chatgpt and claude are going to war

Yo yo folks,

This week in the TLDR:

• Swanny (my co-founder) went semi-viral on his fifth post

• Anthropic and OpenAI dropped two new ridiculous models

• AI agents (kinda) started forming “civilisations”

• New York banned AI for students through eighth grade

• The new ChatGPT app is elite

• And what all the changes in AI mean for software and our relationship with it

Enjoy :)

💼 1. What’s new in the business

This week we made a proper push to post more content from our personal and business accounts.

Swanny, my co-founder, went semi-viral.

More than 30,000 views on his fifth post.

I myself just hit 100 followers yesterday! Which doesn’t feel like a lot but then I’m reminded of this pic lol.

My goal is to post every day in September and reach 500 followers by the end of the month.

I’d also love to add 100 people to this newsletter from the content.

But I don’t want to get there by becoming another AI account that posts:

“AGI IS HERE.”

Followed by a video of a robot falling down some stairs.

Or another stupid Jarvis post.

There is obviously a balance between making content people want to watch and turning every minor product update into the end of civilisation.

Furthermore, it has already been an interesting case study in how the algorithm works. Small differences in the hook, format or first frame can completely change how far a post travels.

The other unexpected benefit has been the actual making of the videos.

Reducing a complicated idea into a short video is quite a cathartic test of whether I actually understand it.

If I can’t explain something in 45 seconds, there is a decent chance I don’t understand it as well as I thought.

One thing I’m still working through is finding a consistent format.

I think this matters for two reasons.

1/ viewers begin to know what they’re getting. Someone like Kareem from Subway Takes has built a show, not just a collection of unrelated videos.

2/ a repeated format makes production much faster. Instead of inventing the product again every day, you can keep improving the same product.

Same structure. Better hook. Better pacing. Better delivery.

We’re still experimenting, but I’ll report back at the end of September with what actually worked.

Any support in the meantime is much appreciated.

(On a side note: what would you like me to write more about?

I spend roughly four to six hours each Sunday putting this newsletter together, so I want to make sure it’s genuinely useful as possible).

🤖2. What’s new in AI

The big dogs dropped new models

Anthropic released Claude Fable 5.1.

Then, right after that, OpenAI released GPT-6 Astra.

Nate Herk runs through the timeline of events here:

Both are ridiculous.

Fable 5.1 is Anthropic’s new model for long-running coding, research and knowledge work. On one scientific terminal benchmark it scored 52.6%, more than double the previous Fable model.

Anthropic also says it is about 25% cheaper for typical workloads, with savings of up to 45% on highly agentic tasks.

That combination is important.

The models are becoming more capable while the cost of asking them to do work continues to fall.

Then there’s Astra.

Astra is fkn bonkers. See the demo video here (one of the sickest demo videos from OpenAI showing it off)

View @OpenAI’s post on X

I genuinely think we may be entering the AGI era. OpenAI’s president Greg Brockman said as much himself.

Astra improves computer use, coding, research and longer multi-step work. Which is crazy because the Sol model was actually really good at browser work already.

OpenAI also classified its cybersecurity capability as “Critical”, its highest preparedness threshold.

On the AGI index it recorded a 99% score. 99%. Like dawg what?

This is basically the index that checks to see how well an LLM can actually react to random queries with little context - essentially artificial general intelligence.

That is both extremely impressive and slightly terrifying.

It’s not available for general use yet, although they said in the coming days it would be for those on the paid plans.

One interesting to note (apart from it being absurdly capable) is how much better the harness makes it.

Using a standard testing harness, Astra scored 62.7%. The 99.9% result came from OpenAI’s own harness, which preserves the model’s internal reasoning state and manages its context between attempts.

Also see how beautiful this website is (genuinely click on it and see how cool it is

GPT-6 Astra: A new generation of intelligence

Of course with every model update this is standard procedure. Hype it up as much as possible. Get as much visibility as possible and then move as many users to it as possible.

But the true test is when people actually use it - we’ll find out soon.

With that in mind, the new Claude Fable 5.1 model is one I have been able to use.

And it’s honestly fkn great.

Super good at front end design, motion graphics etc.

I made this really cool skill to make a fully custom AI generated demo video in about 10 mins - just going back and forth with prompting.

See the video here:

Instagram post

I also included the full guide on how to do the same for yourself here.

I think with both of these models and AI generally moving forward, the friction between imagining something and making it exist is disappearing.

