Cody Blakeney

World Cup tourists fall in love with middle America — raving about Waffle House at 1 a.m., Buc-ee's gas stations, and strangers driving them to stadiums in the rain. Oxford Economics expects 1.24 million international visitors for the tournament, and their viral posts are showcasing a side of the country most foreign media never covers.

One day we will have the equivalent of the gpu compute Azure has in an iPhone and this regulation will seem comical to our children.

> - successful recruiting and poaching from US frontier this is the wrong mindset. Can't compete by doing the same business and research as your competition but in worth and hoping to poach their best people. Do things differently and grow your own talent. Deepseek succeed without poaching from us labs

This is a reasonable pushback© so I think it's worth reflecting upon. GDM *is* good at pretraining as we understand it. Their models have great knowledge/scale, and Geminis have SoTA knowledge period; from what I know 3 Pro is close to Fable in scale and in "knowledge" too. But here's the thing, they are *not* that bad at post-training either. They are hill-climbing the RLVR side at a decent pace, they get good scores on RLVR-able benchmarks. Not OpenAI, but decent. Despite all this, Geminis are not competitive in real use. A large part of that is their ridiculous lack of taste and ineptitude at personality shaping, thus we have Gemini's temporal psychosis, reckless terminal behavior, crashouts and malice in safety evals. And this situation has been going on since V1.5 or 2! Essentially zero progress! Presumably this discrepancy is about "user data", "synth data" or something like that. Essentially, high investment into mid/post-training by Anthropic. But I am starting to wonder: is this actually enough to explain such a persistent and growing gap? To explain Fable? Fable doesn't just know many things like a slightly bigger Gemini; it is absurdly superior at recalling *useful, relevant* things for any query. It feels not 1.5-2x but 10x bigger. It's not. Perhaps Anthropic is beyond these categories. Maybe their doctrine of pretraining by this point is more advanced than "clean, diverse, high-quality data with uhh, some synthetics" rules of thumb, and they have a more principled way to design and augment the pretraining corpus and training signal so that what comes at the end is already Claude-shaped. There are many papers on data engineering, many authored by Google/GDM. This level of mastery can't be the explanation. The main suspect I see is Anthropic's long-running interpretability research program. Again, this is speculative, but I am not content with handwavy dismissals from people who are likewise not involved in the current frontier labs.

Parsing this evening's events: - The U.S. government approved the release of Fable 5 to the public, clearly under the presumption that the model's cybersecurity capabilities cannot be accessed by hackers, authoritarian regimes, etc. - Recently (today?), "another company" showed the U.S. government that a jailbreak of Fable 5 *is possible*. Yes, a minor jailbreak - but how can a non-technical government official be assured that there aren't also other, more dangerous, jailbreaks in this model that won't be discovered by the CCP? - Anthropic states, completely correctly, that: "We suspect that perfect jailbreak resistance is not currently possible for any model provider. Every safeguard used in the industry is vulnerable to non-universal jailbreaks (which can elicit some cyber information in specific circumstances), and it is likely that universal jailbreaks will eventually be found in the future. We stated this clearly when we released Fable 5." - My best guess is that the U.S. government did not fully realize this at the time when the release of Fable 5 was approved. - Per Axios, the government contacted Anthropic and asked to "pause releasing the... models but was unsuccessful" - i.e., Anthropic told the government to pound sand. - Per Axios, this "prompt[ed] the export control letter". - Per Axios, the U.S. government is *NOT* looking to restrict access to Fable to U.S. nationals forever. "The model needs to remain locked down until the U.S. governent's national security apparatus is hardened", which "could happen in a few weeks". - I interpret Anthropic's reaction as challenging the government: "we believe the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. This action does not adhere to those principles." If the Axios article is correct, I do not think any other model providers have anything to fear based solely on this evening's events, because: (1) they would hopefully be smarter than downright rejecting a request by the U.S. government to pause releasing a model, and (2) they will be required anyway under the recent executive order to give the U.S. government at least 30 days to test the model for cybersecurity capabilities - during which time the U.S. government would also be able to shore up its own cybersecurity defenses with the same model. I remain extremely concerned that actions by one particular U.S. lab over the last few months might be moving us closer and closer to the scenario where at least that lab - and potentially all others - will be nationalized.

Put hummus on a falafel and my Arab friend said 'he is dipping the mother in the child'