I'd like a concrete list of "frontier labs", but don't want to exclude newcomers. I'll write a list below, but if there are labs that aren't on the list but plausibly could be considered frontier, I'll make a manifold poll and rule based on the markets decision.
Also please comment if you think I've missed one:
OpenAI
Anthropic
Meta
Amazon
DeepMind/Google
DeepSeek
Alibaba (NOTE: this used to say Qwen, but Qwen is the name of the model, not the lab)
Mistral
XAI (Elon musk lab)
Nvidia
Update 2025-02-07 (PST) (AI summary of creator comment): Preprints Included:
Preprints (e.g., arXiv) count as a valid form of releasing a paper.
The focus is on whether the big labs deem that an AI significantly contributed to the paper, rather than on traditional journal criteria.
1,000While this doesn't count, you might find it interesting: https://sakana.ai/ai-scientist-first-publication/
@zsig IMO it's unlikely that a frontier lab will risk their reputation on this, but place your bets/make your markets to prove me wrong
@MalachiteEagle Good question. I'm assuming the wording "release a paper" vs "publish a paper" implies preprints are included. Note that most publishers don't allow AI to be listed as an author because it can't take responsibility for authorship.
@MalachiteEagle Yes, preprints such as arxiv will count, for the reasons cited by @WilliamGunn . I'm more interested in whether the big labs think an AI significantly contributed to a paper, and less interested in journal politics
@BoydKane thanks 👍
A few suggestions based on LM Arena (https://lmarena.ai/)
Qwen seems to be a model name, I think the lab is Alibaba
StepFun
Zhipu
XAI
01-AI
Nexus Flow
Nvidia
Ai2
@TheAllMemeingEye I'll make the Qwen -> Alibaba edit, although I've never heard of the others except for XAI and Nvidia. I'm open to changing frontier labs to be defined as "any lab with a model in the top XYZ of models on lmarena.ai"? Maybe that's better, and less subjective than what I've currently got
@BoydKane yeah in this case I was just listing labs that had models higher up on the leaderboard than the best model from the worst performing lab you had previously listed (in this case Amazon)