Foreword
I had to resist making that first post an Abbott and Costello bit about Huang and Huang — no relation, except for the relation.
Chinese spy has “close relationship” with Nvidia CEO Jensen Huang
Investigative reporters from The New York Times say American spy agencies were tracking Alice Huang (a.k.a. Huang Le), a Chinese citizen, for her prominent role at Megaspeed, an Nvidia customer strongly suspected of procuring chips for Chinese firms in violation of U.S. export controls. This surveillance confirmed her role as a Chinese intelligence asset — and revealed her relationship with Jensen Huang.
Jensen Huang, of course, has a close relationship with President Trump. He has been seen as holding substantial sway with the president over AI policy and chip exports.
The President is said to have been briefed about the intelligence findings roughly a year ago and done nothing with them.
Sam Altman: “We are going to have to agree to some limits about how smart we let these single models go.”
This was added as clarification to his assessment that OpenAI’s math research pipeline “will probably work” on AI research.
He doesn’t say who “we” is or how this would be enforced, but I will keep encouraging lawmakers to call such bluffs and make this happen.
This interview, with John Collison on the Cheeky Pint podcast, also saw Altman express a preference for society to be joined by “a trillion AIs of slightly superhuman intelligence” over “one AI of one trillion IQ,” but I don’t know how one could get to that point without soon going well beyond “slightly superhuman.”
Altman says OpenAI has a robotics program it will put to work making data centers.
He also confirmed the rumors that OpenAI is holding more unpublished math results.
The media remains bizarrely silent about major math and science breakthroughs from AI
None of the major news outlets I monitor appear to have covered OpenAI’s Monday announcement of three discoveries that are each pretty momentous within their subfields.
That’s probably because they’re over almost all our heads and have been dropped with a minimum of human drama. We saw the same thing with the company’s release of hundreds of math papers three weeks ago.
One of the new findings, a disproof of the Exponential Time Hypothesis, caused a double take from the AI agent I had reviewing the press release about it: Like most humans, apparently, it had assumed the ETH was true. The disproof suggests avenues for further progress in areas long considered settled.
A new discovery involving quantum field theories seems likely to have implications both for physics and mathematics.
I see no sign that these breakthroughs are going to stop any time soon. This is worrying, because the science of developing stronger AI is unlikely to be radically more challenging than the math and science problems AI is already cracking. Monday’s results add weight to AI company claims about closing the AI self-improvement loop within the next year.
Engadget - OpenAI Announces New Discoveries In Mathematics, Computer Science And Physics
OpenAI - Working with academic communities to build scientific understanding
Fast Google release gives weight to its talk of accelerated, AI-assisted development
The new model, Gemini 4 Carbon, appears to be a true frontier model from a company that had spent a year well behind the competition. That only changed two weeks ago, when it released Argon, a near-frontier model. In the weeks leading up to this one-two punch, employees’ social media posts hinted that Google had entered an AI-driven improvement loop.
CEO Sundar Pichai, in an earnings call this week, explicitly said as much:
This rapid progress underscores our incredible model momentum. We are at the point where AI is helping us increase our training and launch velocity. I would expect the next twelve months to see continued acceleration because of RSI.
(RSI = recursive self-improvement, a term traditionally used to refer to a closed AI-developing-AI loop that humans aren’t meaningfully contributing to. It can reasonably be called “losing control on purpose.”)
It may be a while before Carbon can be independently assessed for real-world capabilities. As with frontier flagships from Anthropic and OpenAI this year, the company is doing a staged roll-out where cyber defenders get a window of exclusive access.
Google Blog - Gemini 4 Carbon: our next frontier model
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.