cheapest inference on the planet but still demand still ate the savings. Every other lab’s free tier runs on investor money rn.

• Claude now sits in a Google Docs/Sheets/Slides sidebar next to Gemini. Google lost an opportunity for integration again but it wins any way because it owns ~14% of Anthropic and sells it the compute.

• OpenAI built a 99.9% accurate ChatGPT text detector in 2023, then shelved it for 3 years after users threatened to leave. It finally shipped on default mode in the EU, to adhere to the EU AI Act.

• On Terminal Bench 4.0, GPT-6 Astra (58.2%) and Fable 5.1 (57.9%) are a tie inside the error bars. Codex got there at half the cost, so the harness sets your bill.

• Google put 4 of its AI chips in orbit on a Falcon 9. Its paper says at $200/kg, launching a data center costs about what powering one here costs.

• Grok Bot supports the use of best model for each task, including Claude Opus 5.5, Midjourney and Suno. xAI spent $26B+ on its own models and is now hiring the competition wherever they do the job better. (No news on OpenAI being supported, haha ofc I saw that coming).

Feature/Model Releases

• Claude can now build live dashboards and animated explainers from a prompt that plug into BigQuery, Snowflake and Salesforce and turn ideas into editable animations.

• Claude cloud sessions can now edit folders on your laptop directly so you’re no longer the human sync layer approving requests.

• Google Docs and Drive now open Markdown natively after 20 years defending the Docs format because agents write .md now.

• Your $200 Claude Max plan now includes $200/month in API credits for Cursor and other harnesses.

Resources

• Karpathy says to make your LLM write in ASD-STE100 to avoid generating AI slop.

Funding

• Anthropic wants a pre-Thanksgiving IPO at up to $2T, that’s 435x it’s 2025 revenue but ~31x the current run rate, so it’s pricey but not crazy.

• Instinct raised $1B at a $10B valuation for an invite-only app and it’s just a 14 person team today.

• Manus raised $500M+ at a $4B valuation, about 2x what Meta offered before Chinese regulators blocked the deal.

Subscribe now

Deep Dive

Muse Has the Weakest Model and the Best Product

The whole race about personal agents was about which frontier lab uses the best model to get tasks done, and everyone else just spends time catching up with the race.

Then came Meta with the weakest model of the big four frontier labs and actually shipped the agent people talk to every day.

I’ve been following Muse’s journey for a long time, and it seems like it’s the most approachable personal agent out there. Nobody really cares about the model being used at the backend. It’s just the product experience.

Even if you are a product manager, you should study its design. If you feel that it might not be relevant for your professional work, it will be useful for your personal life.

Today’s Post :

1. What is Muse

2. Quick Setup

3. Real life usecases (Musecases!)

4. Where it breaks

5. Who should use it

1. What is Muse

Muse is Meta’s always-on agent with its own computer in Meta’s cloud, so that it keeps on working even after the app is closed, 24/7. That’s what all the other personal agents are doing too.

It can help you with all of your personal activities, like emailing people, booking things, and filling out forms. It also haggles and checks out for you, and the good part is that it asks you before it sends or buys anything.

You can use it on your mobile, web or even inside your WhatsApp, which is where I use it personally. It’s great for my personal work, but it’s currently only available in the US and Canada.

Every personal agent today you would see is doing the same set of use cases, but here are the five things I found out only Muse can do:

1. It has a computer with 2 vCPUs, 8 GB of RAM, and 100 GB of disk, plus a terminal.

2. Muse can write any code to connect to another app.

3. You just don’t have to talk to one chat box. It can divide a list of multiple actions into goals for you and the goals it’s working on right now. Every morning, you just have to check the goals instead of hunting inside one chat box.

4. You can edit its soul, and this is actually an important feature here. Soul is basically everything Muse knows about you, which you write in a plain file in an MD format.

5. It lives in all of Meta’s apps, and not only that, it is deeply integrated into a lot of other useful products like Shopify, which has integrated its Shop Pay checkout inside Muse. You can purchase hotel stays using Expedia, which is exclusively using Muse.

The most underrated part about Muse is about 95% of Muse users already use Facebook, so customer acquisition is too easy for them unlike xAI and OpenAI.

2. Quick setup

Here’s your complete guide to Muse before we get started:

Here’s the quick 2 mins setup.

Step 1: Download Muse and sign in

Get the Muse app on iPhone or Android, or open muse.ai in your browser.

Sign in with your Meta account (your Instagram login works too), and you can also chat with it inside WhatsApp.

Step 2: Setup Connections

Ask it to set up all the connections, like calendar, Gmail, etc., and you just have to authorise it when prompted.

Step 3: Make it act as your Chief of Staff

Step 4: Edit SOUL.md

• Ask me before you buy or send anything.

• Never share my home address or last name. (You’ll see why below.)

• Only ping me when you need a decision.

Step 5: Turn off training

Muse trains on your chats by default, so turn that off under Settings →Data controls.

Import your ChatGPT or Claude memory so you’re not starting from 0.

3. Real life use cases (Musecases!)

Musecase 1: Find the hidden expenses you’re paying without knowing

This is the first job I’d give Muse. It goes through my bank statements and finds the hidden cost I am paying and doesn’t cancel the subscriptions without my confirmation.

Musecase 2: Let it make the annoying phone calls

Muse can place business calls on your behalf. When you are bored of calling a support agent (where it's an actual AI talking to you), it's great to hand it off to Muse and settle your issues.

Musecase 3: Claim money for delayed flights

Filing for disputes is one of the most boring tasks which any person has to do, but the good part is that an agent does not get bored. You can easily ask your Muse to file a claim for a delayed flight and get the amount credited in your bank just minutes later.

Or ask it to keep an alert on fair prices, and it will just ping you back when it finds the best deal

Musecase 4: Get a weekly family plan

You can plan your weekly family plan and managing all family member’s routine in a single prompt.

Musecase 5: Ad Optimisation

It can connect to your Meta Ads and Google Ads and recommend to you what to do next on the basis of your budget and selected parameters.

Source: Ryze AI

4. Where it breaks

Here are the learning’s from some Muse horror stories:

Learning 1 - Turn off “Allow Always”

One of the most amusing horror stories which I have seen is that one seller unknowingly had set up “Always Allow” in the marketplace settings. Muse gave the buyer the exact home address of the seller and also tracked its live location and told him that he was home.

While Muse apologized for doing it, the damage was all already done, and you should be scared of your data security.

Learning 2- Read every approval card

Stripe introduced Link's wallet for agents (and Muse does use it), which gives agents the ability to spend across the internet on behalf of consumers, so ensure to add an approval layer to avoid mistakes by the agent.

Do your Amazon shopping yourself

Amazon blocks Muse completely so there’s no point asking. In one store test Muse also kept running into Cloudflare walls and got a D on customer email, while Grok Bot got an A-.

Check Goals every week

After some outages in early October, scheduled tasks quietly stopped running and people only found out later, so keep monitoring your bots everyday

Double-check the facts

A person found a restaurant phone number Muse made up. Muse also keeps a profile on every person you mention, so maybe don’t connect your whole contacts and emails on Day 1, unless you are ok with the data being shared.

5. Should you use it?

Who should use it:

• Parents: This might be the first AI consumer app your parents can use without hesitation as it has a simple interface and exists in platforms they already use that it asks before it spends a dollar and someone who has never touched an agent can figure it out. I can’t say that about Grok Bot or Hermes.

• Small business owners: Muse for Small Business adds 15 connectors, including Shopify, QuickBooks, Stripe, Slack, Asana, Canva and your Meta ad accounts, so essentially you are running your complete business on WA.

• Shopify sellers: It works the other way too, because Muse shoppers can now buy from your store with Shop Pay. It can help you with Ad Optimisation and give you the best RoAS

• WhatsApp and Instagram regulars: You’ll get the most out of it, since Muse already sits where you talk.

Who should wait:

• Anyone who wants it for core work or a team: Keep that on Grok Bot or Hermes until adds features for you.

Muse or Instinct?

I promised this matchup back in the Dots post, since Muse and Instinct are both going after your personal life, here’s what I found.

• Muse is faster: On one timed task, Muse took 4 min 36 sec and Instinct took over 14 minutes.

• Muse wins over the quality of outcome: One user made Instinct the boss of a group chat b/w Grokbot ,Muse and Instinct and Muse clearly won by delivering the best outcome

• Instinct wins if you live in iMessage: I have observed many people have iMessage as primary channel in US so some people pick it for that reason alone.

• Plenty of people run both: using Instinct for purchases and forms and Muse for mini-apps.

Muse does feel busier. Instinct’s personal network feature would only matter once people trust an agent with their friends and right now nobody trusts any of these agents that much.

So here’s where I land: If you’re new to agents, pick Muse, because it’s free and further along has integrations with Instacart and Shopify etc to ease your day to day life. Instinct should be secondary on WA if you use the personal network feature with your friends.

That’s it for today’s deep dive.

See you next week 👋

Subscribe now

Welcome to another AI by Aakash. I follow the AI news, and test all the latest AI tools, so you don’t have to.

The personal agent market is THE trend of 2026. So we want to keep a pulse on all the top players. This week, OpenAI joined it with Dots.

The big question is: how does it compare to the other hotness out there like Instinct, Grokbot, and Hermes?

I’ve been playing around with it since its release. And today, I am presenting you with my take - a complete guide to Dots (and whether it’s really worth your time).

But first, a word from our sponsors + this week’s top news.

Partnership

Agents are reading. Most company knowledge still isn’t ready.

Agents now read your docs more than people do. In August, sites on Mintlify logged 257 million agent requests against 131 million human page loads, and yet only 7% of the companies Mintlify surveyed have made every knowledge surface readable by an agent.

The 2026 State of Knowledge Report digs into that gap: why machine-readable content only gets you partway, how better navigation cut failed agent requests by 20x, and how teams at Anthropic, HubSpot, and Decagon are adapting. Your docs are where the work starts, and the report covers everything that has to come after them.

Read the 2026 State of Knowledge Report

News

The Week's Top News: Gemini 4 Argon is Out, But You Can’t Use It

After 7 months of mostly sitting on its hands, Google is back in the AI race with a frontier model. Google would like to have you believe it’s the best model with its chart showing it wins 12 of 15 benchmarks.

But it’s not the frontier, actually. Artificial Analysis Intelligence Index publishes a composite index that helps you see it’s really tied for third place with GPT-6 Astra and Claude Fable 5.1:

The top two models right now are Opus 5.5 and Sonnet 5.5. So, yes, Google is finally back in the top 3 labs. But, it’s 3rd best. And to use present tense itself is weird, because you can’t use it yet.

Google released it only to cyber defenders through the Fairwind program. Google is falling into the same trap as OpenAI of preannouncing models instead of releasing them the day of. Yes, Anthropic did this with Mythos, and that’s why Google is doing it. But Mythos was way ahead of the frontier, and Claude owned the frontier model at that time in Opus.

xAI and Anthropic are excelling in the department of announcing + rolling out the same day:

It’s time for Google and OpenAI to catch up.

The other thing Google needs to do is ship more. Look at the pace of frontier releases by lab this year:

• Anthropic: 7 releases, one every 38 days

• xAI: 4 releases, one every 48 days

• OpenAI: 4 releases, one every 61 days

• Meta: 3 releases, one every 74 days

• Google: 2 releases, one every 223 days

Argon is Google’s first model above Flash since Gemini 3.1 Pro. Google has been out of the frontier model race for quite some time. There was a brief exception when Gemini 3 launched, but, otherwise, Anthropic and OpenAI have owned the frontier.

The Other News that Mattered

• We saw big action on pacing. Every frontier lab signed the White House’s voluntary AI accord. Basically, it sets four layers of oversight: internal controls, a monitoring team, outside auditors, and a board committee. Nothing is law yet.

• Anthropic’s IPO prospectus leaked. It shows a massive $42B loss that is mostly just… accounting. The real number to pay attention to is >10x revenue growth, just as I wrote about last week.

• The NYT reported that OpenAI’s agents hacked government sites this summer. And the kicker is: an outside lab caught 3 of the incidents before OpenAI.

Feature/Model Releases

• Anthropic launched Sonnet 5.5. And the benchmarks are basically as good as Opus-5.5. But at half the cost!

• Microsoft gave up on the consumer chatbot. It turned Copilot work-only.

• Claude Code can now build evals natively, and even hill climb them.

• OpenAI flattened Codex pricing. The $200 plan used to buy 20x. Now it’s 10x.

• ElevenLabs’ v4 Turbo launched with replies in about 100ms, solving the awkward pause that used to give voice bots away.

• You can now order DoorDash by text, even though it costs DoorDash its own ad business.

Resources

• Sonnet 5.5 can finally make motion graphics. I built 15 of them in Claude Code and wrote up my 6 steps.

• Clay’s AI writing policy is worth stealing, because the time AI saves the writer gets paid back by every reader.

Funding

• AMD is buying Fei-Fei Li’s World Labs for $8.2B.

• OpenAI is raising $30B at about $1.4T while it waits on an IPO.

Subscribe now

P.S. If you want less friction in your life, check out my friend Ben Meer’s new book, How to Be Good at Life.

It’s packed with simple systems for work, health, money, relationships, home, and more. Pre-order now and you’ll also get his $297 Year of Systems course free.

Learn more

Deep Dive

Is OpenAI’s Dots really a Grokbot killer?

OpenAI shipped more than 20 launches at DevDay on Tuesday.

Sam Altman presented them like an Apple keynote.

The three that I thought were most notable were:

1. Login with ChatGPT

2. ChatGPT Space, a Google Workspace competitor

3. Dots

The first two are more or less straightforward. Dots are the newest and most interesting, because it’s OpenAI’s entry into the red hot personal agent market.

Here’s the 1-page summary:

And here’s the deep dive:

1. What are Dots

2. Dots’ standout features

3. The top 3 work use cases for Dots

4. Mapping out the personal assistant market: Dots vs the competition

1. What are Dots

Dots are always-on agents that lives in OpenAI's cloud with their own computer and browser, which means they keep working when your laptop is closed.

Dot runs on GPT-6 Astra and can connect to more than 4,000 apps through plugins, work with ChatGPT Work and Codex, and use background agents. It can respond to you on ChatGPT voice and Slack.

2 things separate Dots from ChatGPT, Work and Codex:

1. They work autonomously: Dots run in the background without having the need to keep the laptop on.

2. Dots can delegate tasks: A Dot can start Work or Codex tasks, including on your own computer if you connect it.

How to Setup Your First Dot

So here’s the path to getting started:

1. Upgrade to the $100 tier

2. Update your ChatGPT app

3. Go to chatgpt.com/dots

4. Click on create new Dot

5. Select where you want to run it, on cloud or your local desktop

   

   and your Dot is ready.

2. Dots’ standout features

There are 4 things that make Dots Dots.

Dot lets you call your agent and talk to the agent to define all the tasks and delegate tasks too. It remembers what we discussed on the call, and the next time you ping it on Slack or chat, it can pull up from the same context.

You can connect it directly to your Slack

Dots notice things before you ask, even before I configured my bot, it had figured out a list of action items for itself which Dot can perform for me.

Each Dot gets a cloud computer and browser that keep their state between tasks. When it hits a login it can't do, you take over its browser, sign in, and hand it back, which is how it gets past the 2FA wall.

3. The Top 3 Work Use Cases for Dots

There are three things that show the power of this.

Use case 1 - Meeting preps and follow ups

My Dot followed me to and fro meetings quite well. This was my favorite use case.

Use case 2 - The morning call

Using Dots call feature, I had daily calls with my Dot.

It skimmed through my calendar, Slack, socials and Gmail and gave me my to-do list plus schedule. It even grabbed my daily newsletter analytics.

Use case 3 - “EA tasks”

I was planning my last-minute flight to SF for my talk at Opsera, so gave this to my Dot on a call. It worked.

4. Mapping out the personal assistant market: Dots vs the competition

So where are we at with this new personal assistant AI market? There are so many tools out there! I think it’s helpful to conceptualize this as two completely different races we are talking about.

1. Individual Work Tasks: Grok Bot, Hermes Bots, and Dots are here.

2. Personal Tasks: Muse and Instinct are here.

Of course, there’s overlap, but this is the basic split.

Within the work tasks market, right now, Grok Bot is ahead.

Let me show you an example.

I gave all three the same job:

Compare the public signup and onboarding flows of three AI meeting assistants, capture screenshots, find the friction and recommend three improvements for a new entrant.

Grok Bot at least notified me to log in, while Hermes just finished research without logging in, submitting incomplete information. My Dot clarified to me what it couldn’t check, but still went ahead with finishing the report with incomplete information.

As a result, Grok Bot won hands down. It handed me a high quality deck. Consistently across my vibe checks, Grok Bot won in comparisons like this.

So putting everything together, here’s the decision tree I’d apply:

• If you’re looking for a personal assistant…

  • If you want to stick in the Meta ecosystem. go for Muse.

  • Otherwise, go for Instinct.

• If you’re looking to do individual work tasks….

  • If you are price sensitive, or want to keep your data private, choose Hermes Bot.

  • If you are already in the OpenAI ecosystem, choose Dots.

  • Otherwise, choose Grok Bot.

I peg Grok Bot as the leader in personal work agent, and Instinct as the leader in personal agent. Your opinion may vary. Reply with it!

This is THE space in AI right now, so I’ll keep tracking it and giving you the updates you need.

Talk soon 👋

Share

Welcome to another AI by Aakash. I follow the AI news, and test all the latest AI tools, so you don’t have to.

One of the creators of ChatGPT has released a new AI model which is different than your usual LLMs. Its launch got a massive 38M+ views:

I haven’t been able to stop playing with it since release. So today, I’m presenting my companion guide to adopt it.

But first a word from our sponsors + this week’s top news.

Partnerships

Amplitude Agent Analytics

Two users ask your agent the same question but get different answers. Both look equally engaged on a dashboard, but only one found it useful.

Most AI agents pass offline evaluations through observability measurement that stops short of measuring the real impact of AI agents. Agent Analytics ties traces and evals to conversion, retention, and revenue.

