Ezana’s cynical HOT take 🔥🤨
The frontier labs’ efforts to “slow” AI is not only about safety.
If intelligence becomes a commodity like electricity, frontier labs risk becoming interchangeable utilities while their customers capture much of the value. Keeping their best models to themselves would help them avoid that.
I think their goal is to create a mechanism for delaying model release without risking loss of market share to other frontier labs. Restrictions that delay public access more than internal progress would preserve an advantage for the labs (and their chosen partners).
The safety concerns are legitimate. AI has real risks. However I call bullshit on safety being the whole explanation for an “AI” slowdown.
Consider two commercial incentives:
1. Make it harder for Chinese competitors to catch up.
Anthropic has accused Chinese labs of training competing models on Claude’s outputs (distillation). Withholding newer models limits that shortcut. If the Chinese are distilling models several generations behind what the labs use internally, Chinese model development will be significantly hampered, depending on how much those labs truly rely on distillation.
2. Compete against their own customers.
Imagine buying AI for drug discovery from a lab using a model five generations ahead of yours to develop competing drugs. The lab would still need capital and industry expertise, but you would be competing against your supplier using inferior tools. I think that’s the endgame envisioned by the frontier labs.
But there’s a problem. A lab that delays model releases will lose customers to rivals that keep releasing better models. The obvious solution is for frontier labs to coordinate delays. However, such coordination violates antitrust laws.
Anthropic CEO
@DarioAmodei has proposed a narrow antitrust waiver for safety discussions. Musk, Hassabis, and Altman agree.
I wonder why???? 🤔🤔🤔🤔
Maybe it’s because a safety exemption can easily become a route to restricting public access while the labs pursue those commercial advantages using internal models that are generations ahead of what the public has.
Am I wrong???
#AI
#AIsafety
@elonmusk
@demishassabis
@sama