SITUATION BREWING: The Trump administration is considering plans for an independent regulator to evaluate the safety of frontier AI models, based on the model of securities regulator FINRA.
DAILY SITUATION RECAP:
Xi Jinping speaks at the 2026 World AI Conference in Shanghai. The speech from China’s paramount leader explicitly endorses open-source, accessible AI; global governance; and AI control and risk, and says China will make sure that “this fine steed of AI
Radiant's Tori Shivanandan: "Electrons are to Space Force what jet fuel is to the Air Force."
@torishiv, President and COO of @RadiantNuclear, on the reason America lost the ability to iterate on nuclear and the reactor going to Buckley Space Force Base 18 months from now.
IFP infrastructure policy director Will Poff-Webster on the housing shortage:
"We have the biggest shortage of smaller, cheaper houses in America. If you want to build a huge McMansion in the exurbs, it's not hard to do that. The problem is those houses aren't affordable to a
This is one of the most impressive and surprising parts of ROAD to Housing's journey to become law. By the time it was enacted, the bill included almost every bipartisan pro-housing bill anyone had introduced
It's so different from other congressional processes that whittle down
METR president Chris Painter on the alignment failure in AI coding agents right now and what AI 2040 proposes to do about it:
"In San Francisco, in this community of people who think about the future of AI, there's this idea that if the AIs soon can take over the AI research,
METR president Chris Painter on the specific classroom analogy he uses to explain how AI monitoring actually works:
"If you imagine that today you were like, ChatGPT, can you help me fix my car? And then you go and look at what links it read, and it read something about nuclear
Caleb Watney on the reason he thinks the SF view that DC is clueless about AI is completely wrong:
"There's been this vibe for a little bit in SF that DC is completely clueless. They're not going to pay attention. We're just going to build AGI in SF. I think the federal
Some bittersweet news to announce today:
I recently left @IFP to join @coeff_giving as Managing Director of Public Policy, where I'm building a new US AI policy team, overseeing the Abundance and Growth Fund alongside @mattsclancy, and managing CG's government affairs work.
.@deredleritt3r on why he thinks Thinking Machines Inkling is the American answer to DeepSeek:
“I am really glad they’re doing an open source model. The more different things we try out, the more likely we are to hit on an approach that ultimately works.”
“For Thinking
.@deredleritt3r on why Kimi K3 is the best open source model but still not close to the frontier:
“It is by far the best open source model I’ve tested to date. It is much better than GLM 5.2, which the timeline was in love with two weeks ago.”
“I kind of think of this model
A few thoughts on Kimi K3:
- The Moonshot team is absolutely cracked. Some of you might remember that K1.5 was released on January 20, 2025 - i.e., on the same day as DeepSeek R1, and it was ~just as good as R1. IIRC, the model wasn't open-source, which is why no one ever
SITUATION EXPLAINED: DeepSeek and Moonshot's models are mostly trained on 3-year-old, export-controlled chips.
• Huawei showed a real Atlas 950 SuperPoD at WAIC 2026 in Shanghai: 1 exaflop FP8, 2 exaflops FP4, 256 terabytes of unified memory, more than enough to serve Kimi K3's
SITUATION EXPLAINED: AI detectors have gotten really good, except at catching AI pretending to write like a specific human author.
@EpochAIResearch tested three leading detectors, Pangram, GPTZero, and Originality, on both AI and human text
• On plain AI-generated text from
We stress-tested some AI detectors and found that they rarely flag human text as AI-generated. But asking LLMs to mimic a specific author causes detectors to misclassify text as human-generated ~13% of the time. For scientific writing, false negatives rose to ~26%.
