And they've provided no technical explanation of how...
This may be an odd reference, but I remember this song and dance with Roomba Vs. Roborock. "Roborock stole our IP!!!" but the problem was that song was slightly believable in the very early days, then Roborock started releasing products far better than anything Roomba had, but the tune in the US never changed. Literally for years after Roborock's products were two generations better than Roomba, but somehow they "stole" that?
China has money, talent, and cheap electricity/superior infrastructure. The US's big AI built their products on stolen copyright material and social media scraping. China likely got a catchup boost via distilling, but recently a lot of their AI ventures have come fully online, and they've just throwing tons of compute/power at the problem, and we're starting to see legitimately competitive products without shortcuts/cheating.
But the US won't change its tune; and that's convenient because they don't really want to compete, but rather roll out protectionism and fear per normal.
Z.ai found a way to crack/overcome the cryptography that safeguarded Claude/Fable's thinking tokens in Claude Code, so Z.ai had the ability to peak into decrypted thinking tokens that Fable produced in Claude Code. Z.ai also shared their findings with other big LLM companies like Moonshot AI, which had 20,000+ accounts speed running through Fable to get their hands on as many high quality thinking tokens as possible for distillation.
Given the fact that we seen not any evidence beyond people "see, it one shot X game looks similar to Fable" (a lot of one shots look similar to other models).
You expect to see something from Fable its restrictions/fallback behavior popping up in K3. You do not filter out all that behavior. Especially on such a short notice.
For me the milk example, where somebody asked if some EU cheese's had type A/B milk type (they are allergic), triggering bio restriction prompts. That just shows you how hard it is to find all those restrictions because its not in obvious spots where Fable goes off.
Yet, not a single person has shown this behavior in K3...
The funny thing is, other Chinese companies can train on K3 when the open weight are released next week.
So even if K3 was or was not trained on Fable, they already have a alternative to get their models to the same level.
Distills are bad for the US economy because they compress margins that otherwise go to pushing the frontier. You'd see model progress dramatically slowed if US companies had to compete on distills.
> You'd see model progress dramatically slowed if US companies had to compete on distills
We’re already competing on distills—they exist. The question is whether America will actually compete or if it will lean into its populist tendency and cede the market globally to China.
Yeah, I don’t buy this is happening. I’m open to being wrong. But folks who believe that can put their money where their mouths are in terms for stumping up cash for the training.
> distills force you to move your compute off the frontier
I guess I don't know how you can use the latest models and feel that we are not on the verge of some kind of intelligence. Maybe it's your harness, maybe your problems aren't complicated enough. I find it bizarre to outright dismiss the possibility of AGI in this day and age.
Almost certainly not in the long term. If AI is a productivity enhancer, which I think we have evidence for, then the American economy’s heft (and continuing dynamism) get a tailwind. The minority of the new, leveraged economy getting hit will be a short-term painful adjustment, but it need not be anything more.
The distills are almost certainly bad for the US economy.
AIs promise is undercutting knowledge work, which is our primary export. Manufacturing left ages ago.
In theory, if OpenAI/Anthropic win then the US still wins since they’re headquartered here.
If open or Chinese models succeed, it undercuts our knowledge work market and that money starts flowing out to inference providers instead of in to US companies and their employees.
We do, we export almost twice the amount of goods. The margins on services are so much better that the majority of the profit from exports is from services. I can’t find a study that has traced export revenue to export profit, but the services sector floats around 50% margins and manufacturing is around 15%. That would ballpark goods export profit around $30B and services around $50B.
> America is the second-largest manufacturer in the world by a long shot.
America is the third most populous country in the world. Normalize by population and we’re pretty middling; America is 53rd in merchandise exports per capita (63rd in overall exports per capita). South Korea exports over double our manufacturing per capita.
> Where are most of these?
The most companies, or where the most tokens are going? If you mean “where are the most tokens going?” it’s probably not the US.
The best data I can think of is OpenRouters usage leaderboards, which probably don’t include enterprise customers, and that’s dire.
You don’t hit a US model until #7 and it’s Nemotron which Nvidia is serving for free. The first US model that actually charges is Opus 4.8 at #9 and it handles about a sixth the traffic of the top model (Mimo-V2.5).
There’s really not a ton keeping inference in the US on a moderate time horizon imo. Tencent has Hy3 up for free right now, all Chinese inference, and it’s doing double the tokens Nemotron and Opus 4.8 are doing combined. For free, via the API.
It might not be China, but their aggressive investments in power generation and green energy make it seem likely. They just need GPUs and they seem to be nearing the point where they can offer them based on the Hy3 inference.
The longer the US AI strategy continues to be to bet on openai/anthropic the bigger it will be the US economy impact. The US bounced back from the dotcom crash and the 2008 crash, it does not mean it will bounce back after an AI crash.
It's the exact parallel of how it'll play out for U.S. automakers: destined for irrelevance given their inability or unwillingness to compete against China in the EV market. Instead they pretend EVs are bad and cry for protectionist policies.
The sad fact here is that OpenAI and Anthropic are only looking for a max 1-year runway (Anthropic's is only a few months away) so that they can IPO and their VC investors cash out immensely. They do not care about AI safety or fairness, only to appease their VC overlords. But the public is being lied to and pitched this narrative that "open weight AI models are all bad".
Winning strategy.