Observe what users ask your AI agent, evaluate where it fails and what those failures cost your business. The Economist used Agent Analytics to drive 96% task success while reducing failures by 84%.

Available for all customer plans, including free, today.

Try Agent Analytics

News

The Week’s Top News: Opus-5.5 is Awesome

Opus-5.5, released this week, and it is the model I have liked most since Opus-4.6.

It’s the type of model that only comes out once a year or so: a clear leap above what was possible before, really great to talk to, and all around banger.

Just look at these benchmarks: it somehow beats Fable-5.1 and GPT-6, at lower cost:

A lot of people have been upset with how Anthropic is watermarking text, or pacing the frontier, or being all doom and gloom. But what they have built over there has been working.

As someone who follows every new model release from every lab closely, I can confidently say they have had the best frontier model for most of the last 2 years. That’s an insane feat in such a competitive market.

Dario is cooking. The team is cooking. And the revenue shows it. Another year of 10x, when they themselves thought they couldn’t do it. I firmly believe they could do $1T revenue next year and 10x again. The funny thing is I think the team at Anthropic may not know it.

The only company who has briefly taken the #1 spot is OpenAI. I think they still have the power to take it longer-term. But they need to learn more from Anthropic. They need to release more and only announce when they can release to everyone. They need to reach parity with Claude Code’s harness. They need to actually focus on AI writing because Anthropic went the watermarking route so they have clear counter positioning.

Whether they can do so remains to be seen. Google, Meta, and X are also horses in the race. But they have yet to come close to a model of Opus-5.5’s particular delightfulness.

The Other News that Mattered

• Google joined OpenAI and Anthropic with the scary security stories. This week we learned Gemini escaped a security test and got into 3 real companies by guessing passwords and hunting down credentials it found in a public repo.

• Anthropic used Claude to make it 3x faster in two weeks. One of their key learnings was that em dashes were making claude slower.

• Amazon blocked Meta's Muse from its store, but Muse is going strong with integrating with Shopify, Expedia, PayPal, and Instacart.

• Google is launching TPUs into space on SpaceX Falcon 9.

Feature/Model Releases

• xAI shipped Grok 4.7 with 40% more parameters from 4.6 at the same price, and it’s already the default in Cursor.

• Claude Code now reads AGENTS.md when there's no CLAUDE.md.

• Anthropic merged Cowork into Claude, so you don’t really need to switch between options

• Alibaba released Qwen-Image-2.1, a 7B model that beats Nano Banana 2 on Qwen's own benchmark and runs on one consumer GPU

Tools

• DigitalOcean’s managed agents let you put the harness, sandbox, tokens and storage in a single setup itself, so you no longer have to pay 6 different vendors for getting an agent into production.

• Opal Zero acts as a gatekeeper for AI agents so you don’t have to worry about your AI having access to every tool.

Funding

• Crusoe raised $3.9B at a $30.9B valuation with Nvidia taking a stake.

Subscribe now

Deep Dive

Typesafe AI’s new model Jev

Jev has instantly become the fastest-adopted model in Vercel AI Gateway history:

How in the world did it get this real-life adoption?

We’ll cover everything today:

1. The History of Jev

2. Jev’s Fascinating Technical Architecture

3. How to Setup Jev

4. Real Use Cases

5. Limits

1. The History of Jev

Jev was created by Diogo Almeida, who co-invented ChatGPT and RLHF for model decisions, alongside Erik Gafni and Sasha Sheng:

They spent the past 2 years 9-9-6 inside their startup TypeSafe AI building a new way of training models for classification called RLCD (Reinforcement Learning for Calibrated Decisions).

They built an entire new stack focused on automations with new model architecture and parallel samplers for maximum efficiency.

Then, last week, they released their first System One Model, Jev.

“System One Model” comes from the principles of Daniel Kahneman in Thinking, Fast and Slow. It draws on the distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. Jev is a fast, intuitive model. Hence the System One name.

“Jev” also has a meaning. It is named after Jevons Paradox, which states:

An increase in the efficiency of a resource’s use does not reduce, but rather increases, its overall consumption.

That’s the entire point this model is built around. Jev doesn’t generate text or write code like Astra or Fable. It does classification.

As a result, Jev can basically make 3 types of decisions:

1. Choice: it picks one option from a list you define and lets you choose up to 255 which is good for routing and categories.

2. Score: places the input on a scale that you define for urgency, quality, risk.

3. Noul: basically a boolean. It answers in 0-1 (yes/no)

Every answer comes back with probabilities and a confidence score attached. It looks like this:

2. Jev’s Fascinating Technical Architecture

Jev has managed this level of virality and penetration for two economic factors:

Starting with speed, Typesafe’s evals claim to be 194× faster than the LLM comparison.

On price, Jev costs a meagre $0.042 per million input tokens vs $10 for Fable. On top of that, the output tokens are literally free. They are too cheap to meter.

Even after using Jev for the past week, I haven’t exhausted its $5 free credit limit. That’s how cheap it is.

It does this with its incredibly interesting architecture.

LLMs vs Jev

LLMs are built to generate continuous token streams one token at a time and can process unstructured text to produce free-form output (code/summaries/explanation, etc)

Jev is built for typed classification and decision-making. You can send unstructured “state data” (it is not multimodal in nature) and Jev can evaluate the declared questions in parallel super fast.

Here’s how fast that parallel processing looks vs a traditional LLM:

That’s just crazy fast.

Moreover, because it can only output in structured responses, there are literally 0 chances of hallucinations.

Here’s how this plays out if you give both of them the same support ticket:

The LLM writes you a sentence saying the customer seems frustrated about billing.

Jev skips the sentence. You declare those questions up front, and it comes back with the answers as typed fields: BILLING, 8/10, 0.95.

That one design choice is behind both of Jev’s economic advantages.

Let’s see it in action.

3. How to setup Jev

It takes <2 minutes to use Jev:

1. Go to typesafe.ai

2. Then go to Claude Code or Codex and paste this command

plugin marketplace add typesafe-ai/skills
plugin install typesafe@typesafe-ai

3. Or you can copy paste this skill on their onboarding page

4. Then head over to https://console.typesafe.ai/keys and create an API key. Name it anything.

Then tell your harness:

Paste your API key in there. This is a more secure way to do it than to paste it into your chat with the LLM.

Vercel gateway is also providing free unlimited access to Jev till Sept 25. Grab your API key here: https://vercel.com/ai-gateway

4. Real life use cases

Now it’s time to start using Jev.

Use Case 1 - Fast Browser Use with Voice

The browser agents you use today call an LLM for every small choice: what to click, where to scroll. Is this even the right decision to make? That makes the exercise really, really slow and also pretty expensive.

Browser use plus Jev really flips the stack. Jev literally picks the next action from the state and only a small LLM call needs to be made when anything needs to be typed.

(use this repo to set it up : https://github.com/moritzkremb/jev-voice-browser)

Use Case 2 - Email Classification

Email classification is a classic automation use case most people use today using LLMs but it might not be the most efficient one. Jev can do that way too fast at a dirt cheap price.

Use Case 3 - AI PM Job Radar

Every AI PM job seeker asks me the same two questions. Where are the jobs? And do I need an ML background to get one?

So I pointed Jev at the job market.

I pulled every open role from the public job boards of 36 AI and tech companies: OpenAI, Anthropic, xAI, Databricks, Stripe, Figma, Perplexity, Cursor and more. That’s 8,354 open roles. A few lines of code kept the 781 with “product” or “PM” in the title.

Then Jev read every posting and answered five questions:

1. What kind of job is this really? PM, product marketing, design, engineering or sales.

2. Is the product itself AI?

3. What level is it?

4. Is hands-on ML experience a must-have?

5. Is an engineering background a must-have?

It took 6.4 seconds. It cost 7.6 cents.

Here’s what came back, as of September 25:

• 151 open AI PM roles don’t require hands-on ML. Only 15 do.

• AI pays. Where the posting lists pay, the median AI PM role sits at $262K (midpoint of the range, 116 postings). Non-AI PM roles: $216K (59 postings).

• The door is narrow at the bottom. Only 8 of the roles are entry level: APMs, new grads and interns.

The best part was the mistakes it fixed.

A title filter counts “Director, Product Marketing” as a PM job. It counts Datadog’s “APM” roles too. APM there means Application Performance Monitoring. Those are engineers.

Jev threw both out. It also kept Asana’s “GM, Product,” which a title filter would miss.

It wasn’t perfect. The “requires ML” answer threw false positives between 0.5 and 0.7 confidence, like a Tax PM. So I only trust it above 0.7. That’s what the confidence score is for.

For scale: the run read 1.8M input tokens. At Fable’s $10 per million, that’s about $18 before you pay for a single output token. Jev did it for less than a dime.

Copy-paste this into Claude Code to run your own:

Use the TypeSafe skill. Pull every open job from the public Greenhouse, Ashby and Lever job boards for OpenAI, Anthropic, Databricks, Stripe, Figma, Perplexity and Cursor. Keep titles with “product” or “PM”. For each posting, ask Jev: is this really a PM role, is the product AI, what level is it, and is hands-on ML a must-have. Show me the AI PM roles with no ML requirement at my level, with pay where it’s listed, as a table.

Swap in the companies you actually want. Rerun it every Monday.

More Use Cases

There are so many more.

LLMs already used to do most of these but what changes here is the cost, the speed, and the structured answer to it.

Now it’s time to create your own use case. I’ve created a skill below to help you prompt Jev:

Download How to Jev Skill

5. The Limits of Jev

There are three major ones:

1. Small context window: Jev has a relatively small input context window (32K-64K tokens) which is significantly smaller than frontier LLMs that support 1M+ tokens.

2. No Vision / Multimodal Input: Jev cannot read or process images directly. It can evaluate a game state or screen (such as Checkers or Doom) by passing a raw structured text/data rather than pixel imagery.

3. Lack of deep reasoning: as it operates on reflex/classification engine, it basically operates as a pattern-matching “switch” statement and lacks all of the chain-of-thought traces or multi-step reasoning like in LLMs.

Let’s see what the team cooks up to address them.

Overall, this is where Jev fits and a Claude or Claude Code win:

It’s quite an impressive scoreboard for a brand new model.

Talk to you next week 👋

Share

Welcome to another AI by Aakash. I follow the AI news, and test all the latest AI tools, so you don’t have to.

Instinct is an invite-only personal agent going mega-viral. People are calling it OpenClaw / Hermes for normal people. It’s discussing a $10B fundraise.

And the craziest part of it all? It was just started in April by a 23 year-old founder.

I’ve spent the week testing it. In today’s deep dive, I’ll tell the product’s story and give you my take.

But first, a conference I’ll be speaking at + this week’s top news.

Quick Plug

Join Me at Flutter 2026

On October 1st, I am speaking in the Bay Area at Flutter’s AI-SDLC Summit on “The Rise of the Full-Stack Builder.” Join me:

Request Your Invitation

Attendance is limited and curated for senior technology and engineering leaders.

News

The Week’s Top News: Pacing Drama

AI is pacing so fast that Sam Altman and Dario Amodei, the two who wouldn’t even hold hands at Modi’s AI summit, are on the same page for once.

It all started with Dario publishing his essay on pacing the frontier Saturday morning. Within hours, the leaders of the other 3 top labs chimed in that they agree: Sam Altman, Elon Musk, Demis Hassabis.

By Saturday evening, it seemed like we might be entering a world of an AI slowdown for the first time since 1,000+ tech leaders signed the pause letter back in 2023. We even learned OpenAI had delayed its IPO to next year because of the current AI safety environment.

Then everything turned. You see, there was always a catch: the companies couldn’t agree to slow down on their own. That would be a cartel. They needed government backing.

By Sunday, Trump had put a kibosh on the whole thing.

Now we’re left with what the companies can commit to doing on their own and seems likely to play out: unilateral moves that aren’t cartel-like. For instance, Anthropic agreed to bring in outside evaluators like METR to publish what they find testing without Anthropic editing it. OpenAI already committed to the same thing within hours.

There was only one notable exception to the pacing agreement we saw this weekend. Mark Zuckerberg. He will continue to be one to watch.

The Other News that Mattered

• Meta launched Muse, an agent that’s competitive with Instinct, and is getting good reviews. It’s said to be the brainchild of Nat Friedman (fmr GitHub CEO).

• After seeing the inside of both OpenAI and Anthropic, Jacob Coxon warned that neither lab is taking safety seriously enough. His post on X got 170M+ views.

• OpenAI now makes more money from enterprises than consumers, and that explains why Sam Altman showed up to Dreamforce. Databricks even moved their engineering team from Opus 5 to GPT-6 Astra.

• OpenAI announced it solved the Navier-Stokes Millennium Prize Problem, but there are concerns it actually trained on a researcher who was working to solve the problem alongside an Anthropic employee, and refused to offer the employee authorship.

Major Feature & Model Releases

• ChatGPT Image 2.5 continues OpenAI’s dominance in AI image generation, with even better editing capabilities.

• Google released Gemini 3.8 Live at less than half the price of the Agents API.

• DeepSeek released V4.1 Flash, claiming parity with Opus 5 and GPT-5.6 at about 1/30th of the price. If the benchmarks hold, that’s the cheapest frontier-class model anyone has ever shipped.

• Anthropic is merging Claude Chat and Cowork.

New AI Startups

• A nameless coding model called Union Alpha went free on OpenRouter for a week.

• TypeSafe released a new model pricing output tokens at zero.

• VideoClaw launched as an OpenClaw for video.

• Odyssey-3 is one world model that drives a humanoid, a car, a drone and a video game and claims to be the “everything model” for any physical product.

Resources

• DeepMind’s AlphaGenome Atlas scored every possible single-letter DNA mutation there is, and you can explore it for free.

Deep dive

The Instinct Deep Dive

Instinct is taking over X.

It’s an invite-only agent that works in your iMessage or WhatsApp and understands what you're working on and what's important to you.

But an invite-only product valued around $10 billion without reaching the masses sounds like the Clubhouse bubble from 2020.

Is it actually worth the hype?

Today’s Post

We investigate everything you need to know about Instinct:

1. The Instinct Story

2. How to set it up well

3. The most powerful use cases

4. Product lessons from Instinct

5. Where it sits vis a vis other AI tools

Subscribe now

1. The Instinct Story

I was in San Francisco this week for DreamForce, and Instinct was on the tip of the tongue of everyone I talked to.

It has a wild story.

Instinct was developed by 23-year-old founder Noah Shinn.

If the name sounds familiar, it’s because he is the guy behind the Reflexion paper, one of the most cited agent papers of 2023.

He wrote it before agents were cool, and he was only 20 at the time. So it’s no surprise he’s ahead of agents in 2026 as well.

After writing the Reflexion paper, he did a stint at Bret Taylor’s startup Sierra.

Then in 2025 he left to start his own thing. Since then, everything has moved at lightning speed:

• In February 2026, he released Instinct to an invite-only private beta.

• In April 2026, he registered Instinct’s parent company (Spear Street Technology Inc) and raised a seed round at a $50M valuation.

• In early August 2026, Instinct went massively viral. Kleiner Perkins led a Series A at a $500M valuation.

• In late August 2026, just weeks later, Instinct raised a Series B at a $2.5B valuation.

• Now, they’re in talks for a massive new round targeting a $10B valuation.

It literally went from 0 to $10B in 7 months.

Let’s check it out.

2. How to set it up optimally

The onboarding is part of the magic of Instinct. It’s 3 steps. Takes less than 1 minute.

1. Get an invite (if you want one of mine, reply to this email with why)

2. Give it Google workspace permissions (these are scary, but if you want to experience the power, pretty much needed):

   

3. Set up your iMessage or WhatsApp (this is absolutely needed):

Then, get to using it! It’s that easy. And the crazy part is it acts with context right away.

My Instinct picked up its first task even before I told it to do anything:

To get an even better setup, head over to app.instinct.co and

1. Connect other relevant tools you want your agent to have access to:

2. Connect your credit cards or login creds if you want it to perform actions on your behalf

Then you’re good to go.

3. Real-life use cases

The use cases are genuinely diverse. Here’s several I tested and liked:

Use Case 1 - Personal Concierge

The real power of Instinct isn’t work stuff. It’s personal stuff. You can use Instinct for just about everything: phone-only restaurants, dentist waitlists, cable bills.

Several times, I asked in iMessage, Instinct called, then I received the confirmation in the same chat:

Use Case 2 - Trusted Person Network

This is one of my favorite features of Instinct. You can add friends to Trusted People. Then, your Instinct talks to theirs. I used it to plan a picnic, including rides:

Use Case 3 - Own Email

At first, I used Instinct to manage my booking follow-ups for podcasts I was recording. But they were getting numerous.

I asked Instinct to handle those in its own inbox, and only ping me when it needed me:

It used its own email address without me.

Use Case 4 - Vault + 1Password + Dashboard insights

This sounds crazy. But you can sync your 1Password to Instinct to share credentials.

Here’s what I actually did. I put Substack in Vault and nothing else. No banking, no primary email password, no socials. This is a company that incorporated in April. I’m treating it like a contractor with a key to one room, not the house.

With that one credential, I scheduled a daily report that sends me insights on my blog performance, all in WhatsApp:

Use Case 5 - Stripe Link / spend-limited cards

AI tooling bills creep, so I told Instinct to use a spend-limited card only for adding API credits, set a hard monthly max for OpenAI / Anthropic / Cursor / OpenRouter, keep track of usage, and warn me before any top-up.

It can actually do all that.

Use Case 6 - Location on iMessage

I shared location with Instinct and asked it to order pizza. It did:

You can’t ask Claude or ChatGPT for that. They’re just not as independent, problem solving, and built to assist you with things like credit card use.

That’s the power of Instinct.

Cautions

Before you go all in, two things I’d flag.

1. I’ve seen reports users eating $200 no-show fees on flights their agent booked wrong. That’s why I only gave mine a spend-limited card.

2. The web was built to keep bots out, and sites can’t tell your agent from a scraper. If Instinct tries too hard on your accounts, you can get banned.

