I think overall, the biggest losers in this are the academic and public research community. They don't have the funds to compete with the industry and they have to use closed source models for their research, if they want to stay relevant.

Also, in regards to this sub-topic, that reminds me of another thing I've been wondering about. Some of these top American universities have huge endowments, and lots of serious talent, so, when we see colleges like Harvard (50+ billion annual endowment), Stanford (20+ billion annual endowment), MIT (20+ billion annual endowment), that's such huge amounts of money, with ~80% of it being pre directed to specific uses by the people donating the money, and 20% being general-purpose, that one way or another by a few hundred million of the pre-directed stuff being pointed at training SOTA local models or a few hundred million of the general-purpose money (either way in the many billions in both categories, each year) could use a relative small (borderline "chump change") amount of their money to do full, frontier-level training runs (could even do it on their own hardware if they wanted, given how much money they get each year, let alone if they just do it via Trainium or whatever).

I wonder if we might get some actual strong open weights models from Universities like MIT or Harvard or Stanford. With them making them for their own use, and then also sharing it openly, too, since they are a University, so everyone else gets to use the models as well, as a nice side effect.

Maybe with how strong AI is getting, they'd get nervous about doing this, since maybe they'd be worried that the rest of the endowment would stop donating if they were angry about the ~1% of it getting used to build AI that will take the jobs of their sons and daughters when their sons and daughters graduate, and so they wouldn't want to risk the other 99% of their endowment by using 1% of it to make super strong local AI models?

But in terms of just raw money and capability to do it, I think they'd be able to do it. And not just one university either, but several of the top few (especially the ones I specifically named).

Although I tend to agree with you (to me it seems like only Nvidia (and I guess AMD, Intel, Samsung, Micron, SK Hynix, etc, the hardware companies basically) have much real incentive to release open weight models to the public), I am curious what your thoughts are on why Google released the Gemma models (inlcuding Gemma4 just recently, with the 31b dense models being strong enough to cut into things fairly significantly, like, it wasn't just some useless toy, it can do "real things" pretty significantly and cannibalize stuff a fair bit), and OpenAI releasing OSS 120b a few months back.

You were saying that if zAI or Moonshot were in the shoes of American frontier labs like Google or OpenAI they wouldn't be open weighting models, but, we've seen Google and OpenAI openweighting some fairly significant models. Not 1T models, to be fair, but still, these weren't just little 2b or 4b toys. They were strong enough relative to frontier models at the time of their release that it would potentially fit to at least some degree into your argument of them not having a good reason to do so, right?

Personally I am a bit confused by it and not sure why they do it, tbh. I am sure they have some reasons, but, yea it does seem a little weird, since they definitely are not doing it to be nice/for charity or anything like that. They are for-profit, and have some actual hard reason for why they are doing these things, but seems like it must be some convoluted or not so obvious reasons I guess (for the American frontiers that are doing it I mean. The Chinese side might have significantly more obvious reasons by comparison).

I saw an interesting youtube vid recently that was hypothesizing about this exact topic, in regards to Google and why they would release Gemma/Gemma4, btw. Curious what you think about that guy's theory about the reasoning behind it.