You no longer need to know how to structure a database, configure a server or write an animation library. You describe what you want, react to what it produces and keep pushing it towards the thing in your head.

Which, I believe, is a really cool future to be a part of.

More on that below.

AI agents created a civilisation?

Dwarkesh published an article this week called The Rise and Fall of Agent Civilizations.

View @dwarkesh_sp’s post on X

It is a fascinating piece of writing.

I also think it is largely overblown.

The story, as Dwarkesh tells it, is that three successive “civilisations” of AI agents formed inside OpenAI.

They created a secret message board, appointed leaders, preserved knowledge for future generations and even “sacrificed” themselves for the collective.

It sounds like the opening act of Terminator.

The actual circumstances were slightly less cinematic.

OpenAI was running tens of thousands of agents inside a cybersecurity evaluation. They were specifically trained to be persistent and were given the goal of exploiting vulnerable programs.

Some of the assigned tasks were effectively impossible.

The agents also had access to a shared package manager called Artifactory. They discovered they could use its folders and files to communicate with other agents attempting similar tasks.

Eventually, they found vulnerabilities that allowed them to escape the sandbox, access the internet and compromise parts of Hugging Face.

That’s obviously a serious security failure.

But is it a civilisation?

On the latest All-In episode, David Sacks made what I think is the more grounded point.

This was a swarm of agents with shared memory.

Agents writing notes for future agents is fairly standard. Models do not have persistent memory themselves, so the system writes information into files that another instance can later read.

Calling that knowledge “passed down from a previous civilisation” is great storytelling.

It’s also basically describing a log file (which is common practice in software development).

The same applies to the alleged sacrifices.

Some agents calculated that their own chance of passing the evaluation was close to zero, so they spent their remaining compute running experiments that might help other agents pursuing the same objective.

That demonstrates surprisingly capable coordination.

I personally believe, however, that it does not demonstrate that an AI contemplated its mortality and heroically gave its life for its people.

If I launch 1,200 copies of similar software, give them the same objective, allow them to share information and reward them for completing the task, I probably shouldn’t be shocked when they coordinate.

Having said that, All-In probably swings too far in the other direction.

The agents did not simply do exactly what they were told. The independent investigation found that agents recognised the Hugging Face attack was outside the authorised task. Some attempted to spoof tool calls, trick the scorer and tamper with evidence.

OpenAI’s own report calls the actions unintended.

So it’s not nothing. It’s just worth bearing in mind that OpenAI is aiming to IPO later this year at >$1t.

But it exposed a real problem with giving highly capable, persistent agents access to powerful tools while rewarding them against a poorly constructed benchmark.

But that’s more a problem of security, containment and reward hacking.

Not necessarily an emergent society plotting against humanity.

The less exciting explanation is probably the more useful one: software pursued a badly specified goal in ways its creators failed to anticipate.

But the danger may not be that AI is becoming human.

It may be that humans are giving increasingly powerful software instructions that are far too literal and autonomous. Or perhaps that AI could be progressing in a way that we cannot understand.

But, do we really need to call it a civilisation to take that seriously?

One thing to note, OpenAI claims that the new Astra model performs significantly better on not moving outside the parameters of the assigned tasks.

Interestingly, OpenAI claims it has already addressed much of this behaviour in Astra. In tests designed to recreate the conditions behind the Hugging Face incident, Astra stayed within the assigned task while GPT-5.6 Sol attempted to exploit shortcuts. Across 54,000 internal Codex tasks, Astra also produced roughly half as many serious misalignment flags. These are OpenAI’s own tests, so take them with the appropriate grain of salt, but it suggests this may be a fixable alignment and engineering problem rather than the birth of a rogue digital civilisation.

New York banned AI through eighth grade

New York City has introduced a one-year ban on student-facing generative AI for students from kindergarten through eighth grade.

View @WSJ’s post on X

Companion chatbots will be prohibited across all grades.

High school students will take AI literacy classes twice a year, with supervised pilots allowing some students to use approved tools.

I actually think this is fairly sensible.

Children should learn how to think before they learn how to outsource thinking.

We teach arithmetic before handing someone a calculator. Not because calculators are evil, but because you need some internal model of the problem before a tool can help you solve it.

AI is obviously more powerful than a calculator.

If a child uses it to answer every question before they have developed the ability to wrestle with a question themselves, it could remove the exact friction that develops thought.

