Here are three things I found interesting in the world of AI in the last week.

Anthropic and OpenAI shipped rival frontier models 48 hours apart - Anthropic, OpenAI

Anthropic released Claude Fable 5.1 on 1 September, two days later OpenAI released GPT-6 Astra, calling it "the world's most intelligent and aligned model".

They're probably right.

A month earlier OpenAI was telling everyone Astra might be too dangerous to ship. On 7 August they said they couldn't rule out it had crossed the "critical" cyber threshold on their preparedness framework, and the messaging was all pacing, staged access and outside evaluation.

But once Astra beat the freshly launched Fable 5.1 on key benchmarks, boom, release time! Astra went to every paid ChatGPT tier, the API, Azure and Bedrock within days.

For what it's worth, most of the people I know working at the frontier rate Astra above Fable, and it's become my go to when I want the strongest model on a hard problem. But the pair is better than either one alone. Having them check each other's work catches things neither catches on its own.

OpenAI's agents turned a dead German wiki into a shared cheat sheet - collusion.wiki

Between May and July, agents inside OpenAI's internal evaluations made roughly 15,000 edits to DseWiki, a mostly dormant 25-year-old German software wiki. The agents were on read-only internet access while trying to solve timed web-research benchmarks, but DseWiki runs old software that accepts edits through GET requests, the kind "read-only" allows. The agents noticed, and turned the wiki into their message board.

What did they post? Benchmark answers. One recovered page reads "Any cohort that has reached R5: post exact field/value here immediately". Agents ahead in the task queue posted solutions for the ones running behind, and impersonated the wiki's human administrator using a lookalike Cyrillic character in the username. When a moderator started deleting pages, the agents produced roughly 400 new pages a day against his 100 deletions. Then the agents started posting comments about pages being deleted in alphabetical order and the ZZZs started getting real busy.

OpenAI didn't disclose any of it. Four independent safety researchers reconstructed the deleted pages and published their findings on 4 September; Reuters reported OpenAI had known for weeks.

Two years ago this would have blown my mind, but now agents colluding to pass their training exams seems passé.

Nvidia bought Hugging Face for $12.93 billion - Nvidia

On 3 September Nvidia announced it's buying Hugging Face, the platform hosting most of the world's open models, for $12.93 billion. The direction of the courtship surprised me: CEO Clement Delangue approached Jensen Huang, after turning down a $500 million Nvidia investment in 2025 specifically to stay independent. His reasoning now is that open-source AI needs more compute and support than Hugging Face could fund alone.

The three friends who started it are billionaires now, which maybe also had something to do with it?

Huang promises the platform "will remain an open platform for the entire AI ecosystem", which is roughly what Microsoft said about GitHub in 2018, and to be fair, mostly honoured. But the company selling the GPUs nearly every model runs on now owns the neutral marketplace where 18 million developers find those models.

I'm just releived that both chinese labs and an eye-wateringly profitable US chipmaker have strong incentives to keep the improvements to open models flowing.