> You should. Maybe they have it open now just to destroy competition and then they close it. Now you have a closed system, with no alternatives.

What makes you think there would be no alternatives, though?

Put it like this: Kimi K3.1 is downloadable and can be run locally. (If you have a datacenter.) This is a very capable, ~2.8T parameter model.

Now let's say that every Chinese firm goes closed in six months or a year and jacks up pricing... What's stopping a bunch of enterprising upstarts from updating all 2.8 trillion parameters, extending them, distilling the model into very competent flash models, or applying targeted RL?

Once the principles are understood and there are models that can be pulled off the open internet, it doesn't seem that hard to spin up your own LLM. You just need some financial resources and a reason for doing it. And that would come into play if the Chinese kill every closed American provider and then go closed themselves.

▲

vlovich123 3 hours ago | parent | next [-]

Training is really really expensive and open models are not getting funded by VCs. Distillation also doesn’t improve performance. Until training gets exponentially cheaper so that you can do it for a few hundred grand to train a competitive model, open models are labs trying to flood the market to get market share / undercut more successful labs, not a sustainable strategy that will extend indefinitely necessarily. It’s like claiming Uber and DoorDash are fantastic because their prices are so low - once they eat the market the prices become higher than ever.

	▲	A_D_E_P_T 9 minutes ago \| parent [-]
		Sure, but once they release an open model, it's open forever. You can download it, share it with friends, make a torrent, whatever. So in this sense every open model is a gift to the commons. It's not quite like Uber. It's more like... well... imagine there are two computer operating systems, Gray and Teal. Gray is always slightly better, but it costs $50/user/month for a license. Teal is free. Eventually, when both happen to be in version 9.0, Teal puts Gray out of business. Starting from version 10.0, Teal starts charging $100/user/month from a license. But nothing's stopping you from sticking with 9.0, which remains free forever. Perhaps what's "sustainable" is not clear, given that Gray was eventually dominated by Teal? In any case, versions 1.0-9.0 were, effectively, gifted to the public. We can appreciate (and even disapprove of) that commercial strategy while benefitting from its fruits.

▲

ericmay 4 hours ago | parent | prev [-]

> What makes you think there would be no alternatives, though?

There could be, they can actually come from open American companies too. Ask yourself why we aren't seeing a thousand or more American AI startups spring into life overnight to build "open" models. Follow the economics.

> What's stopping a bunch of enterprising upstarts from updating all 2.8 trillion parameters, extending them, distilling the model into very competent flash models, or applying targeted RL?

> And that would come into play if the Chinese kill every closed American provider and then go closed themselves.

What's stopping American or European tech companies from doing this today? It's economic and technical capability. I'm very suspicious of any suggestion that it is in any way not incredibly difficult to build and scale up LLMs. Maybe in the future it can be much easier, but in that case you run into a sort of tautology where it becomes easier but the leading companies continue to lead and do so by a significant margin because they're the ones that have the most talent and access to capital.

The premise seemed to be "an open China" versus a "closed USA". I'm not sure that's a great premise. If China can create open-weight models so can the US. If the US can close their models (as clearly evidenced by the article) so can the Chinese.

	▲	A_D_E_P_T 18 minutes ago \| parent [-]
		> What's stopping American or European tech companies from doing this today? It's economic and technical capability. I'm very suspicious of any suggestion that it is in any way not incredibly difficult to build and scale up LLMs. Seems more economic than technical. If the economics are covered, the technical aspects seem straightforward enough, and may perhaps be fully solved. I think there's proof of this in that there are a lot of different LLMs on the market. Even Xiaomi -- in-between making cars, e-scooters, backpacks, and home appliances -- spun-up an exceptionally capable LLM! (And one which is rather idiosyncratic. It's weirdly good at music composition and generation, which both Opus and Astra cannot do.) The economic case doesn't exist for VC-funded US corporations, nor does it exist for cash-strapped nonprofits. In China, all such ventures are likely funded by a government development bureau. If Chinese ventures kill expensive closed providers and universally go closed themselves, there's nothing stopping the US or Europe from copying that same playbook -- unless, as may well be the case, they're ideologically opposed to subsidizing strategic industries.