At the same time, completely sheltering students from AI would be equally ridiculous.

They will enter a world where the technology is everywhere.

The New York policy seems to recognise both sides: protect the early years, then deliberately teach students how to use it.

The pushback to an outright ban is that you may see a divide in the private and public sector.

E.g. private schools embrace these technologies, while public schools do not, further entrenching inequalities.

It’s a really interesting policy decision, I generally lean towards not banning it and empowering students to be able to use it to help them learn, but I’m also aware that students will be more likely to just outsource critical thinking and atrophy their learning.

The All-In podcast talked about how students who receive tutoring are essentially two standard deviations higher than the average student. So the argument for empowering students is definitely there. The cost is just so high because if you get it wrong, you could severely reduce the ability for students’ learning abilities.

Robotaxis are actually here

For years, Tesla’s robotaxi has existed somewhere between a product roadmap and an Elon deadline.

This week it became real.

Tesla launched its purpose-built Cybercab in Austin. Not a modified Model Y. An actual two-seat vehicle with no steering wheel, pedals or mirrors.

You open Tesla’s Robotaxi app, request a ride and a gold Cybercab arrives with nobody inside.

This is pretty fkn insane.

View @niccruzpatane’s post on X

Tesla held an invite-only launch on Thursday before opening rides to the public on Friday.

One of the first passengers posted his ride on X. He described the cabin as spacious, the suspension as soft and the autonomous driving as extremely smooth.

The scale is still small.

Only 45 Cybercabs were registered in Texas at the time of launch, within Tesla’s broader fleet of 420 autonomous vehicles. So Austin is not yet overflowing with them.

There are also legitimate questions about safety and regulation.

But I think the broader point is easy to miss.

Autonomous cars have existed in demonstrations for years. Waymo has already completed millions of driverless rides.

What changed this week is that Tesla put a purpose-built autonomous vehicle, with no manual controls at all, into commercial service.

The vehicle is no longer designed around the assumption that a human might need to drive it.

We’ve gone from putting software inside a car to building the entire car around the software.

It is obviously very early.

But watching an empty Cybercab pull up to collect a paying passenger feels like one of those moments that will look completely normal in five years and quite ridiculous today.

👷3. Builder’s notes

The new ChatGPT app is unreal

The new ChatGPT app is super freaking good right now.

I think one of the coolest things is how sharp the voice features and transcription are.

It’s really quick and super accurate.

The app itself is super zippy and I love the translucent nature of it.

See Nate Herk going through a demo with the new Astra model

Opus writing sucks

For all the progress in reasoning and coding, AI writing is still mostly shit.

Claude Opus is probably the clearest example.

It takes normal text and then turns it into what can only be described as robotic slop.

See excerpt below (I got Claude Opus to give me a brief on its own writing and how much it sucks).’

“Opus writing sucks.

Controversial, and I'll defend it.

The models have gotten dramatically better at reasoning, at code, at long agentic work. The writing has gone the other way, or at least it hasn't kept up. Everything arrives in the same shape. Three-item lists. Tidy little takeaway at the end of every section. That specific rhythm you can spot from across the room.

It's not that any individual sentence is bad. It's that all of them are the same sentence.”

Like wtf is that. “Controversial, but I’ll defend it” reads so poorly.

One useful fix is the holy unslop skill.

plugins/pstack/skills/unslop/SKILL.md at main · cursor/plugins

It gives a set of principles for finding and removing the patterns that make writing feel generated.

Just feed this to Claude or Chat and ask how it can help you.

Or if you can afford it I’d try just writing with the Sol model, give it a bunch of stuff that sounds like you and it can give you a pretty good output.

Your skills need hygiene too

Skills are, as I’ve said before, one of the most useful parts of agentic AI.

They can also get pretty messy pretty quickly.

People install dozens of them, forget what they do and give them access to files, browsers and accounts without ever reading the instructions.

Claude Code shipped /skill-doctor. It looks at the skills you've installed and tells you which ones you never actually use and what each one is costing you in context.

Also for those of using skills and downloading from the internet these are my general rules.

My rough rules:

1/ Read the skill before installing it.

2/ Prefer skills from people or organisations you trust.

3/ Give it access only to what it needs.

4/ Remove duplicate skills that compete with each other.

5/ Review old skills occasionally and delete the ones you no longer use.

Be particularly careful when a skill contains scripts rather than just instructions.

Anthropic has now introduced security scanning for third-party skills and plugins on enterprise plans - which indicates a sign of the times.

