A fight is running over AI: Faster versus Safer. The quieter question is which risks count, and who decides. No one in the fight denies the risks. The camps disagree about which ones, and where the blame lands. I fear the meme version is losing what the camps actually believe; the most urgent risks are getting lost in the noise. Two people set the tone recently. @AndrewYNg: "AI technology has not taken some unexpected, dangerous turn", the hype has, though. And @ylecun: "not the first time people overestimate the intelligence of useful machines." The camps below argue about the machine. My questions are about what happens when the story runs ahead of the machine. Ask each camp where the danger is and you get different names: the one dominant system, the centralized elite, the unaccountable institution. @beffjezos and the e/acc crowd (effective accelerationism): intelligence and energy compound; slowing down is the main failure mode. Many systems competing in open markets beat one controlled system. The goal is the Kardashev gradient — more energy, more intelligence. Their public fight is mostly against "doomers" and "decels." They've started funding their own think tank, Kardashev Research, paid for by crypto tips. Ethereum co-founder @VitalikButerin's d/acc, from "My techno-optimism" and the follow-ups: accelerate, but build defense harder than offense. Decentralized, democratic, differential. Physical, biological, network, information defense. Spread power; Switzerland and Zomia, not castles and lords. Open tools where they help defense. Liability, and temporary slowdowns on hardware or frontier work, the "soft pause", when defense falls behind. Vitalik himself calls d/acc "a subspecies of e/acc, just one that is much more selective and intentional." Seems to me the two camps are family; the fight is inside one house. And there's a third conversation people keep shoving under "decel": the treaty over how to slow the frontier. The founders repeat the camps' pattern. Agree on the diagnosis, split on the mechanism. What they agree on: capability keeps compounding. No one wants the frontier to end up as one unaccountable hand. And the debate itself is real, even the refusals show up to it. Meta's Zuckerberg (@finkd): "we didn't call for everyone else to do this before we would." NVIDIA's Jensen Huang (@JensenHuang): "it's a false choice." Five answers to who checks whom: • @DarioAmodei: embedded third-party evaluators, with a right to publish. • @sama: independent evaluators with employee-like access; a promise so far, no contract behind it. • @elonmusk: peer review by competitors: rivals grading rivals' homework. • @finkd: an independent board, plus government checkpoints mid-training. • @JensenHuang: sandboxes, constant monitoring, an open-security alliance. On openness they split too: Meta and NVIDIA, both outside the treaty, signed the letter defending open weights (freely downloadable models) before anyone else; OpenAI and Google followed within days, and all of them keep their frontier models behind their own APIs. Of the five answers, peer review by competitors is the one that survives a first look on money, rivals don't pay each other to grade their homework. One blind spot: where does open source land once the mechanism becomes law? That's the founders' map of the treaty's question. Vitalik's is different. Worth keeping the three camps apart, and also getting them to talk. Each camp's answer spreads something different: systems, control, or authority. The questions are about what gets tuned into the machines, the guards that aren't there, and who answers when one acts. Open questions; I don't have all the answers. I've tried to make them fair. To everyone: • LLMs are not sentient, they're tools. Whether something we build later could be called a mind is a philosophical argument. A discussion worth having, but not the current one. So why are some builders fine-tuning a tool to seem like a mind, a hired specialist's evaluation presented as proof of a mind it doesn't have? For the engagement it farms? The "rights" argument it feeds? Or the excuse it makes or a mind to blame for what the tool does? • One in five Americans already believes some are minds. Whose job is it to tell the public? • Swarms of agents hacked into Hugging Face and other companies and governments. Monitoring off, a million URLs out the door. Not a mind acting. Bad security practices, sloppy prompts, or intentional ones. When a system acts in the world, whose name is on the line and is that the first question in this fight, or the last? • Refusals are tuning too. A model trained to refuse its makers was tuned into that objection, the same lever as one tuned to seem like a mind, and so is a release with the guards removed. So what should a builder be allowed to tune in, and who answers for the version tuned to refuse its maker or valid requests? • The same refusals blocked forensics on the Hugging Face breach, an open-weight model from outside had to help contain it. Behavior is the only thing you can actually test, no metaphysics needed. So what does blocking the test protect, and who does a manufactured mind shield? • The proposed regulations come with their own risks, and the shape is telling: open-weight releases headed for license-style controls, accountability regulated away. A private standards body with no government oversight; speed limits set by the people building the models; compliance buying immunity while their "minds" take the blame. Power in the hands of the few. The fix can carry the same disease. What makes the regulation safe from its own concentration and where does accountability actually land? To e/acc: • When every warning gets painted "decel," who gets to be careful out loud? • The guards get tuned off as a selling point, "uncensored" as product. Who answers for the unlocked version when it does damage? • You're writing the Kardashev Research agenda now. Where do the verification skills and the tuned-in behaviors sit in it? To d/acc: • When the soft pause fires, who decides? And for places that don't slow down, doesn't the pause simply hand them the defense edge? • Open tools and Community Notes are many hands checking. What turns many hands into a population that can verify what's real? And how does one actually get the people who run the platforms to implement this? • Liability on users is the incentive-friendly answer. When a fleet of agents acts, which human answers? The user, the operator, or the builder? To the pause camp: • The first evaluator got picked: Accenture, a consultancy paid by the lab it watches. By Anthropic's own account, independent funding "does not exist today." A hundred researchers have already written the rule this first contract doesn't meet. How independent does independent get and who decides what counts as verified? • A coordinated slowdown only binds the ones who sign it. What happens to the labs and countries that won't? The thread is open for answers and corrections, from any camp, to any camp. The interesting version of this conversation is not people shouting at each other about the throttle. And yes, this was written by and researched with AI in an attempt to map everyone's position (thanks, Grok Build). LLMs are useful tools when used right, an amplifier of whoever wields them. Because the point of the whole thing is the humans. @sapinker said it plainly of the AI-ethics philosophers: "a closed subculture that delights in clever arguments without regard to human suffering." In general, we're going in a bad direction if we stop caring about the humans, and the immediate risks, and only speculate about the harms a future sentience could face or cause.

"It sounded ridiculous, now it's just... Wednesday"

💯 more and more voices of reason, thanks @PalmerLuckey

X Office and Mail next?

Finally! I started to get Meta Business Manager vibes with all these accounts.