前沿 AI 实验室会迎来自己的 WeWork 时刻吗？
Could this be the Frontier Labs’ WeWork moment?
Gary Marcus 质疑前沿 AI 实验室的高估值：tokenmaxxing 热潮已退，中国开放权重模型引发价格战，OpenAI 收入预测存疑，加之失控智能体带来的责任风险和美国上诉法院新的版权裁决，使这些公司寻求的巨额估值越来越难以成立。OpenAI 已将 IPO 推迟至 2027 年，Anthropic 据称下月上市但尚未公开 S-1 文件。
In May, at the peak of tokenmaxxing, people were talking about OpenAI IPO’ing at a trillion dollars, Anthropic at two trillion dollars. Maybe that will still happen, but it probably shouldn’t.
Tokenmaxxing died, Chinese open-weight models forced price wars, there was some hinkiness around some of projections OpenAI’s revenue and the so-called “rogue AI” (really just bad software) fiasco hadn’t yet made the news. With the errant, and possibly criminal, nature of poorly controlled agents now front and center, potential liability is clearly becoming a major new issue. LLM’s might well someday be renamed Large Liability Modules.
A significant new copyright decision by a U.S. appeals court also just undermined the generative AI companies:
Against all of this, it’s harder and harder to fathom how the companies could really merit the enormous valuations they are seeking. This is perhaps part of why OpenAI has pushed back their OpenAI until 2027. (Anthropic is supposedly going to IPO next month, but notably still has not publicly released its S-1.)
§
Two things I read this morning, jet lagged in New Dehli, seem even more ominous.
The paywall below is your occasional gentle reminder that I work hard on this newsletter and appreciate paid subscribers.
来源：Gary Marcus：The Road to AI We Can Trust · garymarcus.substack.com