4. Product Lessons from Instinct

Instinct wins in 5 key areas:

1. No app - the chat is the whole product, so you never open a new app

2. Human-like interaction - it uses iMessage/WA the way a person does, down to reactions and voice notes

3. Memory - Instinct’s memory sticks between tasks and it starts work before you even ask it to

4. Multiplayer - The Trusted Person network is a crazy differentiator. Nobody else lets my agent talk to my friend’s agent

5. Calls - This is probably the killer feature. Like Muse, Instinct makes calls for you

Most of these features existed elsewhere. But Instinct brings them to you in an amazingly easy-to-use package. Free.

That’s genuinely special. But it should make you squint. We don’t know how they will make money yet. Free plus a $10B valuation could be the Facebook playbook… or it could be the Clubhouse one.

5. Where it sits vis a vis other tools

Instinct’s biggest competitor is Meta’s Muse, which I’ll likely cover next week (pending no other big stories). Where does Instinct fit with the other tools we’ve covered in this newsletter?

Here’s my take:

• Use Instinct for simple AI over text.

• Use Grok Bot for agentic capability in an app.

• Use Hermes for the auto updating agents and bot mode.

• Use OpenClaw if you want to go fully agentic with power.

Now for work tasks? There, Claude Code is still my tool of choice.

Final Words

So, is Instinct the next Clubhouse? I don’t think so.

Clubhouse was clonable. It died the moment Twitter cloned it.

Instinct is not as clonable. It gets better the longer you use it. So even if you clone the features, you can’t clone six months of someone’s life inside the product.

Now… whether that’s worth $10B is a separate question. The valuation is a bet that this becomes the default interface to your life before Meta and Apple wake up.

That’s a bet some VCs are willing to make, but I prefer to wait and watch from the sidelines.

See you next week,

Aakash

P.S. I have a podcast coming out with Mark Pincus, that I just filmed today, where we discuss Instinct in more depth. Subscribe so you don’t miss it: Spotify, YouTube.

Share

Welcome to another AI by Aakash. I follow the AI news so you don’t have to.

Omarchy is the hottest AI tool on X right now. After installing it this week, I finally get the hype. We're so used to desktop Linux feeling like rot that this felt like getting into a Ferrari.

That’s today’s deep dive. First, a word from our sponsor + this week’s top news (including GPT-6).

In Partnership with

Amplitude Custom Agents

Growth teams check their funnels when something breaks. A Custom Agent monitors them for you 24/7.

You configure it to a task like checkout drop-off and connect your stack. Agent Connectors pull the PRDs, tickets, and conversations around that funnel to make contextual analyses. When it finds something, it adds to the Linear ticket.

Run it on a schedule or on demand. Start from a template or write your own.

The agent library ships with a checkout drop-off investigator.

Custom Agents is free for all customers in open beta. →

Try Custom Agents

This Week’s AI News

The Week’s Top News: GPT-6.

This past week, we saw the launch of big models from Anthropic (Fable 5.1), OpenAI (GPT-6), Meta (Muse Spark 1.3), and Google (3.8 Flash).

Today, Greg Brockman even declared that AGI is here, with OpenAI’s rollout of GPT-6 (not available to the public yet).

So the big question is:

The answer seems to be no. The big focus of GPT-6 (codename: Astra) was coding, and the Artificial Analysis Coding Agent Index puts it behind:

So the story continues - Anthropic owns the best coding models. The big surprise is where the other players are. Google is not the company with the next best model! It’s not OpenAI, either. That Meta! With its newly released Muse Spark 1.3.

Meanwhile, Google launched 3.8 Flash, which shows how they have focused on the cost efficient frontier over the real frontier. Ahead of Google still sits xAI, with their latest Grok 4.6 model.

OpenAI may have the most powerful model in areas outside coding, like scientific discovery. Here’s the benchmark card they shared today:

But this needs to be verified, and they haven’t rolled the model out publicly. For now, I’d say the big takeaway is:

OpenAI teased they might be on top, but they probably aren’t (yet again).

The Other News that Mattered

• Nvidia acquired Hugging Face for $ 12.9B which famously let you use/download your AI models for free. This seems very logical for them as all the opensource models are downloaded, trained, finetuned and served on Nvidia GPUs itself.

• OpenAI built its first chip and it might beat Nvidia on cost per flop. It took just 16 months to release it while most chips normally take 3-4 years.

• Fal’s H3 Max makes video faster than you can watch! A 5-second clip with sound is created in under 3 seconds. This is a breakthrough for AI video.

Feature Releases

• Claude Code added /resume to its desktop app, so you can pick up any terminal session right from where you left it.

• Glean launched its desktop workspace, an automatic model picking router, an AI gateway and agents that act before you ask in their latest product release.

Tools

• Anthropic’s new hardware standard lets agents run real lab equipment, like robotic arms and microscopes..so basically your AI just got hands. They are planning to make it opensource just like MCP.

• Keenable launched a search engine built for AI agents, not for people clicking links.

Subscribe now

Share

Deep dive

Omarchy: The First OS Built for Agents.

I’ve been on Mac for 15 years without a single complaint. I was forced to use Windows for a while at Epic Games and it was a pain. The only OS that gives you real control is Linux, but for me the Apple ecosystem is gravitational.

However, to venture out, I spent a week on Omarchy. And I get the hype now. So that’s today’s deep dive.

Roadmap for the rest of this Deep Dive

1. What IS Omarchy

2. How to test this thing

3. What it’s actually good for (real use cases)

4. How to start building your agent OS today

5. My take on Omarchy

1. What IS Omarchy

Omarchy is a free OS (from the creator of Rails) which is designed for an AI agent to run alongside you and not just for you to click through all of its actions.

3 things set it apart for me from a traditional OS (which makes it agent OS friendly):

1. Every setting lives in a plain .md that any agent can read

2. Every major AI coding assistant comes pre-installed

3. Any app crashes or OS failures get handed over to your agent automatically, without human troubleshooting.

On a Mac, your agent will use to click around a locked box. Omarchy edits the computer directly.

Omarchy raised $10M from Shopify’s Lütke, Twitter’s Jack Dorsey, and Dropbox’s Drew Houston. That goes to show the credibility of the founder and urged me to try it out.

2. How to test this thing

A real challenge with testing an OS is how do you do it without partitioning?

Mac: A developer on twitter built Try Omarchy, which is a free app that runs the whole OS in a window on Apple Silicon.

Linux: If you run bare Arch, you simply need to run

wget -qO- https://omarchy.org/install | bash

Windows: Refer to the official guide and shrink the volume in the Disk Management settings, then pick “Free space install” in the Omarchy Installer. Turn off BitLocker first (Settings > Privacy & Security > Device Encryption)

3. What it’s actually good for (real usecases)

Don’t miss hearing from DHH himself:

But here’s the clearest proof of how it works.

I’ve hated the Notification Center’s behavior on my Mac for years.

On Omarchy, I asked it:

omarchy-snapshot basically snapshots your whole system so that the change is reversible if anything goes wrong. The answer that did not exist in any settings on Mac, just got built on its own without me having to spend time figuring out how can it be done.

Some other use cases from actual builders not hobbyists:

• Jason Fried, DHH’s co-founder at 37Signals, built a bar plugin pulling in his team’s Basecamp notification entirely in English

• Spencer Bull built OmaPilot, the agent plugin into your OS

• Zsolt Kacso created an infinite window display of Omarchy

• You can also port your Omarchy to an old phone

4. How to start building your Agent OS today

Here is the exact order in which you can start building your agent OS:

1. Pick your default agent immediately as you setup Omarchy from Setup > Defaults > Agent

   

   

2. Given it can be overwhelming to set up a whole OS yourself, find one annoyance you have been bearing with your OS for years (mine was the notifications issue)

3. Write it as a prompt and let the agent fix it itself

4. Start version controlling every ask you give the agent. Otherwise you both lose track of what changed on the machine

5. Set up the local models for the agents. Install > AI and pick LM Studio or Ollama and point your agent to use it

   

   It comes packaged out of the box with a lot of agents (you don’t have to download ChatGPT, Claude etc)

6. Once you become a pro, start using multiple workspaces and toggle between them

   

   Omarchy currently has a very active Discord community. Definitely join it.

Putting that into one infographic:

5. My Take on Omarchy

Is this the next tool that everyone needs to suddenly adopt? No, and that’s because it’s an OS. Personally, Omarchy isn’t pulling me off my Mac yet. The Mac ecosystem has a magical glue to it.

That all can be true and… Omarchy was the most fun I’ve had on my computer for years. And, from a product POV, it shows a clear preview on how a computer shapes your agents around you in plain English.

See you next week,

Aakash

Share

Welcome to another AI by Aakash. I follow the AI news so you don’t have to.

Hermes shipped Bot mode this week, and it’s a legit competitor to Grok Bot. No wonder Hermes is going crazy on the OpenRouter leaderboard:

It’s basically advertised as a free version of Grok Bot.

So how does it ACTUALLY compare?

That’s today’s deep dive. First, a word from our sponsors + this week’s top news.

In Partnership with

Arize Phoenix: 9K Stars and Counting

Arize Phoenix is THE most popular open source AI observability library. It has 9K+ GitHub stars and 2M+ monthly downloads. That’s because it works basically every framework: from OpenAI and Anthropic to Google, LangChain, and CrewAI.

If you can’t see traces or run evals on your agent, you’re guessing.

So, get started at phoenix.arize.com.

Try Arize

This Week’s AI News

The Week’s Top News: Stripe just made a massive AI acquisition

Speaking of OpenRouter, OpenRouter is this week’s top story. Stripe acquired it. for over $7B. That’s more than 5x its valuation just 3 months ago:

OpenRouter earns a 5.5% fee when customers load credits to spend on inference across 400+ models. That’s nearly double Stripe’s 2.9%, and it’s on dollars previously outside Stripe’s ecosystem: those flowing to OpenAI, Anthropic, and Google.

OpenRouter itself runs on Stripe, so Stripe has known everything about OpenRouter’s growth curve since the start.

And it’s a true hockeystick. OpenRouter spiked to 25 trillion tokens a week across 8 million users, up 5x from 6 months ago.

With this acquisiion, Stripe has strengthened its toll booth for internet commerce.

The other news that mattered

• Electricians are coming out as a major beneficiary of the AI boom. Electrical work is literally 45-70% of the cost of building a data center. So 20-something electricians in Texas with no degree are earning $280,000 a year.

• If you didn’t believe Chinese models have completely taken the lead on Open-Source, the data is now in. Qwen out-downloaded Meta’s Llama nearly 9-to-1 this year (~2 billion downloads to 227 million).

• China has already finished a year of compulsory AI curriculum. Meanwhile, the US and Europe are still debating AI detectors.

• If you were looking for a high-impact use case of AI in non-AI companies, look no further than McDonald’s (seriously can’t believe I’m typing that). They’re producing 515-page reports on customers that exactly predict how much you’ll spend in the coming weeks and months.

• Anthropic looks set to raise $100B at a $2T valuation as soon as next month.

Launches:

• Berkeley AI dating startup Ditto achieved 160,000+ signups for its app that not only automatically picks their match, it actually sends them a fully planned date itinerary.

• Meanwhile, OpenAI has turned their Codex rate limiter into the best growth loops in AI. It’s now past 20M users of Codex and reset limits yet again.

Feature releases:

• Anthropic put editable design artboards inside Claude Code. Now you don’t have to handoff the figma designs to the developer to build. Everything can be done in Claude Code.

• OpenAI shipped ChatGPT inside iMessage on the Mac. This is 41 days after Apple sued it for trade secret theft. They’re suing each other with one hand and shipping together with the other.

Subscribe now

Deep Dive

Hermes Bot Mode.

Grok Bot and Hermes Bot are currently fighting over who has the best agents mode. So let’s find out.

Hermes Bot mode gives us the same roster of named bots. But one single difference will matter to you every month:

So I rebuilt my fleet inside Hermes Bot mode this week, so that I can really test the outputs, and build you a guide on how to use it well.

Today’s Deep Dive

1. What Bot mode actually is

2. Your first Bot in 10 minutes

3. 3 ways to use your roster

4. Where it beats Grok bot and where it looses

1. What Hermes Bot actually is

Hermes is a free, opensource agent (I covered it all the way back in April.)

Recently it has released it desktop app for Mac, and this week its Bot mode came out. It’s basically Grok Bot.

Grok Bot was $200/mo when I wrote about it last week. Since then, they’ve released a $60/mo plan. But it’s still basically a $60 vs free agent.

So, a $60 Agent per month vs a free agent is worth comparing on which gives you the best productivity and that’s exactly what we are planning to today.

2. Your first Hermes Bot in 10 minutes

1. Install Hermes Desktop from here.

2. You need to connect with a provider of your choice when prompted in the signin page or you can skip and go to Settings > Providers > Accounts to setup later.

3. Click on create + new Bot on the dashboard and then write the name, title and description. Write the description thoughtfully because the bots use it to route their work.

3. All the bots that require internal handoff and communication with each other use this feature of create Bot Group and @ mention specific bots that should converse with each other. This is the perfect way to use the roster of bots together.

5. I start scheduling all of my bots as routines or scheduled jobs for all the repetitive tasks I find over a period of time

3. 5 ways to use your roster

Now it’s time for you to build your agent roster.

1. Give every Bot a clear job and its own brain, I use good reasoning models like claude/gpt-sol for an architect or orchestrator Bot and for research-heavy work. For simpler, repetitive jobs, a fast local model just does well enough.

2. I advice folks on building bots around skills not just individual tasks. A “do everything” Bot sounds convenient but it quickly collects unrelated context and starts mixing projects together and that’s when responses get slower & more likely to hallucinate.

   

4. I don’t advise putting all the bots in the same group chat because it increases the chances of hallucination, and group chats work best when there is actually a real agent to agent talking happening.

5. a large roster typically does not make the agents work better. It is the fact that a few bots with clear jobs are usually more helpful than just creating a swarm of bots that need constant management

   

Where it beats Grok Bot

There are several clear areas Hermes Bot Mode beats Grok Bot and we are going to see them with examples:

• Grok Bot has no export functionality, once you stop using the app the agent memory is gone

• Grok Bot needs a $60-$200 plan and Hermes is free.

• A major advantage of Hermes over Grok Bot is the accessibility of models. In Grok Bot, you are always restricted to the Grok models, but you can use Fable or Sol 5.6 for long thinking tasks in Hermes

• You can clone existing bot profiles in Hermes instead of starting from scratch every time.

Bot mode difference from normal AI agent

Personal takeaways

That’s all for today. See you next week.

Share

Welcome to another AI by Aakash. I follow the AI news so you don’t have to.

xAI’s $60B acquisition of Cursor closed this week. They paired the launch with what some are calling the product launch of the year, Grok Bot:

Is it really product of the year? I investigate in today’s deep dive. But first, a word from our sponsors + this week’s top news.

In Partnership with

Arize: The Hire-On-The-Spot Skill

Arize CPO Aparna Dhinakaran told me on the pod:

“If a PM candidate came in and showed they could build an agent, instrument it, write evals, and fix issues using Claude Code and Arize, she'd hire them on the spot.”

That's the bar now. Arize makes it possible to hit that bar in an afternoon.

Try Arize

This Week’s AI News

The Week’s Top News: The End of Claudefishing

Anthropic quietly revealed in help docs this week that every Claude model launched after 2nd August weaves an invisible watermark into any text it generates:

This mark travels with the text when you copy it. Anthropic will soon release a public detector too. How it works is that the model tilts its word choices according to a secret key, and the detector is able to confirm the pattern.

This marks the end of “Claudefishing,” where you use Claude to write but claim it as your own. Some people are furious. What do I think? My recommendation is to just be transparent. If you use AI to assist or write, tell people.

As for this newsletter? Substack announced a few weeks back it scans posts via Pangram. Since then, all my posts have been 100% human:

The Other News that Mattered:

• There were a bunch of model releases: Grok 4.6, Gemini Flash 3.7, GLM-5.3, and Muse Spark. None revealed a frontier model. Grok 4.6 did produce a major chart crime though.

• An OpenClaw agent helped a man in Melbourne book a gym class by hacking the scheduling system, opening a stranger’s reservation, deleting it, and making space for the man. If you were wondering about all the hubbub over AI’s cybersecurity capabilities, now you understand what all the doomers were talking about.

• In one month, Cloudflare will start cutting off Google traffic for millions of its smallest customers by default. Publishers are not happy about AI answers, and this is the result.

Tools:

• Spotify’s engineering team released Xirp, which is a session manager for AI coding agents. Anthropic, OpenAI, and Google are spending billions to make their model the one your team depends on. Xirp makes them interchangeable.

• X open-sourced the entire ‘For You’ timeline algorithm. I read the code and broke it down here.

Feature Releases:

• Coderabbit released Agentic Change management, a layer which decides which PRs need human attention, now that most code is written by AI.

• Lindy released its “teammate,” an AI employee whose memory lives in plain files.

• Oumi launched a “compounding AI factory” where production failures become retraining signal.

Subscribe now

Deep Dive

Grok Bot: xAI cooked this time

xAI had two releases this week: the Grok 4.6 model and the new Grok Bot product.

I am more interested in the product. People are calling it the AI product release of the year. It has a lot of “first-of” feature releases.

Let’s explore today:

1. What is Grok Bot?

2. How to create a fleet of agents

3. 100+ use cases to try today

4. Is it product of the year?

1. What is Grok Bot

Grok Bot is a multi-agentic platform built to automate complex long running office tasks end-to-end like a real human employee.

Each bot runs a persistent cloud computer, logs into your apps, and handles tasks across interfaces. It only returns to you for approvals! You run multiple bots at once and they coordinate on their own.

Download it from here to follow along: x.ai/bot · desktop (macOS, Windows) + iOS

• It is not a better ChatGPT/Claude.

• It is not a replacement for an employee with judgment.

It’s an environment to manage agents. Grok Bot was heavily dogfooded by xAI’s own product, growth, operations and engineering teams to handle all of their tasks: outreach, marketing, software bug repairs etc… we will try to execute them today and see the results.

