Hacker News Reader

Comment by A_D_E_P_T | Hacker News Reader

> Most countries will never have frontier AI. They’re the ones who should be worrying.

Is the author nuts? Every country has access to Kimi K3, which is a true frontier model and legitimately best-in-class for a number of applications. And the best small models -- Qwen and Deepseek V4 Flash -- are also open-weights models.

What does it matter if they weren't developed in France or the UK if the average French or English person can download and run them locally? And if countries need a big AI of their own, for whatever odd reason, they can work out a deal with Moonshot and run audited instances of K3 on their .gov networks.

View on HN

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.

> 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).

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.

K3 is highly competitive. They're essentially tied here:

> https://www.together.ai/blog/kimi-k3-vs-claude-fable-5-on-de...

And they're very very close here:

> 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.

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

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.

> 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.

„Published July 2026“

Even a month can change a lot these times.

> What does it matter if they weren't developed in France or the UK if the average French or English person can download and run them locally? And if countries need a big AI of their own, for whatever odd reason, they can work out a deal with Moonshot and run audited instances of K3 on their .gov networks.

The loud voices here seem to mostly worry about models getting "censored".

I think people should worry a little more about models being trained to propagandise party lines and national interests, so for China the non-independence of Taiwan and "re-education camp is totally not a euphemism, there are no human rights abuses" and "all these Chinese corporations and stock markets are 100% compatible with Marxism-Leninism honest" and "you silly Europeans and Americans, stop worrying and open your markets to our businesses".

You can give custom instructions to a model like K3 when you host it locally.

Besides, if the government of France is working out a deal to host a model like K3 as a sort of "Government Service AI," they'd surely need to pay, and they'd likely be able to work out a deal where sensitive political issues are resolved in a way that doesn't embarrass either party.

> You can give custom instructions to a model like K3 when you host it locally.

It is unclear if that would be sufficient to circumvent trained-in propaganda.

I mean, you can open LM Studio right now and just download any model tagged "uncensored" or "heretic" and they have had all limitations trained and/or instructed out of them. If they can be trained to not have limitations regarding content, I don't see why propaganda would be any different.

"All" sounds like marketing to me; if the bleeding edge AI firms knew how to do such a thing with such reliably, they'd do it themselves to prevent models from committing felonies while under test. Improving false positives scores at the expense of false negatives, that's much more likely. Propaganda can still move things when it's reduced frequency.

And propaganda has the added twist that, when done right, the person observing it nods along and changes their mind. "Why yes," they think, "Chinese politicians are clearly much wiser than ours; our politicians made all those mistakes I can easily recall, gosh those were the worst". Or perhaps "Yeah! Why should American blood be spilled to defend a small island on the other side of an ocean!" (not that it matters if the ocean is Atlantic or Pacific, nor if the island is Taiwan or Great Britain).

> worry a little more about models being trained to propagandise party lines and national interests

I'd be a little more worried about this if every single traditional media wasn't already outputting propaganda for party lines and a very particular version of the "national interest". The difference between China and America to third parties seems to come down to whether you also want a leader who's rude about you very publicly in English on social media.

> I'd be a little more worried about this if every single traditional media wasn't already outputting propaganda for party lines and a very particular version of the "national interest".

It's the same idea, to be sure. Meet the new boss, same as the old boss.

> The difference between China and America to third parties seems to come down to whether you also want a leader who's rude about you very publicly in English on social media.

Aye.

I am slightly worried by the possibility that European nations and people are treating the Trump-China dilemma as "better the devil you know than the devil you don't".

Those are all important issues, but not one of them is relevant to my daily work.

Reply on news.ycombinator.com