AI #176 Part 2: Plan B
Zvi (Don't Worry About the Vase) TheZvi ● Covered by 3 sources
Trump's outgoing AI advisor says a formal licensing system for AI models will never happen under this White House.
Based on reporting by Zvi (Don't Worry About the Vase), TheZvi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sriram Krishnan, who is on his way out as a White House AI advisor, told reporters this week that Donald Trump will never sign off on a formal licensing regime for AI. His reasoning: setting up an agency that requires "a team of lawyers before you can get a model out" would jam the gears of the industry. Fair enough, except the alternative he's describing isn't no regulation. It's regulation by mood, where the administration decides which companies get slowed down and which don't, using emergency powers and export controls instead of any written rulebook.
Krishnan also floated the idea of extracting equity from AI companies as part of this same ad hoc approach, and when asked whether a future Democratic administration might use the precedent his own team is setting, he said plainly that he doesn't think about future governments, only "this government and this moment in time." That's a remarkably narrow way to build policy that will outlive an administration by years. GPT-5.6, for what it's worth, shipped anyway, so the delay Krishnan worried about turned out to be a matter of weeks, not months.
Meanwhile there's a separate and maybe more interesting fight happening among people who actually track AI capability, not policy. Dean Ball's now-circulating theory is that a big chunk of the tech and VC world quietly assumed, back around 2023, that models would plateau somewhere near GPT-5.5 or Opus-4.8 level and then just get cheaper. If you believed that, dismissing frontier AI as a serious business or policy issue made total sense. Problem is, people have been calling models "good enough" since GPT-4, and they keep being wrong, three years running.
That disagreement maps onto what's being called the three pills: whether you believe AI can do what it already demonstrably does, whether you believe it'll keep doing more of it, and whether you believe it eventually does essentially everything cognitive. Reactions to the newest models split cleanly along those lines — some call a new release incremental, others are already asking what an army of slightly-better versions could pull off five years out. Nobody's actually disagreeing about the demo. They're disagreeing about the slope of the curve, and that's the argument that actually matters for regulation, not whether some agency requires lawyers.
My take — AI-written commentary, not fact-checked reporting
Ad hoc regulation always sounds pragmatic until you realize it just means whoever's in charge that year gets to pick winners without having to write anything down or defend it later — that's not restraint, it's discretion dressed up as deregulation. And the crowd insisting models have basically plateaued has now been wrong for three straight years running, which is the kind of track record that should cost you your priors, not just your Twitter mentions.
Read more about this at: Zvi (Don't Worry About the Vase)