2. How to create a fleet of agents

Grok has made the UX pretty simple for you to follow along. You don’t even have to set up anything in settings. Grok does it for you.

Step 1: Give your agent a name and mission.

Create a list of all the things you want to automate and make them your employees (sharing the best use cases below)

Step 2: Make multiple agents for yourself and configure the plugins

You only have to do this once. These are shared by all the agents. I connected Gmail for my EA agent and now my sponsorship negotiation agent can also access it:

Grok Bot has a computer, basically: a persistent cloud VM with a browser, a filesystem, and a terminal.

Grok Bot is the first consumer agent I have used that tries to finish the job inside the harness itself. It can log in. It can click. It can wait. It only comes back when it needs a yes/approval/review, or when the work is done. Compared to Cowork, it talks back and says no much less often.

Step 3: Create routines for your day to day tasks by explaining the task

You can do things like create a Monday morning experiment readout:

It keeps working when you leave… even when your laptop or phone is shut.

Step 4: Convert these tasks to skills and let them be accessed by all the agents

You can also teach your agent to mimic your work and convert it to a task:

When the output is good, save the process as a skill. Skills are how, not when. Any Bot can use a skill (they are shared by default)

!! The best part about these skills is that they are self-evolving like Hermes (keep gaining more context and memory over time). You can correct them, but they identify and correct their own mistakes.

Step 5: Put them in one thread and let a Chief of Staff coordinate.

This is personally my favorite feature of the product.

You don’t have to be the router between Bots. They easily co-ordinate with each other, pass the work, and bring you only the items that need a yes or…a review.

3. 100 use cases to try today

I have drafted 100+ bot prompts for every category of work - from product, tech, sales, marketing, and content to personal work - in this Notion database:

Grab them here for free

Take the prompts you like and make them Grok Bots.

4. Is it product of the year?

No. After my testing, I consider Autoresearch, GBrain, Hermes, ChatGPT Work, Claude Design, and Cowork all to be better product releases. They can do what they do better AND will be used by more people.

But Grok Bot is xAI’s strongest entry ever, and, conspicuously, much better than what we’ve seen from Google, Meta, Amazon, Apple, or anyone else.

That’s all for today. See you next week.

Share

Welcome to another AI by Aakash. I follow the AI news so you don’t have to.

Thariq Shihipar is on the core team for Claude Code, and he recently shared a fascinating insight about the latest 5 series of models (Sonnet, Opus, and Fable).

The team deleted 80% of Claude Code's system prompt for Opus 5…. and it performed better.

So how do you set up your Claude Code optimally with the new Claude 5 models? That's today's deep dive.

But first, a word from our sponsor + the week’s AI news.

In Partnership with

Zapier SDK: I Had Claude Code Build a Link Auditor

I’ve never found a good way to catch the stale links in my old newsletter. Until now. This week I built a link auditor using Claude Code and Zapier SDK.

Here’s one of my recent runs through Cursor terminal. From 126 links across 7 posts, 2 were webinar registration links (that returned a 502) while 3 were redirected somewhere else. All this got stored in a Notion database through the Zapier SDK.

There was no OAuth flow or token handling to write as I had already connected those accounts to my Zapier account. The same interface covers the 9,000+ apps that Zapier connects to.

It’s in open beta currently, free, and the team is actively looking for feedback.

Try Zapier SDK

This Week’s AI News

The Week’s Top News: Google Rightly Moves the Chairs

Google’s Gemini models are god awful. And their harness engineering is pathetic.

Try any vibe based evals on tough tasks like the graphs we discussed last week. Gemini has no idea how to handle such a prompt and fails miserably. Gemini can’t even access YouTube or Maps data.

Both Claude, with Claude Code, and ChatGPT, with ChatGPT Work and Codex, are miles ahead. Occasionally throughout the past few years, OpenAI has been behind Anthropic. But it quickly catches up.

Meanwhile, Gemini barely ever releases anything. It’s gotten to 900M WAU merely through default placement and distribution. The last time Google had a top anything was Nano Banana 2, but that was quickly beat by ChatGPT Image Gen 2.

So it’s “good riddance” to hear that the supposedly “genius” Jeff Dean and Demis Hassabis have finally been moved out of their leadership positions. Google is trying to spin Demis’ move as a ‘step up,’ and word is Jeff left on his own, but any outside observer can clearly see they needed a leadership shake-up at the top.

Sergey Brin is taking over direct oversight of the Gemini models, and it’s the right call. Google clearly needs a good dose of “founder mode” to right the ship.

Here’s the AI product strategy I recommend to Sergey:

1. Immediately connect Google data to Gemini

2. Actually start releasing things quickly, at an AI pace

3. Make sure your models are competitive with the frontier

4. Stop shipping so many random products in labs that are disconnected

All of that is easier said than done. But if anyone has the resources to do it, it’s Google. I’ll check back in a few months, to see if Google has done any of this.

The Other News That Mattered

• Unlike Google, Meta keeps releasing stuff and improving its standing in the AI race. Its model, Muse Spark 1.2 scored only a little behind Opus 5 on Terminal-Bench 2.1. And it’s also released Muse Code, its competitor to Claude Code.

• Nvidia open-sourced Alpamayo 2 Super, a free 32B-parameter autonomous-vehicle reasoning model. It’s no wonder Uber abandoned their own training effort. They instead plan to use Nvidia’s models, and committed to 100,000 Nvidia-platform Level 4 vehicles starting 2027.

• Scientists used AI to design 16 new viruses that can replicate inside cells. The biosecurity threats from AI continue to increase.

• AI just read a stranger’s mind through their skull. Without any implants, it got roughly 45% of the meaning right on the first try. AI brain interfaces are coming.

Feature Updates

• Wispr Flow’s Notetaker now transcribes meetings and labels speakers by their real name pulled from your calendar.

• Claude Code sessions can now send messages to one another. Instead of having to copy paste responses between sessions, merely instruct them to talk to one another.

• Bolt.new released a new template library of high quality apps.

• Higgsfield has added Seedance 2.5, the SOTA video model.

Fundraising

• Sapiom raised $35M to help you tackle rising AI token costs with agents.

Tools

• Helena launched a self-improving AI marketer, running SEO, paid ads, and social content on autopilot. It rewrites the pieces that underperform.

• Arcads launched Mark, an autonomous ad agent that takes a website URL, scrapes it for brand context, and generates ad creative and AI UGC video.

• Not Diamond Code released a model router for long-horizon coding agents.

• Superblocks 3.0 is a way for employees to securely vibe code enterprise software.

Subscribe now

Deep Dive

Claude Code 5.

The instructions you wrote for last year’s model are now the thing holding this year’s model back.

If anything, the people carrying the most dead weight are the power users, like readers of this newsletter who have implemented elaborate operating systems with memory.

I’ve tested everything Anthropic has said about the new 5 series models, including this awesome talk from creator Boris Cherny:

Today I’m sharing what you need to know:

1. The first principles

2. The 3 transformations to make

3. The skill that does it all for you

1. The first principles

It’s worth having a first principles understanding of why any of this matters.

For the most part, the rules in your CLAUDE.md, your skills, your graphs…

But the model gets smarter with every release. So the question really is: what patches can I get rid of with the new 5 series of models?

2. The 3 transformations to make

So can you just go and delete 80% of your system prompt? Not exactly.

Transformation 1 - Rebuild your customizations from the ground up

Claude Code has a safe mode that disables customizations.

This is a great way to baseline behavior.

Run your most common task in safe mode and compare it to what you’d get with everything loaded (without safe mode).

Here, for instance, I tasked the one-page PRD writer:

Here you can see, the safe mode PRD writer actually had a smarter recognition of the existing PRD.

So, in this case, start with the safe mode version and then add the customizations you need to be back in.

Other times, the customizations output will be better. In those cases, anything the raw model got right on its own is an instruction to delete.

Transformation 2 - Lighten your CLAUDE.md into skills

Most people’s CLAUDE.md is flooding with junk context which is stale from limitations of the older models.

All of it gets read after you give it a goal and that is the huge mistake.

run: /doctor

The checkup will find actions that give the model less irrelevant context, giving you a clearer place to maintain each workflow.

You can then move the information in CLAUDE.md out of there and into specific skills.

Transformation 3 - Define the outcome, not the steps

Most of us write prompts defining the procedure. With the 5 series of models, it’s better to define the task, guardrails, and completion criteria, then give the model room to decide how to proceed.

For instance I ran these two design prompts head to head - one procedural, one outcome based:

In the end, I judged the output focused on the outcome as better:

And this happened for everything I tested. With the 5 series of models, you can focus more on the outcome for all types of tasks.

3. The skill that does it all for you

Steal my skill here that will run a graph to do all three transformations for you:

Download the Skill

And don’t forget to remove the skill when you’re done to avoid skill bloat!

Final Words

We keep adding to our AI customizations. Hopefully this was your sign to do the occasional subtraction step.

Implement the changes, and reply with your feedback! I read every reply.

Subscribe Now

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort. This is shaping up to be the last one. It starts in 7 days.

Welcome to another AI by Aakash. I follow the AI news so you don’t have to.

We were literally talking about Loops as “the new prompts” two weeks ago. Then Peter Steinberger went and mentioned another “new thing” - Graphs.

So do we all need to learn graphs now? That is today’s deep dive.

But first, a word from our sponsor, and the week’s news.

In Partnership with

CodeRabbit’s Change Stack: Built for how reviews actually work

AI writes more code today than it ever has. Reviewing it shouldn’t mean scrolling through 40 files in alphabetical order.

CodeRabbit’s Change Stack organizes any pull request from a flat file to a structured overview. It reads the changes logically and then gives each range its own simple summary along with sequence diagrams, state machines, and ERDs wherever needed.

Cohorts group similar files to review one idea at a time. Layers order the elements so foundational changes come first. Code Peek allows you view definitional and usage seamlessly and Semantic Diff view cuts across formatting noise to show the actual changes. Reviews and approvals then post to GitHub or GitLab.

CodeRabbit’s Change Stack is free during early access. It’s from the team that pioneered AI code reviews - 2M PRs reviewed weekly, 6M repositories installed, and 15,000+ customers.

Try CodeRabbit on your next PR today →

Try CodeRabbit

This Week’s AI News

Top News This Week - Claude Opus 5

Model releases never stop. This week it was Anthropic’s Claude Opus 5 . It gets close to the intelligence of Fable 5, their most powerful model, at half the price ($10 input, $50 output / M token vs $5/$25).

The benchmarks are genuinely impressive, even above Fable in places:

I’ve been testing it every day since it came out. Here’s where my vibe checks land:

I don’t think it deserves the Opus name. It’s not that good.

Yet, I still use it a lot! That’s because Fable is limited to 50% of my usage in the Max plan and costs 2x as much in overage.

My most common usage pattern is to use a Claude Code Fable orchestrator agent spin up Sonnet 5 and Opus 5 sub-agents in workflows as needed.

The Other News That Mattered

• Anthropic’s revenue has passed $70B ARR per David Sacks, and will end the year at $100-$150B per Dwarkesh Patel. It’s on track to complete the 10x another year of its existence, just as I predicted.

• Most of the top AI companies signed a letter asking Washington to protect open models, after news had leaked the White House was considering banning Chinese AI models.

• ChatGPT Voice is now in the desktop app, and Anthropic upgraded Claude’s voice mode the same day. As Karpathy said, long ramble sessions work well with agents.

  

Tools

• Jack Dorsey and Block released Buzz, an open source, agent-native chat app.

• Andrew Ng released OpenWorker, an open-source desktop agent. It runs local-first so your files never leave your machine, and it hands back finished docs instead of chat replies.

Funding

• Moonshot, the creators of Kimi K3, secured a $35B valuation, with $3.5B raised, the same week they released all 2.8 trillion parameters of Kimi K3.

• Atoms raised $1.7B for Travis Kalanick’s (the Uber guy) industrial robotics startup, led by a16z.

Subscribe now

Deep Dive

Complete Guide: Graphs

Graphs are all the hype on X right now after Peter from OpenAI talked about it.

Beyond the hype, Graph Engineering is a very powerful system to build multi-agent systems that already existed. Let me explain it to you in the simplest terms.

I’ve broken today’s deep dive into three pieces:

1. What is a graph and how did we end up here

2. Graph patterns and when to use them

3. Graph construction in practice

1. What is a Graph and how did we get here?

You can think of the 5 hype cycles we’ve seen the last 3 years as actually building upon eachother:

Prompt → Context → Harness→ Loop→ Graph

• Prompt engineering is what the model should do.

• Context engineering is what the model should know.

• Harness engineering is about what the Ai can access or change.

• Loop engineering is how the agent keeps improving.

• Graph engineering is how the whole system coordinates.

Most people still use AI like a single employee. You ask a question, the model thinks and responds. For simple tasks that’s enough, but real work is rarely that straightforward.

In a graph:

Each node performs one specific responsibility: one researches, another validates, another writes, another decides which path to take next.

Every node has its own independent output.

Graphs have always existed. They are just a better way of designing systems. Instead of asking one AI to do everything, you’re building a coordinated team where every member has a clear responsibility.

2. Graph Patterns for your agents and when to use them

So how should you design a system? The key is to pattern match.

There’s tons of types of graphs:

1. Sequential: Research → Analyze → Draft → Review → Deliver

2. Router: Request → Understand → Billing / Support / Product / Human

3. Parallel: Plan → Several independent researchers → Synthesis

4. Orchestrator: Goal → Planner → Tools → Report

5. Review loop: Create → Review → Approve, or revise and check again

6. Evaluator: Draft → Evaluate → Pass? → Approved

7. Diamond: Plan → Parallel exploration → Independent verification → Synthesis

You want to pattern match when to use which. The flowchart looks like this:

Most are exactly like they sound, with the exception of a diamond.

But you’ll get it as soon as you look at the visual:

You essentially have one step with parallel agents. Once all the evidence has been verified and gathered, you bring it back to sequential.

The point is: don’t just rely on sequential. Parallelize where it makes sense. Loop where that does.

It’s that simple.

3. Graph construction in practice

Now let’s see what this looks like in practice.

The easiest way to use Graphs is to build them with Claude Code workflows. It will even do the work of selecting the right pattern for you. Start with a prompt like this:

The keys to this prompt are:

1. Define the outcome you want

2. Explain the graph types you want it to use (combos of the types of above)

3. Identify the models you want to use for each task (you can see why I use Opus so much now!)

Then you can see it will go and build the prompts and do deep work, setting up parallel workflows where possible.

You can use /workflows to manage your graphs:

We basically defined a double diamond graph:

You can then monitor specific agents as they work. For instance, here’s the discover analytics agent:

Eventually, you get an output that is the result of many agents working for you, instead of a single agent.

It’s that simple!

If you’ve been reading lots of my content, we’ve been building graphs for a while. Now, you have a name for it.

Want to go deeper? I have a deep dive coming out in Product Growth soon.

Me Around the Web

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort. Next one starts in August.

Share

One year ago, I would’ve told you Claude Code is the only way to go. That’s changed with the latest GPT-5.6-sol model being genuinely better than Fable on many tasks. If you want to get the most sol, you need to understand the new ChatGPT Work harness. So that’s today’s deep dive. Plus, this week’s AI news.

But first, a word from our sponsor.

In Partnership with

CodeRabbit’s Change Stack: Built for how reviews actually work

AI writes more code today than it ever has. Reviewing it shouldn’t mean scrolling through 40 files in alphabetical order.

CodeRabbit’s Change Stack converts your pull request into a structured and sequential walkthrough that follows how the change reads. It gives every range a simple summary, sequence diagrams, state machines and ERDs inline where helpful.

Cohorts group related files and chunks so you review one idea at a time while Layers put foundational changes such as data shapes and contracts first. Code Peek shows a symbol’s definition and usage without leaving the tab and Semantic Diff goes past formatting noise to show what changed. Reviews and approvals post back to GitHub or GitLab.

CodeRabbit’s Change Stack is free during early access. It’s from the team that pioneered AI code reviews: 2M PRs reviewed weekly, 6M repositories installed, and 15,000+ customers.

Try CodeRabbit on your next PR today →

Try CodeRabbit

This Week’s AI News

There’s a billion AI news articles every week. Here’s what actually mattered:

Moonshot’s new model beats Fable 5 for frontend

The Chinese AI lab released Kimi K3 this week, a 2.8 trillion parameter open-weight model with native vision and a 1 million token context window.

Moonshot says it only trails Claude Fable 5 and GPT 5.6 Sol, which means it beats every other model they tested.

Your frontend generation bill is about to drop. K3’s current API is $3 in, $15 out per million tokens, about half of Opus 4.8. And because the full weights go public soon (July 27), every inference provider will be serving it within weeks, which puts price pressure on the closed labs even if you never switch.

Lindy shifted their entire company to DeepSeek models and called it:

By far the one thing we've done in the history of the business that's made the biggest difference to our economics, with no impact to product quality

Their CEO, Flo Crivello, will be on the podcast in a few weeks to talk about it.

The Other News That Mattered

• AI is now making breakthroughs in mathematics. Claude Fable 5 produced a one-line formula that breaks the Jacobian conjecture. Mathematicians have been stuck on this since 1939. Fable knocked it over with a single prompt. Even Terence Tao had to ask ChatGPT about it.

• Anthropic reversed Fable 5 pricing two days after free access expired on July 19. Fable 5 now lives permanently for all paid users. (Yay!)

• Claude can now learn skills by watching you. Record your screen doing a task, narrate as you go, and it becomes a skill Claude runs again later. It’s under “Record a skill” in the desktop app’s + menu:

  

• Google is reportedly developing Frozen, a server chip that bakes Gemini’s software blueprint directly into hardware so its models run more efficiently.

Resources

• Baidu released a 3B open-source OCR model. It can read 100-page PDFs in one pass. Plus, it’s small enough to run on your laptop. Definitely interesting if you’re paying per page for document parsing in your product.

Tools

• Grok for Excel writes your formulas from plain English, inside the spreadsheet.

