Astra, fly brains, & World Models

TLDR;

Welcome back! This week, we have a long but fascinating stream of news in research, markets, products, and everything in between. From the launch of GPT-6 Astra to a 1:1 mapping of the male fruit fly’s brain, passing through AI-generated 3D worlds and proposals to ban AI, there’s a little bit of everything in this newsletter issue.

Enjoy!

THEWHITEBOX

OpenAI’s Astra is Here. It’s Big, but not for the reasons you think

This is a seminal moment in the industry, and this release deserves its own piece. In the next Leaders segment, not only will we go into great detail to cover how Astra actually works and all the misunderstandings I’m seeing, but we’ll analyze how this intertwines with the recent HuggingFace and German wiki cybersecurity incidents that may tell us that, in some way, the world has already changed. Here, I’m laying out the key things you need to know for now, but rest assured there’s a lot more to talk about in my next newsletter.

It’s finally here. OpenAI’s GPT-6 model, Astra, is generally available. And if benchmarks still mean anything, it’s the best model in the world, and it’s not particularly close, setting record after record.

Perhaps the most striking one is how Astra completely saturates ARC-AGI 3, achieving a 99.9% score using OpenAI’s harness. But the most impressive part is that François Chollet, the benchmark's creator, claims Astra is more “action efficient” than humans, meaning it takes fewer steps to solve than the humans running the tests.

This is a crucial milestone because, while AIs are nowhere near the same power-efficiency ballpark as humans, they are becoming as efficient at problem-solving/exploration as we are.

Another very interesting characteristic of this model is that it has become so powerful on a forward-pass basis (the quality of each individual prediction) that, for most tasks, using high thinking thresholds is probably massive overkill.

Source: TogetherAI

But what's become clear is that standard ways of analyzing models no longer cut it. For parsing emails or searching the Internet, the outcomes you’re going to get will rarely feel like step changes.

Instead, we need to really push them to new tests; reshuffling the same tests again and again makes little sense. Luckily, people are pushing Astra to pretty extraordinary lengths:

Having Astra use Paint to draw the user based on an image

Making it play a browser piano using computer use at the correct speed (this one is pretty remarkable)

A game rendered on a browser using 3D software Blender

one shotting a 3D animation of the Codex Micro hardware from a simple prompt an image

And many others.

Whether this progress will translate into greater utility is another story. This industry still conflates “cool” with “useful.” Is the Paint example cool? ‘Fohshure’, as Macron would say. Is it something that will make all enterprises adopt Astra tomorrow? Obviously not.

But what makes this release “different” is that it has freaked out many researchers who think we’re moving in the…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.