字号 ·· | 护眼
semafor

OpenAI的新定价层级凸显算力瓶颈OpenAI's new pricing tiers highlight compute bottleneck

点「原文对照」整页切到原文,或双击某段只看那段的原文。

尽管OpenAI的新工具dots在周二的年度开发者日上抢尽了风头,但其定价层级的调整更能说明该公司和整个人工智能行业的现状。OpenAI每月200美元的套餐现在只提供原来一半的token。而且还推出了令人瞠目的每月500美元套餐,让如今的柴油价格都显得不值一提。

While OpenAI’s new tool, dots, stole much of the spotlight at its annual developer day Tuesday, it was the change to its pricing tiers that says more about the company and the AI industry. OpenAI’s $200-per-month plan now comes with half the tokens. And there’s now an eye-popping $500-per-month plan that makes the price of diesel these days seem like nothing.

这种成本攀升再次表明,人工智能领域存在严重的算力瓶颈。提供始终在线的智能体助手工具将让OpenAI和Meta——后者最近推出了其可爱的智能体Muse——花费巨额资金,而且目前尚不清楚这些智能体能否良好运作。人工智能产品往往过度承诺,因为token根本不够用(而且成本太高),无法让它们发挥最大潜力。

The cost creep is another indication there is a severe compute bottleneck in AI. Offering always-on, agentic assistant tools will cost OpenAI and Meta — which recently launched its cutesy agent Muse — a ton of money, and it’s not clear yet that these agents will work well. AI products often overpromise because there just aren’t enough tokens to go around (and it would be too expensive) to make them capable at their maximum potential.

我还没有用过dots,但人们目前发现Meta的Muse会遇到付费墙和CAPTCHA验证障碍,需要用户自行排查解决。与此同时,我的Codex正在嗡嗡运转,为我做着酷炫而有用的事情,而我的dot能否替代我已经投入时间构建的一些项目,目前尚不清楚。Dots很可能像许多新的人工智能工具和功能一样:令人惊叹,但同时又还不够完善。

I haven’t yet worked with dots, but people are so far finding Meta’s Muse runs into paywalls and CAPTCHA hurdles they have to troubleshoot. Meanwhile, my Codex is humming away doing cool and useful things for me, and it’s not clear that my dot will be able to replace some of the projects I’ve already put time into building. Dots will likely wind up like a lot of new AI tools and features: completely amazing and, simultaneously, not quite there yet.

OpenAI完成其更宏大的任务也需要花费巨资。它最近在大约一个周末的时间里花费了约1000万美元来解决纳维-斯托克斯方程的部分数学问题。如果人工智能真的发展到如此高效、以至于能在圣诞节前消灭人类的地步——正如许多人所担心的那样——那么token的价格应该下降而不是上涨,解决数学问题的成本也应该低得多。

It costs OpenAI fortunes to do its more ambitious tasks, too. It recently spent about $10 million over roughly a weekend to solve part of the Navier-Stokes math equations. If AI was getting to the point where it’s so efficient it can take out humanity by Christmas, as so many fear, the price of tokens would be going down, not up, and it would cost a lot less to solve math problems.