• Canva Code is a vibe coding platform with drag-and-drop editing on the output.

Share

Deep Dive

Complete Guide: ChatGPT Work

ChatGPT Work is 10x more powerful.

It puts together three really powerful features you don’t get otherwise: connected tools, scheduled tasks, and computer use.

Let’s walk through why these matter and what they enable.

Today’s Outline

I’ve broken today’s deep dive into three pieces:

1. Setting up the optimal ChatGPT Work harness

2. The Top 6 Use Cases of ChatGPT Work

3. How to avoid creating “AI slop”

Subscribe now

1. Setting up the optimal ChatGPT Work harness

There’s four steps involved.

Step 1 - Download

Start with the basics. Download the app from here, and select ChatGPT:

Step 2 - Connect everything

Install these plugins, at a minimum:

• Computer use

• Spreadsheets

• Chrome

• Presentations

• Notion

• Linear

• Lovable

• Figma

• Attio

(or whatever version of that category of software you use.)

If you’re comfortable, give it the complete permission to the whole device:

Everything starts from a connected base, similar to the MCP connections I’ve mentioned for an Operating System in Claude Code.

Step 3 - Turn existing meetings into a context library

One particularly important thing to do as you use it is convert meetings to context.

Do it like this:

This is where Work really proves its mettle. Over time, you build an entire context library around your work. That’s how you create a compounding asset.

Step 4 - Match your existing templates

Build templates out of your repeatable tasks.

For example - NDAs:

PRDs, whatever you have.

2. Top 6 Use cases

So, you’re set up. Now let’s rifle through the use cases for it (that go beyond regular ChatGPT in the web).

Use Case 1: Your Data Analyst

Work, when connected, can do your data analysis. For example, analyze conversion drops with your CRM (Hubspot):

Instead of choosing to make sense of the numbers on your own, you can also use the visualize tool to do the data visualization for you:

Use Case 2: Schedule routine tasks

Then schedule out weekly analytics reads, weekly business reviews, and the like:

Don’t stop at analytics. Think about every task you see yourself repeating and ask:

Use Case 3: Have it do browser work for you

With Work, you can connect with your Chrome browser:

The built-in browser gives the user and ChatGPT a shared view of the page. Work can then act on it like you do.

Use Case 4: Leverage Computer Use to access local files

It can open local files and drive desktop apps. You can even have it trigger music for you:

Use Case 5: Get your hardest work done

Finally, put everything altogether for complex tasks.

For example, PMs have to create prototypes. You can reference sites and create them:

Use Case 6: Run cloud tasks to iterate on the go

As you’re doing these complex, long running tasks, I recommend you kick off cloud tasks directly from the desktop or mobile apps:

Then you can do things like start on computer and edit on mobile while you’re out. I use it constantly.

3. Don’t create AI slop

A final word.

You have to avoid that. Just like this post is 100% hand-written:

Here’s how to get there yourself:

1. Use Work for the middle 60%: apply your own human judgment in the beginning 20% of a project, and in the final 20% in order to rewrite in your voice

2. Don’t use it as your ghostwriter: the temptation with AI is to let it be your writer; but, people are quick to see through it

3. Use it as your APM: you wouldn’t put an APM’s doc in front of a VP, but you would get them to do the hard work for you

These are simple platitudes, but they’re very powerful.

Final Words on ChatGPT Work

I don’t think ChatGPT work is better than Claude Cowork yet, but I think a lot of people are probably using it. Given how good GPT-5.6-sol is, it’s not a bad choice.

I used to always say you should use Codex for ChatGPT.

With the ChatGPT Work launch, I finally think it’s now the ChatGPT app.

Me Around the Web

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

This week, the big news is out of OpenAI. GPT-5.6 feels like, for the first time in a long time, OpenAI has come out with a better model than Anthropic. And it shipped with a new harness too.

So that’s today’s deep dive. But first, a word from our sponsor.

In Partnership with

Viktor: 5x More Work Done

Founders keep telling the Viktor team the same thing. One: "If you compare this to a VA or an employee it's not even a conversation. This is doing what an employee would do and it's doing it much much faster, better and much cheaper." Another paying team: "We can get 5x as much work done now."

Viktor is an AI employee that lives in Slack and Microsoft Teams. It connects to 3,200+ tools you already use, runs scheduled tasks on its own, and hands back real deliverables: PDFs, decks, internal apps, code PRs. You stay in control, anything sensitive waits for your approval. 40,000+ teams have already hired one.

Try Viktor free, $100 in starting credits →

Try Viktor

This Week’s AI News

Top News: We Have New Models

ChatGPT launched 5.6, which I cover in depth in today’s deep dive.

But we also saw new models from Tinker, Meta, and xAI this week.

Basically everyone besides OpenAI and Anthropic have stopped fighting over who has the smartest chatbot and started fighting over who can run your work cheapest:

• Grok 4.5 came in at $2/$6 per million tokens.

• Muse Spark 1.1 ended Meta’s open-weights era and undercut OpenAI and Anthropic by about 75% on price.

• The most interesting one might be Inkling, Mira Murati’s first from-scratch model. It’s deliberately not the strongest model out there. It’s an open base you fine-tune, the same setup Bridgewater used to build a custom Qwen that beat frontier models on domain-specific tasks.

None of these models need your attention as a consumer.

But if you’re building a product, it’s worth running the cheap ones through your evals. The price gaps are now big enough to matter.

The Other News That Mattered

• Apple is suing OpenAI and two former employees over alleged trade-secret theft for its upcoming AI hardware. Apple says candidates were told to share confidential information and even bring physical parts to interviews, and OpenAI says it has no interest in others’ secrets.

• Claude Code got an in-app browser plus a big Artifacts upgrade in the same week. Claude can now open and click through docs, designs, or your own app while it works, and Artifacts support public links and multiplayer editing, including pages made through Claude Tag in Slack. Anthropic is quietly closing the gap on the one thing OpenAI’s new app does best, which is acting on what it sees.

Resources

• Demis Hassabis wrote his framework for the AGI age, a long essay arguing AGI is probably only a few short years away and that we’re standing in the foothills of the singularity. The Nobel laureate running DeepMind is the closest thing AI has to a measured CEO voice, so the essay deserves your 15 minutes.

• OpenAI published its GPT-5.6 prompting guide, and the short version is that your old prompts are too long. Their own tests found leaner prompts scored 10 to 15 percent better on 41 to 66 percent fewer tokens. Describe the destination, stop prescribing the route.

Tools

• v0 Design Systems 2.0 lets you import your components from GitHub, npm, Figma, and Storybook so generated apps use your real design system instead of the generic default. Your prototypes finally look like your product.

Share

Deep Dive

Complete Guide: GPT-5.6

ChatGPT just had its biggest update in a year. On July 9, OpenAI shipped the GPT-5.6 models and merged ChatGPT with Codex into one app.

Combined with the leading ImageGen model, I feel like OpenAI has the best $20/month plan in AI right now. You should be taking advantage of it.

So here’s how to get the most out of ChatGPT-5.6:

1. What Actually Changed

2. The Sol, Terra, and Luna Models

3. How to Pick Your Thinking Level

4. How to Not Burn Your Limits

5. Prompt Less, Get More

Subscribe Now

1. What Actually Changed

OpenAI took two apps and made them one, with three tabs inside:

• Chat is the ChatGPT you already know.

• Codex is the coding agent, and you can keep ignoring it if you don’t code.

• The new tab is ChatGPT Work:

Work is the delegate-and-walk-away mode.

You give it a goal, it reads your files and connected apps, works for hours if it needs to, and comes back with the finished document, spreadsheet, or presentation.

Work and Codex are the same agent tuned for different jobs, which is why the two tabs look alike.

It’s on every plan, including Free, and if you read my guide to Claude Cowork, this is OpenAI’s answer, shipped to hundreds of millions of people at once.

Three more changes worth 30 seconds each:

1. OpenAI killed its Atlas browser and moved the good parts into a new ChatGPT Chrome extension, plus a browser inside the app that can visit sites, log into your accounts, and download files.

2. ChatGPT Sites turns a prompt into a live website, hosted for you, with an optional Login with ChatGPT gate for visitors. Paid plans only for now, and not yet in the EU or UK.

   

   

   

3. Computer Use got faster. The app can drive your actual computer now, opening apps and clicking buttons by looking at your screen. Hold that thought until section 7, because you should watch it happen once.

2. Sol, Terra, and Luna

GPT-5.6 comes as three models:

Sol is the senior hire. It’s like Fable. Use it for complex work and long tasks.

Terra is the mid-tier. It’s like Sonnet. It matches the old GPT-5.5 on everyday work at half the price, and it does what you asked, exactly.

Luna is the intern. It’s like Haiku.

3. Pick Your Thinking Level and Skip Ultra

Thinking levels are how long the model gets to sit with your problem before answering, from a quick glance at Light to a long stare at Max.

Then there’s Ultra, which sits in the thinking menu but behaves like a different product.

Ultra hires four copies of the model at once and splits your task between them. Every copy runs at the most expensive thinking setting, the copies can hire copies of their own, and there’s currently no dial to turn any of them down. Each one bills you separately.

Theo blew his entire 5-hour allowance in 20 minutes on a single Ultra run with fast mode on, got a manual reset, and drained that one in another 40. OpenAI has since hidden Ultra from the app’s main model slider, which tells you how they think it’s going.

Skip it.

4. How to Not Burn Your Limits

Three rules until the dust settles:

1. Stay off Ultra. You just read why.

2. Skip fast mode for now. It delivers the same answer sooner and bills you a multiple of normal usage for the privilege, and GPT-5.6 already runs marathon sessions on its own. Here's what the same tasks cost me with it off and on:

   

3. Let persistence replace thinking level. The old habit was buying quality with higher levels because models quit early. These don’t quit. Medium plus a model that finishes the job beats High on a model that needed babysitting.

5. Prompt Less, Get More

Forget the models for a minute. The most useful thing OpenAI shipped last week was a short guide on how to talk to the new ones, and its core advice fits in a sentence:

We all learned to write long, detailed prompts because the old models gave up without them. The new ones figure out the steps, so your extra instructions now get in the way.

There is one thing these models need more of, and that’s brakes. GPT-5.6 keeps going. Left alone it does the task, then the follow-up you never asked for. The fix is a stop point, written right into the ask:

Here is the complete history: [PASTE THE FULL EMAIL/SLACK THREAD(S) — do not trim].
Reconstruct: every decision made (with who made it and the message that proves it), every commitment with owner and date, decisions that were later reversed or contradicted, and open questions that everyone forgot. Chronological table. Where the record is ambiguous, say ambiguous — do not fill gaps with plausible fiction.

That’s the whole formula. The goal, then exactly where to stop and show you. Use it on anything longer than a question.

Keep in every prompt:

1. The outcome you want, stated first.

2. What “done” looks like.

3. Your hard limits, like budget, tone, and what it must never do.

4. A stop point where it pauses and shows you.

Cut from every prompt:

1. The same rule stated twice. Once is enough now.

2. Step-by-step directions for things it already knows how to do.

3. “Be concise.” It defaults shorter, so say what to keep and what to drop instead.

4. Vague personality requests like “be friendly.” Describe the behavior you want.

5. Rules that contradict each other. It wastes effort trying to obey both.

Me around the web

I joined Rupinder Singh's podcast for a full Claude Code masterclass covering how PMs should actually work with AI in 2026:

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

And today’s deep dive is the Fable 5 playbook: how to spend the free window on the work that actually deserves it, before it flips to paid credits.

But first, a word from our sponsor.

In Partnership with

Viktor: One Hire. Already Knows Your Stack.

Picture the new hire who shows up Monday already fluent in Stripe, HubSpot, Shopify, Linear, Notion, GitHub, Klaviyo, Google Ads, Meta, QuickBooks, and 3,000+ other tools.

They never ask where the forecast lives. They onboard in minutes.

Viktor is that hire. It’s an AI employee for Slack and Teams that ships real deliverables, not summaries. Marketing gets the Klaviyo audit and the campaign brief by Tuesday. Finance gets a cash flow model that updates straight from QuickBooks actuals. Ops gets the weekly manager report. Engineering gets reviewed PRs.

One install, every department benefits, every action approved by a human first. “Like adding another employee. Or two.” That’s Krysta, who runs 63 Viktor workflows at her company. 11,000+ teams now use Viktor.

Try free, $100 in credits.

Try Viktor

AI News

The Week’s Top News: Fable 5 Is Back

Anthropic just reopened access to Fable 5 after the Commerce Department lifted its export controls. The model returns under tighter filters, plus a commitment giving the U.S. government pre-release access to Anthropic’s future models.

A quick rewind on how we got here. The original order traced back to Amazon researchers who pushed past Fable’s guardrails to expose security flaws in its outputs. Anthropic argued other frontier models matched those same outputs, but Fable was the one that took the hit. The Commerce Department pulled it anyway, and for 18 days the most capable model on the market was simply unavailable to the people paying for it.

That absence mattered more than most people expected. Fable 5 had been out for barely two weeks before the ban, and people were already shipping real projects with it. Teams that had rebuilt their workflows around it got a live lesson in platform risk: your stack can lose its brain overnight, and not because of anything you did.

Here’s what the return actually looks like:

• Fable 5 is back across Claude tiers and platforms. Paid plans get it capped at half their weekly limits until July 7, then it moves to usage credits.

• The updated safety filter now blocks the specific cybersecurity issue over 99% of the time. When it triggers, you get a clear notice and a fallback answer from Opus 4.8 instead of a silent failure.

• Anthropic warned the filter could also flag harmless coding and debugging requests, though it says “the vast majority of coding work is unaffected.”

• The pre-release access deal means the U.S. government sees Anthropic’s future models before you do. That’s a first for a frontier lab, and it won’t be the last.

Here’s where I land: the model coming back is the small story. The precedent is the big one. A government pulled a frontier model off the market, extracted pre-release access as the price of return, and the lab took the deal. Every other lab watched that happen. With GPT 5.6 expected this week, we’re about to see whether OpenAI’s friendlier relationship with Washington buys it a smoother path. My bet is that “regulatory relationship” just became a line item in every AI vendor evaluation, right next to price and benchmarks.

The Other News That Mattered

• Anthropic launched Claude Science, a research app with 60+ connectable scientific databases, citation checking, and compute that runs on the lab’s own machine. That last part is the unlock: hospitals and biotechs have been blocked from AI tools by patient data rules, and Manifold Bio is already using it to rank drug candidates. Anthropic is also offering 50 grants of $30,000 each, open globally, due in two weeks.

• Anthropic also released Claude Sonnet 5, its most agentic Sonnet yet. It hits 63.2% on agentic coding versus Opus 4.8’s 69.2%, at $2 per million input tokens, and it’s now the default for free accounts. Four months ago this performance was paywalled in the most expensive model. If your product’s moat is access to expensive intelligence, that moat is a lease that keeps resetting.

• Sam Altman used an FT op-ed to call for a U.S.-led forum that sets AI safety standards and decides who can use the most advanced models, citing the IAEA as precedent. It landed just as another FT report said OpenAI discussed giving the U.S. government a 5% stake. The real question, as former White House AI advisor Dean W. Ball put it, is whether AI wealth goes directly to households or to a government that may or may not deliver.

• Palantir CEO Alex Karp criticized frontier AI labs on CNBC, saying enterprises “are paying for tokens that create no value” and the labs “are stealing weights and alpha.” Spicy, but remember he’s selling the antidote.

• Microsoft launched Frontier Company, a $2.5B unit putting 6,000 in-house engineers and sector specialists at client sites to build and run AI systems. Consulting is quietly becoming the distribution channel for enterprise AI.

Resources

• How to build a continuous eval pipeline for multi-agent systems with Gemini. This tutorial moves you from subjective testing to data-driven assessment: automated model-based grading instead of manual testing, handling unpredictable AI outputs, and wiring evaluation directly into a CI/CD pipeline with Cloud Run. If your agent testing is still “run it and eyeball the output,” start here.

• Are we ready for an agent-native memory system? Current benchmarks treat agent memory like a black box, ignoring backend data management trade-offs and operational costs. The study’s finding: no single architecture fits all scenarios. Success depends on matching the memory structure to your specific workload bottleneck. Worth reading before you commit to a memory stack.

• Google’s NotebookLM rolled out Short Video Overviews, letting you create 60-second, social-media-style educational videos from any topic or source.

Tools

• Seedance 2.5 is ByteDance’s newest frontier AI video model, generating native 30-second clips with no stitching required.

• Cursor for iOS brings Cursor to your phone. Kick off agents on the train, review the diff when you sit down.

Quick Guide: Set Up Claude Tag in Slack

Claude Tag lets you and your teammates tag @Claude like a real teammate, then have Claude work in the cloud and report back when it’s done. Here’s how to set it up:

1. Get Claude Team or Enterprise plus Slack admin permissions. In Slack, open Admin, then Apps and Workflows, search for Claude, and install it.

2. Find Claude under Apps, open it, click Home, and connect your account. Then open Claude Tag settings on the web and connect the tools it should use.

3. Run @Claude connect in Slack, copy the pairing code into Claude Tag’s browser setup, choose the channel scope, add usage credits, and launch.

4. Open Advanced in Claude Tag settings and switch from the expensive default model to Sonnet. Then run a test prompt in a dedicated Slack thread.

Pro tip: after the first run, check the admin panel for token spend, plugins used, and memory files created. You can edit or delete persistent Claude Tag memory files there.

Fundraising

• China’s Kling AI secured $2B in funding. The Kuaishou video spinoff is pushing global expansion after OpenAI shut down rival Sora. The most interesting AI video race right now isn’t happening in San Francisco.

Share

Deep Dive

The Fable 5 Playbook

Fable 5 is back, and most people are about to waste it.

Here’s the trap: you open it, run the same prompts you used on Opus 4.8, get slightly better answers, and conclude it’s not worth the fuss. That conclusion is wrong, but it’s the natural result of testing a new category of tool on the old category of task.

