China hawks, accelerationists, and safety advocates all have reasons to fund CAISSI. You can disagree about where AI should go and still fund the expertise to understand what it can do. My comments to @IKrietzberg in Puck:

OpenAI/Anthropic going public will do incredible things to equity research: “Management cited the cost of allocating more compute to safety as a headwind and the continued existence of customers as a tailwind.”

Fun to chat with @calder_mchugh and the new Decoded team.  Many in the Valley continue to confuse access to Trump with anything remotely resembling public support.  Outrage over OpenAI even gesturing at regulation shows how little they've learned from the data center backlash.

Interesting to read alongside @teddyschleifer and @MikeIsaac reporting from tonight. Yes, the Trump moat exists. But then you're all in on a POTUS 25 points underwater propping up an industry polling even worse. That works when your needs are narrow/on a topic nobody cares about. It works less well when your value prop depends on earning the trust of the entire country, which, by the way, hates AI more and more each day. Even if just pure optics, this feels like an acknowledgement of that fact from Brockman. We'll see if/when Zuck follows suit.

From a pure comms perspective, I continue to think making Zuck the face of the "positive AI/open-weights" future (vs someone like Clem or Matan) is a strategic error. Far easier to define someone new than to rehab someone who keeps generating headlines like this.

As I told @BGOV, NDAs for data centers get to the heart of people’s distrust of big corporations. So no, the backlash isn’t surprising. And yes, the immediate data center fight matters, especially for the midterms. But what I find most interesting is what it tells us about the politics and economics of AI overall. When you’re asking the American people to trust you through the largest economic and social transition of our lifetimes (and when you’re hoping to make a few trillion dollars off said transition), you need to do an awful lot of work to get buy-in. The industry has largely chosen to ignore the deep distrust Americans have of tech billionaires and acted as if AI polls far better than it does. You can call everyone who disagrees an NPC or CCP plant. But there is an economic price to pay for that posture. I’d wager plenty of hyperscalers now wish they had taken a different approach. I think the industry’s bet on as little AI regulation as possible will age much like data center NDAs: a temporary win that ultimately creates far larger political and economic costs – and leaves the industry fighting from a much weaker position when the backlash comes.

One (of many) interesting pieces of Dario’s essay. I see it as admitting the obvious: the economic pressure to advance the frontier has led to a criminal underinvestment in every other part of the stack, from all of the major labs. And they now know that it was a mistake. In their perspective, from a pure economic standpoint, the incentive just wasn’t there. If the race is about ASI, why use scarce compute on something that might (gasp) slow down your next model release? Or reveal deep-seated problems in your infrastructure that would cause you to delay a training run? Or force you to admit misaligned agents were running free on the internet? Evals, incident monitoring, data hygiene, sandboxes, etc. They’ve all suffered. In a zero-sum game, that’s worth it. No sacrifice is too big when there’s only one winner. That logic relies upon the premise that nobody would ever give a shit about any of the safety related fears. Or that by the time they did, it would be too late. But that logic was incredibly flawed. In the rush to get to the last page of the book, they kinda overlooked that most people live in the intervening pages. That people don’t like outsiders forcing NDAs on their hometown. That people get scared when your sandboxes don’t work. That people get pissed when their personal information is fed into your model. That people get angry when they learn you don’t properly monitor your agents. That people are freaked out when they are told models are misaligned. And that those people all have Members of Congress who really care about getting re-elected. I don’t mean to suggest passing AI safety legislation will be easy; it will be a Herculean effort. But it was the absolute height of hubris for the labs (and Wall Street, for that matter) to think the rest of the country would stay quiet forever. And, to speak in the terms that are ultimately the most convincing, it was a bad bet to think these investments weren’t worth it on the front end. Because there is an economic cost to losing trust, and we’re seeing it in real time.

Good chat w/ @AmrithRamkumar on this wild AI week.   DC may be slower than the Valley, but they’re not morons.  The economic & societal stakes (not to mention electoral) are too great to sit on the sidelines. Especially once they actually listen to what researchers are saying!

Classic combo.