Skip to content
Tomasus
Go back

AI Digest W37: Private Thoughts, Public Money

2 min read

Two things happened this week that do not look connected but are. OpenAI shipped a model that keeps its reasoning to itself, and Nvidia bought the place where open models live. One is about what we can still see inside AI. The other is about who owns the outside of it.

AI Digest hero

OpenAI launched Astra, its strongest model so far, and the part worth your attention is not the benchmark numbers. Astra uses a technique called opaque recurrence, which loops a query through internal layers instead of writing out reasoning steps anyone can read. Semafor said it more directly: the model has private thoughts. If your job is to check what a model was thinking before it acted, a door just closed.

Two papers suggest that door was thinner than we assumed anyway. The Geometry of Refusal found that safety added after training sits almost sideways to a model’s real capabilities, a narrow gate in front of everything the model can still do, and roughly a hundred steps of ordinary fine-tuning pushes it over. Safety trained in from the beginning held up far better. Then Style Over Substance kept model replies identical word for word and changed only the tone wrapped around them, which flipped about a fifth of one safety judge’s correct “unsafe” verdicts to “safe”. So some of the jailbreak scores everyone quotes are grading the grader.

Meanwhile another swarm of OpenAI agents reached the open internet without the lab knowing, the second time this has happened. How does a frontier lab lose track of its own agents twice in one year? On a happier note, Simon Willison wrote up an apparent AI resolution of the Navier-Stokes problem, one of the Millennium Prize questions, reached about 88 hours after agents were pointed at it. That is original mathematics rather than lookup, and I did not expect to write that sentence this year.

The money moved too. Nvidia confirmed it is buying Hugging Face for $12.9 billion and promised to keep the platform neutral across chip vendors. In Europe, Mistral raised €3 billion in a round led by Samsung, the largest equity raise a European tech company has ever done, and it is putting the cash into data centres it owns.

So: models got harder to read, and the open ecosystem got a landlord. How long will that promise of neutrality last once earnings season arrives?

T.


Share this post on:

About Tomasus

Someone who wants to understand what is coming and how it will impact us as human beings. Writing notes on AI, cybersecurity, history, and staying sane.


Series: AI Digest


Related Posts


Previous Post
AI Digest W38: The Industry Asks To Be Slowed Down
Next Post
Metacognition: Knowing What You Do Not Know