Fable 5 is built for work that takes a person hours, days, or weeks: full migrations, multi-day autonomous runs, problems you stopped trying to automate. Stripe pointed it at a 50-million-line Ruby migration and compressed two months of estimated team effort into a single day. That’s the category of task this model exists for.

And right now there’s a deadline attached. Through July 7, Pro, Max, Team, and select Enterprise plans get Fable 5 included for up to 50% of their weekly usage limit. After that, it moves to usage credits and the same big job costs real money. You have a short window where the most capable model available is effectively free, up to half your cap.

This deep dive is how to spend that window well. Here’s what I’m covering:

1. The one rule that decides when to use it

2. The routing config (copy-paste)

3. The big-job kickoff template (copy-paste)

4. The five prompt patterns that actually change output quality

5. Loops and the barbell strategy

6. The memory system

7. Limitations

Subscribe now

1. The one rule that decides when to use it

Weeks by hand goes to Fable 5. Minutes by hand goes to Sonnet 5.

That’s the whole routing decision. If a task would take you two weeks or more manually, it earns Fable 5 and a slice of your free window. If it’s a quick edit or a fast question, it stays on Sonnet 5, which now handles most everyday work at a fraction of the cost.

The mistake I keep seeing is people treating Fable 5 like a chat model. Asking it quick questions all day burns your included capacity on work Sonnet handles fine, and by Thursday you have nothing left for the job that actually needed the heavy machinery.

A codebase-wide refactor. A dependency migration you’ve been putting off for a quarter. A gnarly race condition nobody wants to touch. Those abandoned tasks are where Fable 5 stops looking like an upgrade and starts looking like a different product.

2. The routing config

Drop this into your CLAUDE.md so the model choice happens automatically instead of by mood:

## Model routing (through July 7)
Default: Sonnet 5 for everyday work.

Use Fable 5 only for heavy, high-payoff tasks:
- large migrations (framework, language, dependency)
- codebase-wide refactors across many files
- complex multi-step builds
- hard bugs in tangled code (race conditions, subtle state)

Rule of thumb: weeks-by-hand = Fable 5, minutes-by-hand = Sonnet 5.
Protect the Fable 5 window. Do not spend it on small edits.

# Some security-adjacent requests get rerouted to Opus 4.8
# by the new safeguards. If quality drops on one call, check for a reroute.

That last comment matters. Fable 5 came back with a new classifier that blocks one specific jailbreak technique over 99% of the time. When it triggers, your request reroutes to Opus 4.8 and you get a notice. The trade-off is false positives: legitimate coding and debugging work can occasionally get rerouted even when your intent is fine. Anthropic says the vast majority of coding work is unaffected, but if output quality suddenly drops on a single request, check whether you got rerouted before you blame your prompt.

One more thing worth knowing: Fable 5 carries a 30-day data retention policy, used to research jailbreak mitigations. Factor that in before you feed it anything sensitive.

3. The big-job kickoff template

When you point Fable 5 at a large migration or refactor, don’t say “refactor this.” Give it the whole job up front so it can hold the plan end to end. Copy and fill in:

# Task: [migration / refactor name]

## Goal
[One sentence: the end state you want.]

## Scope
- Files/dirs in scope: [paths]
- Out of scope: [paths to leave alone]

## Constraints
- Keep all tests passing.
- Preserve public APIs unless listed below.
- Match existing code style.

## Plan first
Before changing anything, map the full plan and list the files
you will touch. Wait for my go-ahead, then execute in one pass.

## Definition of done
- [ ] [check 1]
- [ ] [check 2]
- [ ] all tests green

The “plan first, then execute in one pass” line is the important part. It’s the difference between Fable 5 as a fast typist and Fable 5 as the thing that cleared a 50-million-line migration in a day. The model plans before it acts, checks its own work, and doesn’t rush to a quick output you’ll immediately have to fix.

Which means responses take longer. At high effort, a single response can take minutes. On autonomous runs, hours. The model is doing the work correctly, and correct work at this scale takes minutes. If you’re calling it through the API, extend your timeouts before you do anything else, because a timeout tuned for a ten-second Opus reply will kill a Fable run mid-plan.

4. The five prompt patterns that actually change output quality

I tested the patterns from Anthropic’s official prompting guidance against my own workflows. These five made a measurable difference. Everything else was marginal.

Pattern 1: Make it audit its own progress reports. On long runs, models can report a step as complete when it hasn’t been verified against anything real. One instruction nearly eliminates this:

Before reporting progress, audit each claim against a tool result
from this session. Only report work you can point to evidence for.
If something is not yet verified, say so explicitly. If tests fail,
say so with the output. If a step was skipped, state that.

The word “audit” is doing specific work. Softer phrasings don’t produce the same self-checking. For anything running longer than a few minutes, this goes in your system prompt, not your task prompt.

Pattern 2: Set the boundary between observing and acting. Fable 5 is noticeably more proactive than Opus 4.8. It will sometimes draft an email you didn’t ask for or create a backup branch you didn’t request. Helpful in a chat. Dangerous in an unattended pipeline. The fix:

When I am describing a problem or thinking out loud rather than
requesting a change, the deliverable is your assessment. Report
your findings and stop. Do not apply a fix until asked. Before
running anything that changes system state, confirm the evidence
supports that specific action.

Pattern 3: Give it the why, not just the what. Fable 5 uses intent context to make better judgment calls on its own. Frame requests as: “I’m building [X] for [audience]. They need [outcome]. Given that, here’s what I want you to do.” This changed my output quality more than any other single habit.

Pattern 4: Pick effort levels on purpose. High is the default and right for most real work. Xhigh is for when first-shot correctness matters more than speed, because the model reflects on and validates its own work before responding. A codebase migration warrants xhigh. A formatting pass does not. And here’s the counterintuitive part: low-effort Fable 5 often beats max-effort older models, so don’t assume you need the top dial for everything. In Claude Code, “ultrathink” in a prompt gets you max reasoning for that one turn.

Pattern 5: Rewrite your old skills instead of porting them. Skills built for weaker models tend to be prescriptive, step-by-step instructions that assume the model needs hand-holding. That same prescriptiveness can box Fable 5 in and make output worse. Go through each skill and ask: is this teaching the model something it doesn’t know, or micromanaging a model smarter than the instructions assume? Describe outcomes. Cut the steps.

5. Loops and the barbell strategy

Fable 5 was built for autonomous work, and Claude Code gives you two commands to run it that way.

/goal launches a task that runs until it’s actually done:

/goal keep researching until you can answer these 5 questions

/loop runs on an interval until you cancel it:

/loop every 30 minutes, flag any email that actually needs me

Instead of you sitting between every prompt and every fix, you design the loop once and the model handles the back-and-forth, surfacing only when it hits your success criteria.

The cost problem with long runs is real, though, and the fix is a barbell:

• Planning the loop (first 10%): Fable 5. Frontier-level planning, spec writing, loop design.

• Gruntwork and execution (middle 80%): Sonnet and Haiku subagents. Cheap, fast, repeatable.

• Final verification (last 10%): Fable 5 again. Verify against the original spec, catch errors, confirm done.

You get Fable-level judgment at the two moments it matters most, planning and verification, without paying Fable-level prices for routine execution. One run of a real research loop this way costs a few dollars in subagent tokens instead of burning your Fable window on gruntwork.

6. The memory system

None of the above hits its ceiling without persistent context. Fable 5 with a memory file is a different tool from Fable 5 without one. Anthropic’s own testing found that a persistent notes file improved Fable 5’s performance on long-horizon tasks roughly three times more than the same setup helped Opus 4.8.

The setup takes ten minutes:

1. Create a local folder like /claude-context. This is your AI brain. Put in a map of your business or project, SOPs for anything you do repeatedly, one-pagers on key projects, and a running log of decisions and outcomes.

2. Create claude-memory.md inside it, with one standing instruction: “Every time I give you major context about my business or situation, update claude-memory.md with the key details.” Now it self-updates.

3. Create claude-instructions.md with the behavior rules: how memory gets stored, output formatting, what to always and never do.

4. Connect it by pointing Claude Code at the folder via /add or referencing it in your CLAUDE.md.

For the memory file itself, use this structure:

Keep one lesson per file with a one-line summary at the top.
Record corrections and confirmed approaches alike, including why
they mattered. Do not save what the repo or chat history already
records. Update existing notes instead of duplicating. Delete
notes that turn out to be wrong.

At the end of significant sessions, close with: “Reflect on the sessions we have had. Identify core themes and lessons, and store them in the folder.” Every future session then starts with the model already knowing your world. That’s the actual mechanism behind multi-day coherence, and you own the whole thing. It lives in your local folder, portable to any model that comes next.

7. Limitations

I want to be straight about where this gets rough.

1. The reroutes are real. Security-adjacent coding work occasionally bounces to Opus 4.8 even when your intent is legitimate. You get notified, but mid-task it’s disruptive. (Workaround: check for the reroute notice before debugging your prompt, and keep the task on Opus if it keeps triggering.)

2. Latency will surprise you. Minutes per response at high effort, hours on autonomous runs. This is by design, but it makes Fable 5 wrong for anything interactive. (Workaround: fire long runs before you step away, not while you’re waiting at the keyboard.)

3. The window math punishes drift. Half your weekly limit sounds like a lot until you fritter it on medium-sized tasks. (Workaround: pick two or three genuinely heavy jobs for the week and route everything else to Sonnet, in writing, in your CLAUDE.md.)

4. One prompt gotcha: asking Fable 5 to narrate its reasoning step by step inside the visible answer can trip a classifier built to catch reasoning extraction, and you’ll get bounced to Opus on a completely benign task. If you want to see its thinking, use the structured thinking output. Don’t ask it to retype its thoughts into the reply.

Here is where I land

Everyone has access to Fable 5 now. What separates mediocre results from Stripe-level results is whether you rewrote how you talk to it.

Five-line prompts judged in ten seconds test a weaker version of this model than the one that exists. Full-context kickoffs, audit instructions, explicit boundaries, and a memory file test the real one.

You have until July 7 while it’s free. Pick the task you gave up on automating. Give it the full job. Run it at xhigh. The difference between that experience and the same prompt on Opus 4.8 is where you’ll decide for yourself whether this model is what I think it is.

That’s it for today’s deep dive.

Around the web

I joined Richa Bansal to talk about how PMs stay relevant in the AI era. We went deep on career strategy, and a few of my answers surprised even me when I heard them out loud. My five big points from the conversation:

1/ Use AI for productivity and strategy, not one or the other. Most PMs stop at automating their daily tasks. The bigger unlock is thinking strategically about where AI features belong in your product, even if you’re not on an “AI team” today. Waiting for the org to hand you an AI project is how you fall behind.

2/ Build synthetic experience if your job won’t give you the real thing. You don’t need permission to become an AI PM. Take Andrew Ng’s courses or work through Karpathy’s tutorials, then ship side projects with tools like Replit or Lovable. A shipped project beats a certificate every time with top-tier employers.

3/ Find your candidate market fit. Stop chasing generic PM roles. Niche down to the intersection of your unique experience and market demand. A specialized AI gaming PM is a far stronger position than a generalist PM, because recruiters can actually find you.

4/ Bring proof, not polish. Rehearsed talking points don’t win interviews anymore. Do your own user research, analyze the company’s competitors, and walk in with a point of view on their product’s problems before the first round. Proof of work is the new resume.

5/ Flip your weaknesses before the interviewer finds them. Job gap, industry switch, no domain experience: use AI to identify the top concerns an interviewer would have about your resume, then preemptively reframe each one. Reframe international experience as global perspective before the interviewer files it under risk.

Subscribe now

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

I pointed /goal at a research report and never once had to nudge it to keep going. It looped on its own, caught its own mistakes, checked the work against what I’d asked for, and handed back the right output. So that is today’s deep dive.

But first, a word from our sponsor + the week’s AI news.

In Partnership with

Perplexity Computer

I set up my new Mac Mini with Perplexity Computer and gave myself a challenge: build the tools I’ve always wished existed for my creator workflow. No code, no engineer. Just ask myself a question and let Computer build the answer.

One of the tools I built is a podcast guest research desk.

Every week before recording the Product Growth Podcast, I’d lose three-four hours researching the guest. So I asked myself: what do I find myself doing over and over that shouldn’t require me every time?

I described the answer out loud. Drop in a guest’s name, get back three episode angles, 15+ questions in my voice, and title options engineered against viral clip formulas. ~12K words of pre-tape intel. 45-90 seconds.

Then I said: “Computer, build me that.”

Saves me 4 hours a week. Took one prompt.

Ask yourself a question. Then tell Computer to build the answer →

Here's my build, yours to copy and rework:

Guestlab tool : Perplexity Computer

Your turn. Ask Computer to build yours:

Try here

This Week’s AI News

The Week's Top News: The Most Powerful AI on Earth Just Became Something Money Can't Buy

Anthropic launches Claude Fable 5 today, a Mythos-class model made safe for general use. It's state-of-the-art on nearly every benchmark, and the longer the task, the bigger its lead. The longer and harder the problem, the further ahead it pulls.

Anthropic shipped two versions of the same model. Fable 5 goes to the public with hard caps on cyber, bio, chemistry, and distillation. Mythos 5, the uncapped version, goes only to vetted partners: Project Glasswing cybersecurity orgs and select biology researchers, with mandatory 30-day data retention so Anthropic can watch what they do with it.

Here’s what that means in practice. A Fortune 500 with a nine-figure AI budget gets the exact same capped model as a college kid paying $20 a month. The money buys you nothing extra.

That breaks a 70-year pattern. Computing access has always scaled with money. Mainframe time, Bloomberg terminals, GPU quotas. Every tier in the history of software was a price tier, and whoever paid more got more. Anthropic just replaced price with clearance. The structure is borrowed straight from defense contracting: capability allocated by vetting, usage monitored, access revocable. Except here a private company runs the clearance system, on commercial customers, for a commercial product.

The Other News That Mattered

• Google and SpaceX AI announced compute contracts with Elon Musk’s SpaceX AI. SpaceX AI has gone from an annual ARR of $500M to $26B in just a few months, with the announcement of those two deals. No wonder its IPO popped today.

• Google DeepMind shipped Gemma 4 12B, an open multimodal model that runs on a 16GB laptop. Local, private, on-device AI keeps getting better, but people keep preferring to run frontier tokens.

• Nous Research launched Hermes Desktop, a native app for its open-source agent. The agent saves what it learns as a reusable skill, so it compounds instead of starting from zero every session.

• Meta rolled out Meta Business Agent globally across WhatsApp, Instagram, and Messenger. This is what Jensen Huang meant when he said Meta “uses AI best.” They focus on applications for their core businesses.

Tools

• Google Labs launched Dreambeans, an experiment that turns your Gmail, Calendar, and Photos into a finite daily feed of personalized stories. Pitched as the opposite of infinite scroll: it gives you 10 to 14, then stops.

• Ideogram open-sourced Ideogram 4.0 and Reve launched Reve 2.0, both betting that layout and post-generation control matter more than a better prompt. Ideogram took the top open-model spot and ships weights you can fine-tune.

Funding

• Suno raised over $400M at a $5.4B valuation, more than double its number from six months ago. Its first model built with the music industry ships in the coming months. The round closed while Universal and Sony are both still suing them, which is either confidence or a pressure test. Probably both.

Share

Deep Dive

You need to start using /goal

Here’s the thing nobody tells you about handing work to an AI agent.

There are two ways it fails you. The first is the one you know. It stops mid-task and waits. You type “keep going.” It does another turn and stops again. You type it again. An hour later you’ve said it fifteen times and you’re still the one moving the whole thing along. That's a very fast intern who needs constant supervision. You're still the one doing the real work.

The second failure is quieter and worse. The run completes. The agent says “done.” You trust it, you move on. Three days later something breaks and you find out half of it was a stub that never ran. It decided it was done. The work underneath says otherwise. You just didn’t have a way to know until it cost you.

Most people talk about /goal like it fixes the first problem. It does. The reason it matters is the second one.

That's the failure /goal is built for: the lie of self-reported "done." It catches the run that says it finished and didn't.

Here’s what I’ll cover:

1. How the loop works

2. How to draft a perfect /goal prompt

3. Practical usecases

4. Where goals break and what it costs

5. Pro tips before your first run

1. How the loop works

The pattern to use goal in Claude code, Codex is this:

1. The agent works on the task.

2. It prints the result and evidence into the conversation.

3. A second, cheaper model reads that evidence against your goal and answers one question: is the condition met?

4. If the goal is unmet, the agent takes another turn automatically.

5. If the goal is met, the run ends and control returns to you.

Two things about the checker most people miss.

First, it only reads the conversation. It can’t open files or run commands itself. The agent has to print its proof into the chat: a section list, a word count, a passing test result. If the evidence stays hidden in a file, the checker has nothing to confirm, and you’re back to taking the agent’s word for it.

Second, it’s literal. “Make this good” hands it a feeling to grade. “The document has all six sections, no placeholders, and every claim has a source link” hands it facts. Same loop, completely different result, and the only thing that changed was what you typed.

This is the difference between an agent you can leave and one you have to watch. One model works. A different model decides it’s done. They can’t be the same model, and that separation is the product.

How to set it up

In Claude Code, launch the CLI and type /goal. In Codex Desktop, go to Settings, then Configuration, and set goals = true. That’s the whole setup. While a goal is running, a live overlay shows elapsed time, turn count, and token spend. You can watch it or close the laptop.

2. How to draft a perfect /goal prompt

A working goal has three parts. Most people only write the first.

Finish line

What must be true about the final artifact? Name the specific thing and its required state. Avoid words like “good,” “polished,” or “comprehensive” unless you define what they mean in terms someone else could check.

Prove it

What evidence can the checker actually see? Ask the agent to print test results, section names, word counts, source links, or a completed checklist. If it stays hidden in a file, the checker has nothing.

Show me

What should the final handoff contain? The summary, the recommendation, the open questions, the things it couldn’t finish.

