As long as they're a few months behind, it makes sense to publish the models to compromise the competition financially. But once they're frontier, the incentive for this strategy will be gone.
Ofc, the counter argument is that China's central state is commanding these companies. Generally, I don't think the Chinese still believe in marxism themselves. But maybe China is opensourcing AI models as a way to free the "means of production" in a very literal sense. Maybe this isn't pragmatic but like religiously ideological.
For me, the open GLM is consistently outperforming closed Qwen Max, so I am not sure why they would not be "frontier" - they're the frontier of open weights. I haven't seen a Kimi release that beats GLM either this year.
Sure. Frontier as in they overtake Anthropics or OpenAIs top model at least once in major benchmarks as e.g. lmarena: https://arena.ai/leaderboard/agent. But I can see how you could classify GLM5.2 as frontier now already.
Yes. The Agent ranking on Arena needs more entries though - for example MiMo 2.5 (also open) was built for this and performs pretty well, but it is not assessed yet.
I just hope it will be open-weights xD
Likely, because the guy replying "won't take that long" is from Z.ai (and they still open-weights everything - GLM)
Let's hope zAI will continue to publish open weights. GLM5.2 is astonishing. Let's hope it's not the last one.
Have you read anywhere that there are plans to do otherwise?
No, but it's a constant fear in the community.
As long as they're a few months behind, it makes sense to publish the models to compromise the competition financially. But once they're frontier, the incentive for this strategy will be gone.
Ofc, the counter argument is that China's central state is commanding these companies. Generally, I don't think the Chinese still believe in marxism themselves. But maybe China is opensourcing AI models as a way to free the "means of production" in a very literal sense. Maybe this isn't pragmatic but like religiously ideological.
Or maybe all of this is overthinking it
For me, the open GLM is consistently outperforming closed Qwen Max, so I am not sure why they would not be "frontier" - they're the frontier of open weights. I haven't seen a Kimi release that beats GLM either this year.
Sure. Frontier as in they overtake Anthropics or OpenAIs top model at least once in major benchmarks as e.g. lmarena: https://arena.ai/leaderboard/agent. But I can see how you could classify GLM5.2 as frontier now already.
Yes. The Agent ranking on Arena needs more entries though - for example MiMo 2.5 (also open) was built for this and performs pretty well, but it is not assessed yet.
@BlokchainB Thoughts?
As they should!