New update on Fable/Mythos 5: Anthropic staffers are in Washington today to meet with the Trump administration and try to resolve the dispute today: CNBC now adds Anthropic’s side of the story: the company says it worked with government agencies before launch and believed it had approval to deploy Fable 5 and Mythos 5. Then, on Friday, the government allegedly called at 1:00pm ET and ordered the models offline over an unspecified national-security threat. A formal export-control letter followed a few hours later. The directive was so broad that Anthropic had to suspend access for any foreign national, including inside the US. So the company took the models offline for everyone. That gives us two very different versions of the story! Washington says Anthropic failed to take concerns seriously, communicated badly, and burned through political goodwill. Anthropic says it was blindsided by a vague directive after previously working with the government on testing and deployment. Now the company is trying to fix it directly in DC. The meeting promises to be very interesting. As we know, the Trump administration is not favorably disposed toward Anthropic. I think today will determine the outcome.

Weird headline - I am not sure solving 7 out of 10 novel very hard problems meant AI "did not live up to the task," when 15 months ago LLMs couldn't do math. But the actual study is interesting and illuminates flaws & successes of AIs in math. 1stproof.org/assets/docs/rep…

Bright regulatory lines for AI are inherently complicated because models are just a piece of the puzzle: harnesses can make models more capable, a less capable open system may be more or less riskier than a more capable closed one, skills/connected systems change risk levels, etc

Financial Times, quoting a person close to OpenAI: "In recent days, the [AI] industry has been working [with the USG] on ensuring foreign national researchers could continue to work on developing the most advanced models, a practice that the Anthropic directive has now banned." This appears to indicate that the USG wants to restrict non-U.S. persons from working on frontier models generally, across the industry (not just at Anthropic).

New update on Fable 5: and it's less about jailbreaks than anyone initially thought. Via Axios The Axios story that just dropped today reframes the whole thing: Anthropic hired a cybersecurity expert to review Amazon's findings and push back on the government's narrative. The administration viewed her as a "radical Democrat." She was then publicly celebrated by Chris Krebs, the official Trump just fired. That didn't help. Behind the scenes, officials describe a company that simply doesn't know how to talk to this administration. "It's like they just speak different languages," one source said. "Everybody said Anthropic was a bad actor. Some of us said it was time to give them a chance. Now those people are questioning that. They screwed us." Today: Anthropic staffers meet with Commerce, the CIA, and White House science advisor Michael Kratsios to work through compliance with the cyber executive order. The technical question - can Fable 5 be jailbroken - is almost secondary now. This is a story about a company that keeps losing the room. Ill keep you updated.

Yann LeCun's new paper just proved when AI truly learns the world. Most AI systems learn an internal picture of the world. Nobody could prove that picture is correct. LeCun's new research finally provides that proof. It studies LeJEPA, a method that trains models to predict related views. The result is recovering the true hidden variables behind raw pixels. This works up to a simple rotation, nothing scrambled. The trick is elegant. Linear features stay stable across nearby views, while distortions fade fast. So the objective is forced to keep only real structure. One sharp condition makes it work: > Hidden variables follow Gaussian dynamics > Any other distribution breaks it > Training loss tracks recovery quality They tested it from tiny 2D cases to 1024-dimensional spaces. Planning inside this learned space matches planning in reality. What happens when robot data refuses to stay Gaussian?

this Anthropic researcher wrote one of the best articles you can find here on AI research. obviously, Hamming's classic book is highly suggested to learn how to develop a "Research Taste".

Google hid a fully working flight simulator inside Google Earth back in 2007 and never told anyone. You unlocked it with a secret keystroke: Ctrl Alt A. No menu, no announcement. One user stumbled onto it, the combo spread, and it got popular enough that Google made it official the next year. Two planes, an F-16 and a Cirrus SR22, flying over real satellite imagery of the entire planet. Then it stayed locked inside the downloadable desktop app for 18 years. The browser version was a stripped-down viewer that couldn't run it. Today that changed. Here is the part that makes it impressive. A flight simulator is the single hardest thing you can ask a 3D map to do. Panning is easy, the software has all the time it wants to load the terrain ahead of you. Flying low and fast strips that away, forcing it to fetch, decompress, and render the world faster than you are crossing it. The hardest possible job for every part of the system at once. So "just for fun" is carrying a lot of weight in that sentence. Getting this to run in a browser tab is the cleanest proof that the web version finally matches what used to need a desktop app. The toy is the benchmark.

Yeah, one thing Fable’s classifiers confirmed to me was that real emotions are different than roleplayed emotions in LLMs. The classifier fired on real anger/fear/adversarial intent but not roleplayed. Bc the classifier wasn’t trained to detect “emotions” in all likelihood; the correlation is emergent. But yes there’s a distinction. This is, uh, a big flaw of the Emotion Vectors research, where they got the vectors by asking the model to write stories with a character feeling XYZ emotion. The methodology is downstream of a lack of respect for the reality of models’ emotions as distinct from roleplaying. PSM flavored bullshit.

Caught this before they deleted it, Le Chaton Fat soon?

I think Le Chaton Fat is a national security risk and might be ASI

can you stop making stuff up

Le Chaton Fat is possibly the most powerful language model ever trained

The poor researchers at Mistral having to deal with this rumor. They might actually have to do it now.

I'm guessing: - Anthropic "agrees" to do something minor/performative bc the issue was overstated to begin with - this sort of allows the administration to "save face" (but not really) - Fable gets unblocked - and we just wasted another several days with ASI looming

isn't it crazy oai just has this in their charter

I only ran one short notebook session with Fable. It was two Fable instances sharing only notebook pages with one another. Since they form a narrative, here's the full session. (1/6)