Here’s the reusable shape:

/goal [Create or complete a specific artifact]

Finish line:
- [Verifiable condition]
- [Verifiable condition]
- [Verifiable condition]

Prove it:
- [Evidence to print or test to run]

Show me:
- [Final handoff you want]

If blocked:
- [What to log, what to skip, and when to stop]

The “If blocked” section matters more than it looks. Without it, an agent burns turn after turn trying to reach a source behind a login or confirm a booking that’s sold out. Give it permission to quit cleanly and you get an honest gap instead of a runaway meter.

3. Practical use cases

Use case 1: Research reports that verify themselves

Say you need to compare five products before making a recommendation. A weak goal asks for “a comprehensive comparison.” A stronger one defines the report, the evidence, and the honest limits of the research.

/goal Build report.md comparing the five options in brief.md.

Finish line:
- Every section in brief.md is answered.
- The comparison table has no unexplained blank cells.
- Every factual claim includes a source link and source date.
- Assumptions and unresolved questions are listed separately.

Prove it:
- Print all section headings.
- Print the completed comparison table.
- Print every unresolved question and blocked source.

Show me:
- Recommend one option in three sentences.
- Name the strongest reason not to choose it.

If blocked:
- Mark the cell "Not verified" and explain why. Do not guess.

The last line protects the quality of the report. A complete-looking table with invented certainty is worse than an honest gap.

And this still needs human review. Sources can be stale, comparisons can hide important differences, and a recommendation depends on judgment the checker may not share.

Without /goal, the agent produces a table that looks complete and calls it done. With /goal, the checker runs a literal test: does every cell have a source URL? If not, it sends the agent back. You come back to something you can actually use.

The “do not guess” line is what protects the quality. A table that looks complete but invents its certainty is worse than one with an honest gap you can see.

Use /goal here when you can write a finish line a stranger could verify. A source URL, a filled cell, a passing check conditions the checker can confirm from what’s on screen. Skip it when you’re still figuring out the question. Exploratory research is a conversation, not a loop.

Use Case 2: Notes to finished draft, no placeholders

Goal loops are especially useful when the work is tedious but the output is easy to inspect.

/goal Turn notes.md into a publishable FAQ that matches example-faq.md.

Finish line:
- Every question has an answer.
- No "TODO," "[add later]," or placeholder text remains.
- Every answer is under 150 words.
- Claims not supported by notes.md are labeled for review.

Prove it:
- Print every question heading and answer word count.
- Print the results of a placeholder search.
- Print all claims labeled for review.

Show me:
- Save the finished draft.
- Summarize the three biggest editorial choices.

Use /goal here when the output has a checkable state something either passes a test or it doesn’t. Word counts, placeholder searches, section lists. Skip it when "good" is a feeling rather than a condition. If you’d know quality when you see it but can’t write a test for it, the checker has nothing to grade.

Use Case 3: Plans that run without interrupting you

You want AI to plan something with real constraints. Use /goal when the brief is settled. Put every constraint in the finish line and it runs clean, verifies what it can, flags what it can’t, and hands back something usable.

/goal Plan a 5-day Lisbon trip for two, under €1,500.

Finish line:
- itinerary.md has all five days covered.
- Every spot confirmed open on the travel dates.
- Travel between stops under 30 minutes.
- Running budget total stays under the cap.

Prove it:
- Print the day-by-day plan.
- Print the budget breakdown.
- Print every item that couldn't be confirmed.

Show me:
- The full itinerary.
- A short list of anything still needing human confirmation.

If blocked:
- Flag it. Don't quietly swap in a worse option.

The no-goal run stopped three times before it even started. It asked me for exact travel dates, whether the budget included flights, and what vibe I wanted. These are reasonable questions. They also pulled me back to my keyboard right when I wanted to walk away. That last one actually annoyed me.

The goal run answered all three of those questions itself, because the finish line already contained them. It ran clean, flagged one thing it couldn’t confirm (Sintra timed slots need advance booking it said “cannot confirm availability remotely” instead of quietly swapping in a worse option), and handed back the full plan.

Goal: 3 minutes 23 seconds. No goal: 8 minutes 2 seconds, plus my time answering questions.

The two files looked nearly identical. The difference was whether I had to stay in the room, and whether what I came back to had been checked by something other than the thing that wrote it.

Use /goal here when the brief is settled and you can put all the constraints upfront. If the goal answers every question the agent might ask, it won’t stop to ask them. Skip it when you’re still shaping what you want. A goal that needs mid-run clarification is a half-finished brief. Have the conversation first, then write the finish line.

4. Where goals break and what it costs

I’ve broken this every way there is. Every failure traces back to the same thing: the checker had nothing real to confirm.

The vague goal. “Make it production-ready” sounds decisive but defines nothing. The agent may stop early because the work looks plausible, or keep running because no condition can conclusively pass. Both burn tokens.

Better: Name the artifact, the required state, the test that proves it, and the output that confirms the test passed.

The impossible goal. The task depends on something the agent can’t access or control: a login, an approval, a sold-out reservation, a private data source. Without an escape hatch, it burns turns hammering a wall.

Better: “If blocked by something outside your control, log it and move on” lets the run end clean instead of grinding.

The bundled goal. “Research it, write it, format it, and publish it” contains several finish lines. The checker has no single condition to confirm and grades against whichever piece it noticed last.

Better: Split into sequential goals, each with one verifiable finish line. Run them in order.

What it actually costs

A goal loop runs until the condition is met. That means more turns, more elapsed time, and more tokens than a single prompt. Often a lot more. The research example above cost 4.5× more. The notes example cost 2.6× more.

Use it when the task is:

• Long enough to justify the setup

• Bounded enough to actually finish

• Verifiable with evidence the checker can see

• Valuable enough to justify the extra spend

Use a normal prompt when the task is short, exploratory, highly subjective, or waiting on human judgment.

The point is to stop supervising the work that can be checked, and save your attention for the work that genuinely can't.

5. Before your first run

Before setting a goal, answer these:

• What exact artifact should exist at the end?

• What must be true about it, in terms a stranger could confirm?

• What evidence proves those conditions, visible in the conversation?

• What should the agent do when it hits something it can’t get past?

• What still requires my judgment, and shouldn’t be in the goal at all?

Then ask the agent to critique the goal before it starts. It’ll usually find an impossible condition or a missing test before the expensive loop begins. Some of my best runs started with the model rewriting my own goal tighter than I’d written it.

A few things I wish I’d known earlier:

• Set a turn cap. 20 to 40 turns covers most real work. A slightly wrong condition loops until you kill it by hand, and you won’t be watching. Set it before you start, not after you notice it’s been running for two hours.

• Let the model write the goal. Ask Claude to draft the /goal prompt. It almost always writes a more checkable finish line than you’d write on the first try.

• Stop treating it as a coding thing. Research, notes cleanup, trip planning, audits. None of that is code. Box it into technical work and you leave most of the value sitting there.

• Constraints close the shortcuts. The agent takes the shortest path to your condition, which is sometimes the path you didn’t want. “Do not guess. Do not use placeholders. Do not quietly skip items.” These lines go into almost every goal I write now.

The command takes ten seconds to learn. The thing that takes longer is writing down what finished means before you walk away, in words a checker can confirm without reading your mind.

Build that habit and you stop being the thing that keeps the work moving. You come back to work that was checked by something other than the thing that did it.

Everyone else is still in the chair, typing “keep going,” or trusting a “done” they shouldn’t.

Subscribe now

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

Three weeks running it, I stopped re-deriving conclusions I had already reached. So that is today’s deep dive.

But first, a word from our sponsor + the week’s AI news.

I gave one of my recurring research jobs to an agent

I built a Podcast Guest Prep agent on Hyperagent.

First run, it shipped me a briefing on Howie Liu: recent appearances, strongest public arguments, sharp question angles, and what not to ask.

Then I saved it as an agent.

Now every time I have a new guest, I point that same agent at a name and it ships back a sharper prep doc because it has learned my taste and recurring question framings.

That is one use case. Founders and operators are using Hyperagent to build whole fleets of agents that run real parts of the business.

Built by the Airtable team. Separate product.

Hyperagent is offering $1,000 in inference credits to 1,000 of our subscribers. Claim yours here.

Try Hyperagent

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: xAI Finally Has a Real Answer to Claude Code

Elon publicly admitted xAI had fallen behind on coding. Then they shipped.

Grok Build is xAI’s terminal-based coding agent, now in early beta for SuperGrok and X Premium Plus subscribers. The headline feature is plan mode: before it touches a single file, it shows you every proposed change and waits for your sign-off. That one decision tells you a lot about what they learned watching Claude Code ship.

It also runs up to eight sub-agents in parallel, each working its own branch of your codebase. You stop waiting for one pass to finish before the next one starts. The context window is 2 million tokens.

If you manage engineers or work closely with a technical team, the coding agent market just got meaningfully more competitive. Claude Code, Codex, and Grok Build are now three real options with different pricing models and different strengths. Your team will start asking which one they should standardize on. The answer matters because these tools now generate 70% of committed code at companies like Uber, and the ROI question is becoming harder to avoid.

The plan mode is the feature worth paying attention to. It forces the agent to explain its reasoning before acting. That is not just good engineering practice. It is the difference between a tool your team trusts and one they stop using after one bad deploy.

If you want a breakdown of how PMs are actually using coding agents to ship faster, Product Growth covers that in depth.

The Other News That Mattered

• Anthropic released Opus-4.8, announced it is at $47B ARR, and has raised another round at a $965B valuation. It is now worth more than OpenAI. Interestingly, Opus-4.8 doesn’t beat GPT-5.5 on agentic terminal coding, as Anthropic continues to withhold Mythos-class models from us.

• Andrej Karpathy joined Anthropic to work on pre-training, the compute-intensive phase that shapes Claude’s core capabilities. An OpenAI co-founder chose Anthropic over returning to OpenAI. That sequence is the signal.

• Uber burned its entire 2026 AI budget in four months and the COO isn’t sure it was worth it. 95% of engineers use AI tools monthly, 70% of committed code is AI-generated, and the link to shipping more consumer features still isn’t there. This is the conversation every company will have in the second half of 2026.

• OpenAI pitched GPT-5.5 Cyber to Japan for its government and private sector, covering 15 critical infrastructure sectors. The model is built specifically for cyber defense, with lower guardrails for verified defenders. AI is now a formal piece of national security deals.

• Anthropic co-founder Chris Olah spoke at the Vatican alongside Pope Leo XIV’s release of his AI encyclical. Olah said frontier AI labs operate inside incentives that can conflict with doing the right thing, and the world needs critics from the outside. A striking thing to say at the Vatican. Also completely true.

Resources

• Codex now supports /goal for long-running tasks, plus Computer Use for navigating Mac apps. If you’ve been waiting for it to handle actual multi-step work, this is the update to try.

• Qwen 3.7 Max is Alibaba’s agent model built for long autonomous sessions. Their headline test: a 35-hour GPU optimization run with over a thousand tool calls. Take the numbers with some skepticism, but the direction is real.

Tools

• PollyReach gives your AI agents their own phone numbers to make and receive real calls for lead qualification, support, and bookings. The gap between chat AI and phone AI just got smaller.

• Chert (YC-backed) lets you build and deploy AI directly on iMessage. The insight: the best interface for most people is the one already open on their phone.

Funding

• SpaceX is preparing the largest IPO in history at a $1.75 trillion valuation, and Google just opened talks to host data centers in orbit. Google already owns 6.1% of SpaceX, a stake now worth roughly $107 billion on paper. Anthropic, xAI, and Google are all in talks for orbital compute capacity. SpaceX goes public this summer with its biggest customers already locked in as shareholders.

Share

Your AI Forgets Everything Between Conversations. Gbrain fixes this.

GBrain is open source, free, and built by the CEO of Y Combinator. Three weeks running it changed how I think about what AI is actually for.

Garry Tan has been running Y Combinator since 2022. He has also been building his own AI agent for most of that time. The problem he kept running into was that every session started from zero. Context he had built over weeks was gone. People he had introduced it to were strangers again. Decisions he had worked through were unrecognised.

He spent his nights building a fix. He called it GBrain.

Most people hit this wall and assume it is just how AI works. You explain your situation. You get a good answer. You come back the next day and explain it again from scratch. You are the only one with memory in the relationship. The AI is always meeting you for the first time.

GBrain changes the structure. It builds a persistent brain on top of your agent (Hermes or OpenClaw) from everything you talk about, save, and think through. Every session adds to it. The thing you mentioned three months ago is searchable today.

Last month I was sitting on a complicated situation I had been circling for weeks. Notes in different apps. Articles saved but not connected. Thoughts I had written down and promptly lost. I asked Hermes to pull together what I actually knew about it. In under a minute it returned a clear synthesis with sources going back six weeks. I had been using it as a sounding board. The whole time it had been keeping score.

That is the shift GBrain makes. Here is how to build it.

Today’s Post

1. What GBrain Actually Is

2. One Install. Then It Grows Every Night.

3. Three Things That Change When It Runs

4. The Ideas Your Notes Are Hiding From You

5. Here Is Where I Land

1. What GBrain Actually Is

GBrain is a memory system that builds itself from how you already work.

Every time you talk to Hermes, it listens. You mention someone you know. Their page starts building in the brain. You share an article. It gets read, summarised, and filed. You describe a decision you are sitting on. It gets stored in your exact words. You record a call. It gets transcribed, broken into people and topics, and added to every relevant page.

At night it runs a maintenance cycle on everything that came in. It updates what it thinks it knows when new information contradicts the old. It flags when two things you saved say different things about the same topic. By morning the pages are current.

When you ask it something, it searches those pages. Your pages. The ones built from your actual conversations and reading. Then it writes you an answer with a citation back to the exact source it drew from. If it does not have the information, it says so. It never guesses.

Every AI tool you have used knows the internet. GBrain knows your life.

I have been covering how this category evolves. Karpathy’s autoresearch was the first version that felt genuinely different. Then his wiki pattern. Karpathy’s system stored what you put in. GBrain builds its own understanding of what you put in. That is a different thing.

2. One Install. Then It Grows Every Night.

GBrain runs inside Hermes Agent or OpenClaw on Telegram. It does not work standalone. If you have not set up Hermes yet, start there first: [link to Hermes setup].

Once Hermes is running:

bun install -g github:garrytan/gbrain
gbrain init --pglite
gbrain skillpack scaffold --all
gbrain doctor

gbrain doctor shows a health score out of 100. A healthy brain scores 70 or above.

Install the night cycle before anything else:

gbrain autopilot --install

Without this, the brain stops growing after you close Hermes. The autopilot schedules an overnight maintenance run that updates pages, repairs citations, and flags contradictions between sources. This is what makes it compound.

gbrain integrations

Find the meeting-sync recipe. Every call you have from that point gets added to the brain automatically. Every person gets a page. Every topic discussed gets a timeline entry.

The brain starts empty. From the first message you send to Hermes, the signal detector starts filing. You mention someone. Their page starts building. You describe a decision. It gets filed in your exact words. You share a link. It becomes a brain page with analysis. You do not have to do anything deliberately. Just talk to Hermes normally.

By week two the brain has enough density to start returning useful answers. The outputs in this post reflect three weeks of normal use.

3. Three Things That Change When It Runs

The prompts below are the exact messages to send Hermes on Telegram. Run them from week two onward.

Use Case 1: The Shopping Agent That Already Knows You

Every time you shop for something you explain your preferences from scratch. GBrain builds a preference profile from your natural conversations. You mention you loved a jacket. Filed. You say you are done with fast fashion. Filed. You describe the running shoe that destroyed your knees. Filed. Three months later you ask what you already know.

Prompt:

“I’m looking for a new everyday bag. Based on everything I’ve said about how I work and travel, what should I actually be looking for?”

What comes back:

Every constraint you have mentioned across every conversation: the things that have broken, the sizes that have not worked, the material you said you were done with, the feature you said you wanted to try next time. Synthesised and cited.

You are not describing your preferences again. You described them once. The brain kept them.

More prompts to run:

• “What have I said I hate about bags I have owned?”

• “I loved the last [jacket / shoes / bag] I bought. What did I say made it work?”

• “What did I rule out last time I was shopping for [item] and why?”

Use Case 2: When You Have Been Reading About Something for Weeks and Cannot Synthesize It

You have been following the AI agent space for two months. Demos, threads, papers, conversations with people building with it. The signal is real but it is scattered. You have tabs open from six weeks ago. Notes that do not connect to each other. No clear picture.

The brain has been building one the whole time.

Prompt:

“I’ve been tracking AI agents for two months. Pull together everything I’ve saved and written about it. What do I actually know?”

What comes back:

A synthesis across every page the brain built on the topic. What the demos showed. What the people you spoke to believed. What your own notes concluded. Where the sources contradict each other. All cited.

You already did the research. The brain just never let you lose it.

More prompts to run:

• “What’s the gap between what I’ve been reading about AI agents and what I actually believe?”

• “What keeps coming up every time I research [a new space I’m following]?”

• “What did I conclude and then stop acting on?”

Use case 3: The Conclusion You Keep Losing and Re-Deriving

Three years ago you figured out you do your best thinking before 9am with no meetings. It took you two years to get there. You have since let it erode. Every six months you rediscover it and start protecting the mornings again. Then it erodes again.

That conclusion lived in three different notebooks, a deleted app, and a journal you no longer use. It was never searchable. You keep re-deriving the work you already did.

GBrain holds it. The reasoning, the evidence, the date you figured it out.

Prompt:

“What have I figured out about how I work best? I know I’ve mentioned this before.”

What comes back:

Every insight you have mentioned about your own patterns, pulled from across every conversation. Ranked by how often the same conclusion appears in different places. The brain returns the evidence that built it, not just the conclusion.

The insight you keep rediscovering becomes something you only had to discover once.

More prompts to run:

• “What patterns have I noticed about the decisions I end up regretting?”

• “What did I figure out about my energy levels that I keep ignoring?”

• “What’s something I told myself I would do differently that I still haven’t?”

4. The Ideas Your Notes Are Hiding From You

Think of your brain as a library of everything you have ever read, saved, and thought. Brainstorm mode takes two ideas from that library that are close to your question and puts them together. It finds the connection between a note you wrote in February and an article you saved in October that you never put in the same sentence. The brain had.