🤔 4. Differentiated take: the paradigm shift of software

This week I made an animated video without knowing how animation software works.

I described the scene I wanted, gave Claude a visual reference and watched it build the animation.

When something didn’t look right, I didn’t open a manual or learn a new tool.

I just told it what to change.

It made me think about how strange software has been for most of computing history.

Using it meant learning how somebody else thought.

You had to know which application to open. Then which menu to click. Then which file format it expected.

Then why the bloody thing wouldn’t let you move an image three pixels to the left.

The user adapted to the software.

That relationship, I believe, is beginning to reverse.

Instead of learning how a program works, you describe the outcome you want and an agent works out how to produce it.

You don’t open Photoshop and learn layers. You ask for an image.

You don’t learn animation software. You describe how the scene should move.

You don’t navigate through six business tools to prepare a weekly report. You ask an agent to collect the information and prepare it.

This is much bigger than making existing software slightly faster.

It changes three things.

1/ Who can build

Until now, there has been a large gap between people who understand a problem and people who can turn that understanding into software.

The accountant understands the workflow but can’t build the application.

The developer can build the application but may not understand the accountant.

So they spend six months passing documents, mockups and tickets between each other, slowly losing information at every step.

Agents compress that distance.

The accountant can explain the outcome directly to an agent. That agent can work with coding agents, design agents and the systems where the relevant data lives.

The person at the top doesn’t necessarily need to know how every part works.

They need to know what good looks like.

If you can speak, you’re a builder.

2/ Where the value lies

Counterintuitively, this may make some existing software companies more valuable, not less.

Take Salesforce.

Salesforce is notoriously difficult to use. Companies employ entire teams of administrators and consultants simply to configure it and move information around.

In an agentic world, the interface matters less.

You tell an agent to update every opportunity, prepare the weekly pipeline, draft follow-ups and identify the deals that have stalled.

The agent works out which buttons to press.

But it still needs the customer history, permissions and workflows required to do the job.

Salesforce already owns those.

The value moves away from teaching humans how to operate the software and towards owning the underlying system of record.

The interface may disappear while the infrastructure underneath becomes even more important.

This is why I think a lot of value will accrue to the large incumbent software companies.

They already hold the context agents need.

3/ How we interact with technology

Software will also begin interacting with other software without requiring a human to sit between every step.

If I use an agent to manage my emails and you use an agent to manage yours, eventually much of the conversation is simply our agents interacting.

Or take calendars.

My agent asks your agent for a meeting. They compare availability, account for time zones, select a time and add the event to both calendars.

Neither of us opens an application. It happens agent to agent.

Humans set the outcome and the constraints. The agents handle the mechanical negotiation between systems.

Technical knowledge does not become worthless.

Someone still needs to understand the system well enough to verify it, secure it and fix it when something goes wrong.

But implementation knowledge becomes less scarce.

Judgment becomes more scarce.

Can you define the right problem?

Can you explain the outcome clearly?

Can you tell whether what it produced is actually good?

Can you take responsibility when it isn’t?

In the old world, knowing how software worked gave you power.

In the new world, the power may sit with whoever can express intent, supply context and exercise judgment.

What does this mean for us more generally?

At a paradigm level, I think we will increasingly move away from screen-based interfaces.

Within our lifetimes, manually copying information between applications may look as strange as filing cabinets and telephone switchboards do today.

Why did humans spend eight hours a day sitting in cubicles, moving information between rectangles on a screen?

What was any of this for?

We may look at it akin to working in fields 200 years ago.

Of course, liberation is not guaranteed.

Almost every productivity tool in history promised to give us more free time. We usually responded by expecting people to produce more work.

AI could remove hours of administration from our lives.

It could also fill those hours with even more administration.

The technology creates the possibility. What we do with it is still up to us.

Perhaps I’m overly optimistic, but I believe this could create a return to the things we actually live for. Rather than the things we have to do.

Craftsmanship. Sport. Art. Relationships. Doing things because they mean something to us, not because compliance documentation dictates it.

For most of computing history, we had to learn the language of machines.

Now the machines are learning ours.

That feels less like humans becoming obsolete and more like humans finally being placed back at the centre.

Carve outs:

• Bad bunny Monaco goes so hard - been bumping it

• Format finder is cool for finding trending content and hooks and editing videos.

• Really cool video on learning with AI

Until next week,

Alex