3 ms·

When China too stops publishing weights you'll never have anything better than K3 / the current SOTA. Your choices for future frontier models will be: use USA A

When China too stops publishing weights you'll never have anything better than K3 / the current SOTA. Your choices for future frontier models will be: use USA API or use China API. Even now open weights are perpetually 6-12 months behind.

A_D_E_P_T 1mo ago
> When China too stops publishing weights you'll never have anything better than K3 / the current SOTA. Hasn't happened yet. Might not happen at all. Also, there's a lot that can be done with K3. > Even now open weights are perpetually 6-12 months behind. This is a very commonplace opinion yet I don't believe it's genuinely true. K3 is hardly much worse than Fable and can surpass it on certain tasks; they read as different, not as one that's 12 months behind the other. And don't forget that access to Fable is highly constrained. As for small/local models: The leading edge is open by definition, and hardly much worse -- if worse at all -- than small models like Luna or Haiku (lol).
- HeatrayEnjoyer 1mo ago
  It almost can't not happen. No one is going to release highly powerful near-future models for the same reason no one releases how to make physically compact but megaton yield fusion devices (which even today is still not publicly known how to do). Any state, especially the two present superpowers, has little to gain and much to lose by doing that. A model that fits on a laptop and is literally Skynet-level capable would not be released by any rational actor, which China is. (Whether someone else would make and release one is another matter.) China doesn't want Uyghur groups taking down their infrastructure, eventually it's not about the USA at all. -- K3 is almost, but not quite, competitive in August 2026, against a model that completed training last winter. There is still a gap.
  - A_D_E_P_T 1mo ago
    K3 is highly competitive. They're essentially tied here: > https://www.together.ai/blog/kimi-k3-vs-claude-fable-5-on-deepswe-cost-and-coding https://www.together.ai/blog/kimi-k3-vs-claude-fable-5-on-de... And they're very very close here: > https://artificialanalysis.ai/models https://artificialanalysis.ai/models And that's if you don't take cost and guardrails into consideration. If you do, K3 beats Fable hands down. That aside, I'm not an insider, and I can't really say when K3 completed training. What we do know is that it released to the public in mid-July, about a month ago. Fable 5's initial botched release was in mid-June, and it was re-released in early July. > No one is going to release highly powerful near-future models for the same reason no one releases how to make physically compact but megaton yield fusion devices So there's essentially an open Mythos-class model right now. Near-Mythos-class if you really want to quibble. Maybe past some threshold such models won't release to the public at all, or will release only in closed form, but we cannot predict when that threshold will be crossed. Fortunately for the world, it remains politically advantageous for China to support open model development. It's a winning strategy, especially in the long-term. Besides, I don't think a Uyghur Underground would be able to acquire the hardware to run a K3-class model, to say nothing of a Skynet-class AI.
JSR_FDED 1mo ago
So these countries will be able to use AI from multiple providers, which means they'll benefit from competition?
roenxi 1mo ago
If China stops publishing weights, that would suggest the US has been knocked down to 2nd place and it is likely that they will rediscover the "Open" in OpenAI for strategic reasons, or something like. It's reasonably safe bet that there will always be someone on the planet with the means and interest to commodify the models. They don't seem very hard to make; we're swimming in options.
- ben_w 1mo ago
  > It's reasonably safe bet that there will always be someone on the planet with the means and interest to commodify the models. They don't seem very hard to make; we're swimming in options. I don't think that is a safe bet; right now we're all benefitting from investors who disagree about who is going to win a monopoly, with each AI stock priced as if it will be the winner and the market collectively therefore priced several times higher than the maximum* return. When the investment bubble pops, I think there will be too many burned fingers to keep training new models. If we're lucky though, it may lead to someone burning weights onto hardware, which has the potential to significantly reduce the energy required per token. * currently possible return at least. I don't think any of the investors are actually pricing for the possibility of an actual technological singularity. Even Musk's rhetoric about this, where he opines about getting humanity to Kardashev type II, the numbers he attaches to this are orders of magnitude too small.

A_D_E_P_T 1mo ago

> When China too stops publishing weights you'll never have anything better than K3 / the current SOTA. Hasn't happened yet. Might not happen at all. Also, there's a lot that can be done with K3. > Even now open weights are perpetually 6-12 months behind. This is a very commonplace opinion yet I don't believe it's genuinely true. K3 is hardly much worse than Fable and can surpass it on certain tasks; they read as different, not as one that's 12 months behind the other. And don't forget that access to Fable is highly constrained. As for small/local models: The leading edge is open by definition, and hardly much worse -- if worse at all -- than small models like Luna or Haiku (lol).

HeatrayEnjoyer 1mo ago
It almost can't not happen. No one is going to release highly powerful near-future models for the same reason no one releases how to make physically compact but megaton yield fusion devices (which even today is still not publicly known how to do). Any state, especially the two present superpowers, has little to gain and much to lose by doing that. A model that fits on a laptop and is literally Skynet-level capable would not be released by any rational actor, which China is. (Whether someone else would make and release one is another matter.) China doesn't want Uyghur groups taking down their infrastructure, eventually it's not about the USA at all. -- K3 is almost, but not quite, competitive in August 2026, against a model that completed training last winter. There is still a gap.
- A_D_E_P_T 1mo ago
  K3 is highly competitive. They're essentially tied here: > https://www.together.ai/blog/kimi-k3-vs-claude-fable-5-on-deepswe-cost-and-coding https://www.together.ai/blog/kimi-k3-vs-claude-fable-5-on-de... And they're very very close here: > https://artificialanalysis.ai/models https://artificialanalysis.ai/models And that's if you don't take cost and guardrails into consideration. If you do, K3 beats Fable hands down. That aside, I'm not an insider, and I can't really say when K3 completed training. What we do know is that it released to the public in mid-July, about a month ago. Fable 5's initial botched release was in mid-June, and it was re-released in early July. > No one is going to release highly powerful near-future models for the same reason no one releases how to make physically compact but megaton yield fusion devices So there's essentially an open Mythos-class model right now. Near-Mythos-class if you really want to quibble. Maybe past some threshold such models won't release to the public at all, or will release only in closed form, but we cannot predict when that threshold will be crossed. Fortunately for the world, it remains politically advantageous for China to support open model development. It's a winning strategy, especially in the long-term. Besides, I don't think a Uyghur Underground would be able to acquire the hardware to run a K3-class model, to say nothing of a Skynet-class AI.

JSR_FDED 1mo ago

So these countries will be able to use AI from multiple providers, which means they'll benefit from competition?

roenxi 1mo ago

If China stops publishing weights, that would suggest the US has been knocked down to 2nd place and it is likely that they will rediscover the "Open" in OpenAI for strategic reasons, or something like. It's reasonably safe bet that there will always be someone on the planet with the means and interest to commodify the models. They don't seem very hard to make; we're swimming in options.

ben_w 1mo ago
> It's reasonably safe bet that there will always be someone on the planet with the means and interest to commodify the models. They don't seem very hard to make; we're swimming in options. I don't think that is a safe bet; right now we're all benefitting from investors who disagree about who is going to win a monopoly, with each AI stock priced as if it will be the winner and the market collectively therefore priced several times higher than the maximum* return. When the investment bubble pops, I think there will be too many burned fingers to keep training new models. If we're lucky though, it may lead to someone burning weights onto hardware, which has the potential to significantly reduce the energy required per token. * currently possible return at least. I don't think any of the investors are actually pricing for the possibility of an actual technological singularity. Even Musk's rhetoric about this, where he opines about getting humanity to Kardashev type II, the numbers he attaches to this are orders of magnitude too small.