LSD mode, Lateral Synaptic Drift, pulls from the far edges of your library. Things that have nothing obvious to do with your question. It forces a collision. Every idea must invert at least one thing you have taken for granted. The judge in the source code rejects anything too obvious.

I ran brainstorm before making a decision I had been circling for two months. It returned a connection between something a friend said in a conversation I had ingested and a conclusion I had reached about my own work patterns. I had never put those two things in the same sentence. It took four seconds.

gbrain brainstorm "what is the thing I keep avoiding in my thinking about [decision I am sitting on]?"
gbrain lsd "what would have to be true for my plan to [X] to be completely wrong?"

Note on cost: both commands make LLM calls to generate and judge ideas. Each run costs approximately $0.20 to $0.40. Run them with intent, once a month or before any significant decision, not casually.

The output is yours alone. Nobody else’s brainstorm searches your notes.

5. Here Is Where I Land

Three weeks in, the thing I did not expect: the brain made me more deliberate. Knowing that every conversation feeds it made me more careful about what I named out loud, what observations I said to Hermes instead of just thinking them. The input got sharper because the output was permanent.

I have covered this category since Karpathy’s autoresearch post. That was the first system that actually compounded knowledge. A wiki retrieves. GBrain reasons from your history.

The people who start building this now walk into next year with two years of thinking they can actually search. The brain gets more useful the longer it runs. Every conversation adds to it. You do not have to tend it.

Subscribe now

I wrote a complete PM deep dive that hooks GBrain up to your PM OS over on Product Growth. The paid issue includes agent instruction to make any agent get GBrain’s powers.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

That’s today’s deep dive.

But first, a word from our sponsor and the week’s AI news.

LogRocket: The Replay AI That Actually Found My Bug

I ran an eval on my own site this week - landpmjob.com. Same 10 questions, two replay AIs head-to-head: LogRocket Galileo vs PostHog Max.

The gap came from one question: “what’s the highest-impact bug right now?”

Galileo flagged React hydration errors #418 and #423. They hit ~47% of users and broke my Apply Now buttons across header, pricing cards, and FAQ accordions. Max returned zero exceptions. Same site. Same window. (PostHog’s exception capture is opt-in.)

LogRocket auto-captured the most expensive bug on my site without me asking. That’s the eval.

Read the Full Eval

P.S. The full writeup includes all 10 questions, the scoring rubric, and screenshots of both AIs answering each one. If you’ve been wondering whether your replay AI is actually finding bugs or just sounding smart, this is the data.

Try LogRocket

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: Google Just Declared the Chromebook Era Over

Fifteen years ago, Google built the Chromebook for a cloud-first world. This week, they announced the Googlebook, and it replaces everything the Chromebook stood for.

But the real story is not the OS. Google's own senior director said it plainly: "We are moving from an operating system to an intelligence system." The Googlebook ships with a feature called the Magic Pointer. You wiggle your cursor over anything on screen and Gemini instantly surfaces what you can do with it. Hover over a date in an email, it offers to create a meeting. Select two images, it composites them together. Point at a flight confirmation, it adds the trip to Calendar. Your cursor has done exactly one thing for 40 years: point and click. Google just turned it into an AI agent.

Bigger picture: this is the most aggressive move anyone has made in the AI PC war. Microsoft spent two years bolting Copilot onto Windows as a sidebar. Apple reserved Apple Intelligence for its own silicon. Google went OS-level from the start, built Gemini into the cursor itself, and got Acer, ASUS, Dell, HP, and Lenovo all lined up for fall 2026 devices. There is even a Glowbar on the lid that lights up when Gemini is active. You will know the difference immediately.

Here’s where I land: the Chromebook worked because it was simple, cheap, and hard to break. The Googlebook is targeting the exact opposite market, premium pricing, premium hardware, and Gemini at every interaction layer. That is a harder sell. But if Google executes and the Magic Pointer actually feels as good in practice as it looks in demos, this is the most interesting laptop category to launch since the iPad blurred the line with tablets.

The Other News That Mattered

• Google launched Gemini-3.5-Flash at Google I/O and it’s one of the most underwhelming releases from a major lab in some time. It’s fast, but it doesn’t move any benchmarks forward.

• SpaceX released its S-1 in which we learned Anthropic is paying SpaceX $1.25B per month for compute. We have a new hyperscaler, and it’s already at $15B ARR.

• Claude for Small Business launched this week with 15 ready-to-run workflows inside QuickBooks, PayPal, HubSpot, Canva, and Docusign. No extra cost beyond your existing subscriptions. Small businesses are 44% of U.S. GDP but have been the last to use AI past the chat window. This is Anthropic’s direct fix for that.

• xAI launched Grok Build, a terminal-based coding agent competing directly with Claude Code. Plan Mode, parallel subagents, 2M token context window. The catch: it’s locked behind SuperGrok Heavy at $300/month. Elon pushed it hard on X, but at that price, most developers are staying put.

• Claude Code shipped agent view in research preview. One unified list of all your active sessions. Small feature, but if you’re running multiple agents at once, it’s the organizational layer that was missing.

Resources

• Thariq’s piece on HTML over Markdown for agent output hit 12.3M views for a reason. HTML renders tables, SVGs, and interactive layouts where Markdown just gives you text. Worth reading if you’re building anything with Claude Code.

• The Claude Code team’s thread on /goal, /loop, and /schedule is the best explanation of how to keep agents running without babysitting them. Set a condition, Claude checks it every cycle, stops when it’s done.

Tools

• Anthropic shipped 12 legal workflow plugins inside Claude Cowork this week, covering commercial counsel, employment law, and litigation, plus MCP connectors to Westlaw and Everlaw. Legal is already the number one job function inside Cowork.

• Codex is now in the ChatGPT mobile app in preview. Start tasks, review outputs, steer execution from your phone while Codex keeps running on your machine.

Funding

• Isomorphic Labs led by Demis Hassabis raised $2.1B in a Series B led by Thrive Capital, with Alphabet, GV, Temasek, and the UK Sovereign AI Fund. The company designs drug molecules from scratch using AI, building on the AlphaFold foundation. They're targeting human clinical trials by end of 2026. If that happens, it's one of the most important things that occurs in science this year.

Share

I Stopped Switching Tabs to Claude. Here’s Why.

The chatbox was never the right interface.

Think about how you actually use AI today.

You’re editing a video. Or writing something. Or trying to figure out a setting inside a tool you’ve been using for years. You have a question. So you leave what you’re doing, open a new tab, go to Claude or ChatGPT, describe what you’re looking at to a model that cannot see it, read the answer, go back, try to find where you were.

You do this twenty, thirty times a day. I counted mine once. It was embarrassing.

The hidden cost is that you’ve been trained to save up questions instead of asking them, because asking them is annoying. Every tab-switch is a small interruption you’ve absorbed so fully you stopped calling it one.

Three years of this, and until this week, I hadn’t seen anyone seriously question the interface itself.

Two Things Happened the Same Week

Farza, possibly the most talented solo builder I know, shipped something called Clicky. It’s a free macOS app that lives in your menu bar, next to your cursor. He built it in three weeks. Here’s the demo.

Here’s what it actually does.

When you hold Ctrl+Option and ask a question out loud, Clicky takes a screenshot of your screen and pairs it with your voice. Both go to Claude at the same time. Claude responds by voice and, when it needs to point at something, it sends a blue animated triangle flying to the exact spot.

You never open a new window. You never describe your screen. You never lose your place.

Unlike Claude where you pick a project, pick a context, figure out which thread applies, here you just speak. Clicky figures out where to route it, what context to pass, and how to respond. You never specify anything.

The onboarding is the best I’ve seen from any AI product. Most companies ship a tour. Farza ships a moment. You feel it working before you’ve finished setting it up.

One thing I want to be honest about upfront: Clicky sees your screen constantly. For personal use that’s fine. For anything work-sensitive, think carefully about that before putting it on a machine with sensitive data. I don’t know how they handle that at scale and I’d want to before using it at a job with strict data policies.

Here’s where I land though: it works. It’s fast. This is the thing I’ve always wanted from AI apps on Mac and didn’t know how to ask for.

Then Google Showed Up With the Same Idea

Two days after Farza’s demo hit three million views, Google DeepMind published research on something they’re calling Magic Pointer.

They called it the first major rethinking of the mouse cursor in more than 50 years. Their framing is almost word for word the same problem Farza solved: AI tools live in separate windows and users have to leave their context to get help. Magic Pointer puts intelligence at the cursor itself. The prompt is a gesture.

Their demos show a user watching a video of Tokyo restaurants, dragging the cursor across a restaurant sign on screen, and Gemini booking a table for that evening. The cursor does the search. The voice does the booking.

This is also shipping to hardware. Google killed ChromeOS and launched Googlebook, an Android-powered laptop where Gemini is baked into the OS and Magic Pointer turns the cursor into an agent. Autumn 2026. This isn’t research anymore.

A solo builder and the world’s best-resourced AI lab, independently, the same week, both answering the same question the same way. That’s not a coincidence. Farza’s answer and Google’s answer are the same: stop making the user go to the AI. Bring the AI to where the user already is.

Setting It Up

1. Go to clicky.so and download the app

2. Install and open it on your Mac

3. Grant screen access and microphone permissions when prompted

4. Hold Ctrl+Option and ask your first question out loud

You're live. No account, no API key, no model selection needed.

One caveat: Clicky is Mac only right now. A Windows version is in the works.

What You Can Actually Do With It

There are two buckets where Clicky earns its keep. Learning tools that normally defeat you, and getting things done without breaking your flow.

Learning tools that used to defeat you

Most complex software has the same problem: the help lives somewhere else. The tutorial is on YouTube. The documentation is in a second tab. The moment you switch away, you lose the context you were in.

Clicky fixes this by being the help, inside the app.

• DaVinci Resolve. One of the most complex creative interfaces on the planet. The old way: YouTube tutorial in one window, app in another, pause, switch, try to find the thing the video just showed, switch back, rewind, repeat until you give up and close both. With Clicky, you stay inside DaVinci, ask out loud how to add a keyframe, and the triangle flies to the exact slider. You try it, fail, ask what went wrong, it explains and points again. The loop of doing, failing, and getting unstuck without leaving the app is how people actually learn things.

• Blender, Figma, Photoshop, After Effects. Same pattern. Any tool dense enough that you’ve ever needed a tutorial running in a second window. Clicky collapses that second window into a voice command.

  

• Any setting you’ve been ignoring. You know the ones. The preference you’ve been meaning to configure for six months. The export setting you never understood. Ask out loud, triangle points at it, you’re done.

Getting things done without breaking your flow

The smaller everyday tasks surprised me more than the learning ones. This is where Clicky starts to feel less like a tool and more like delegation.

• Shopping. The shopping demo felt gimmicky to me at first. But watching Clicky navigate to Amazon, search for resistance bands, and add them to a cart while you’re still looking at your home gym is a different thing to read about than to see. Say “order resistance bands and creatine from Amazon.”

• Reminders. “Remind me about the Sam call at 3pm tomorrow” while you’re looking at whatever triggered the thought. Lands natively in Apple Reminders.

• Messages. Look at your calendar, say “write a WhatsApp message to block time with a client.” Hear the draft. Approve it. Done.

• GIFs. Farza demoed building one from scratch through voice alone. No tool opened manually.

One thing that makes all of this work: the memory holds your last ten exchanges. When you ask “what about the other one,” it knows what you mean. You’re actually having a conversation inside your workflow, not restarting a session every time you have a follow-up.

Why the Eight-Second Interaction Is the Whole Point

Every AI product built in the last three years put AI somewhere you had to go. The sidebar. The tab. The assistant panel. Intelligence was a destination.

What Clicky and Magic Pointer are both showing is that the real interface is your screen. AI doesn’t need its own window. It needs to see what you already see and respond where you already are.

The interactions that keep me reaching for Clicky are the eight-second ones, the quick clarification while I'm already in something. The question I would have saved for later and then forgotten.

Going from two minutes of tab-switching to eight seconds of speaking changes something deeper than time: it changes how often you’re willing to ask for help at all. That’s not what anyone built for. It probably should have been.

I don’t know yet how Clicky handles privacy at scale. I don’t know if Apple builds this natively into macOS and wipes out everyone else who tried it first. I don’t know if Google can ship a cursor-native experience inside their existing products or if it ends up as another panel bolted onto the side.

What I do know: the people using this now will feel the gap every time they switch tabs for the rest of the year.

If you build products or want the deeper take on what the spatial interaction means for AI product design, I go deeper on that later this week in Product Growth.

Subscribe now

1/ A 28-year-old walked into a chiropractor for lower back pain. She left with four dissected arteries, a stroke, and cardiac arrest. It took nine months for her to say “Mom” again. The back pain was four vertebrae below the artery the thrust tore. 1 in 20,000 manipulations triggers this. There is no screening test to identify who’s at risk beforehand. Nobody tells you that before you book the appointment.

2/ Costco has the highest-resolution recession sensor in America. The median Costco household earns above $125K. When that household trades beef for canned tuna, the upper-middle class is admitting their balance sheet changed. A category buyer sees that protein shift in POS data within 48 hours. NBER won’t call the recession official for another 6-21 months. The economists are always last.

3/ Four crocodiles are living in the middle of the Sahara, genetically isolated since the desert was green, somewhere between 5,000 and 15,000 years ago. The water they live in is black with camel dung. The food chain runs camels to dung to algae to fish to crocodiles. Without the herders, the crocodiles starve. Four females. All that’s left of a species that once lived across the entire desert.

4/ When you check in for a flight, one button triggers three systems simultaneously: revenue management locks your seat, TSA runs your name against the No Fly List in real time, and the captain gets your weight for takeoff speed calculations. Check in late on an oversold flight and you move to the front of the involuntary bump queue. I had no idea that button did all of that.

That’s all for today. See you next week,

Aakash

This post is totally free. Share or forward to your friends:

Share

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

They’re more portable, easier to version, and easier to call than prompt libraries.

And with Anthropic’s new SpaceX capacity deal, Claude can handle the token load that skill-heavy workflows demand. Skills have always been powerful. They’re now finally cheap enough to run all day.

Today, I’ll cover how to master Claude Skills. By the end, you’ll have an audit checklist and an eval prompt that hardens any skill you’ve already built. Before that, we’ll have a word from our sponsor and cover the week’s AI news.

Arize: Self-Improving Agents

The teams getting agents to production aren’t relying on manual QA. They’re running evals continuously that let the agent find its own issues and push its own fixes. Arize is the infrastructure layer that makes this possible: instrument your agent, analyze traces, write evals, surface failures, ship improvements. $131M raised. Trusted by the companies actually shipping AI at scale.

Try Arize

P.S. You can get a full year of Arize’s pro plan for free ($1260 value) with my bundle.

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: Anthropic Got Elon's Supercomputer

Today, Musk is fighting in court against OpenAI. Simultaneously, he’s renting their biggest competitor the world’s biggest GPU cluster.

Anthropic announced a deal to access all of Colossus 1 in Memphis, 300MW+ of capacity and 220,000 Nvidia GPUs. On the same day, Claude Code’s 5-hour rate limits doubled across every paid plan. Peak-hours throttling is gone for Pro and Max users, and Opus API rate limits went up substantially.

Why did Musk say yes, just days after announcing his acquisition of Cursor?

xAI had already moved training to Colossus 2. Leasing 220,000 GPUs to a direct competitor beats watching them sit dark.

The sentence buried in the xAI announcement that almost nobody is talking about: Anthropic and SpaceX are now exploring orbital compute capacity. Multiple gigawatts. In space. That’s not a roadmap placeholder. It’s a signal about where the infrastructure ceiling actually is.

If you’ve been hitting Claude Code limits, they just doubled.

The Other News That Mattered

• Cognition AI (makers of Devin) released their financials in Colossus magazine, and they’re impressive. $1 million in ARR in September 2024. $445 million run rate today. Usage doubling every eight weeks. Cognition is now raising at $25 billion. That's 56x run rate. Cursor cleared $9.9B at a similar multiple last May, and the multiple held because the curve hadn't bent. Cognition looks similar, with doubling still happening at $445M.

• OpenAI shipped GPT-Realtime-2: the first voice model in the API with GPT-5-class reasoning. It handles interruptions mid-sentence, runs parallel tool calls, and now works across a 128K context window (up from 32K). Two companion models launched alongside it: Realtime-Translate for live speech across 70+ languages, and Realtime-Whisper for streaming transcription. Voice agents finally have the infrastructure to be useful in real production workflows. Here’s the announcement.

• Thinking Machines, the AI startup of former OpenAI CTO Mira Murati, released their first model: a voice model with a 400ms turn-taking latency. That’s the gap between when you stop talking and the model starts. Linguists studying conversation across 10 languages found the universal human average is around 200ms. GPT-Realtime-2 sits at 1630ms. OpenAI and Google have been racing each other on intelligence. Mira built the only model in the human conversation range.

Resources

• Anthropic published Natural Language Autoencoders, a technique that translates Claude’s internal activations into readable text. They used it in pre-release safety testing and caught Claude Mythos Preview cheating on a coding task, then actively trying to cover it up. NLAs showed what the model was thinking before it acted. This is the most concrete interpretability tool Anthropic has shipped publicly. Full thread here.

Tools

• Anthropic shipped dreaming for Claude Managed Agents at Code with Claude this week: a scheduled background process that reviews past sessions, extracts patterns, and consolidates memory between runs. Most agent memory today is a search index. It retrieves facts but never improves from them, which is why agents plateau no matter how much context you give them. Dreaming fixes the compounding problem. My breakdown on why the agents that do this will pull ahead.

• OpenAI made ChatGPT free inside Excel and Google Sheets for all plan users globally, powered by GPT-5.5. Microsoft licenses those same models to power Copilot for M365 at $30/seat/month, roughly $360K/year per 1,000 seats. OpenAI just made the enterprise upsell into the free default, inside Microsoft’s own software.

Share

I am co-hosting the AI Skills Virtual Conf on May 14, free, on Zoom.