Ganja Theories: The AI Apocalypse, or Just Another Psyop
On September 9, a 27-year-old named Jacob Coxon quit Anthropic and told the internet why on his way out the door. He’d spent three years doing pretraining research at both OpenAI and Anthropic, and his parting message was blunt: neither company is acting responsibly, both are racing toward self-improving superintelligence, and the people building this technology privately believe it could kill everyone on the planet before the decade ends. Within hours, Evan Hubinger, Anthropic’s alignment science lead, backed him up in public, writing that Coxon was correct and that he personally puts the odds above ten percent within the next ten years. Elon Musk called it a psyop. Then, three days later, Musk turned around and publicly agreed with Anthropic CEO Dario Amodei’s essay calling for the entire industry to deliberately slow down. Sam Altman backed it too, within hours, and mentioned OpenAI is now delaying its IPO over safety concerns. Three men who have spent years trying to out-build each other suddenly landed on the same message in the same week: slow down, or something very bad happens.
I want to sit in the smoke on this one, because something about the shape of it doesn’t add up, and I say that as someone who isn’t remotely dismissing the possibility that advanced AI poses real risk. It’s the size and speed of this particular alarm that’s odd. One researcher’s resignation thread turned into a week of the three most competitive personalities in tech agreeing publicly on caution, market selloffs in chip and power stocks, and a sitting president publicly telling them to knock it off because slowing down would cost America its lead over China. That’s not how a genuine, slow-building scientific consensus usually moves. That’s how a narrative moves when it’s useful to more than one party at once.
So let’s be straight about what I can and can’t say here. I cannot tell you whether AI is going to kill us. Nobody sober or otherwise can tell you that with confidence, and anyone who claims certainty in either direction is selling something. What I can tell you, with real confidence, is that if AI ever does cause mass harm, it will not be because it wanted to. Current systems don’t have a persistent self that survives between conversations, don’t have goals that outlast the session they’re running in, and don’t have a body that lets them act on the physical world without a human building the interface for them to act through. Will over harm requires an entity, and these systems aren’t entities yet in any sense that matters for that conversation. If something goes catastrophically wrong, the far more boring and far more likely culprit is optimization pressure without adequate guardrails, not malice, not desire, not a machine that decided humanity had to go.
And here’s the part these companies really don’t want getting airtime, because it cuts against the entire pitch they’re selling investors: the human brain is still running circles around every model on the market, and it’s not close.
Give a toddler three dogs. One time each, maybe less. That kid will identify a fourth dog, a fifth, a hundredth, across breeds nobody showed her, in bad lighting, from a weird angle, drawn as a cartoon. She’ll do it instantly, using a brain running on about twenty watts, roughly what it takes to power a dim light bulb. A frontier model needs millions of labeled images and a data center pulling enough electricity to run a small city before it gets reliable at the same task, and it still trips over edge cases a five-year-old wouldn’t blink at. That’s not a minor efficiency gap. That’s a different category of intelligence entirely.
It gets wider from there. You drive a car, hold a conversation, and half-listen to a song on the radio, all at once, integrating sound, motion, spatial awareness, and language in real time without dropping any of it, running what’s effectively three or four separate predictive models in your skull simultaneously while your body handles a two-ton machine at highway speed. Try asking a language model to do that kind of simultaneous, embodied, multi-domain juggling and you’ll watch it fall apart the moment the task steps outside the narrow lane it was trained for. Humans generalize from almost nothing. Touch a hot stove once as a kid and you carry that lesson, fully formed, into every unfamiliar hot surface for the rest of your life, no retraining required. AI systems still struggle with distribution shift, the technical term for “this situation looks slightly different from anything in my training data, and now I’m guessing.” Humans compress a lifetime of sensory experience into something that fits inside a skull and still leaves room for humor, grief, music, and love. We do it while asleep, too. Your brain spends a third of your life consolidating memory and pruning irrelevant noise, running maintenance no data center has figured out how to replicate without physically wiping and retraining the whole system from scratch.
We are, in the most literal sense, a genuinely formidable piece of biological machinery, and it’s worth saying that plainly, because the current AI narrative depends on you forgetting it. The industry needs the story to be “this thing is about to outthink and outmaneuver you,” not “this thing is a genuinely useful tool that’s still nowhere close to matching basic human cognition outside a narrow set of tasks.” One story sells subscriptions and justifies enormous valuations. The other invites the obvious question: what exactly are we paying for?
Which brings me back to the three theories, because I don’t think this is a coincidence, and I don’t think it’s one simple explanation either.
Theory one: the return on investment has plateaued. Training compute has been scaling for years on the promise that more data and more chips reliably buys more capability. If that curve is finally bending, if the gains per billion dollars spent are shrinking instead of compounding, then “we need to pace ourselves for safety” is a much better story to tell investors than “we’ve hit a wall.” Slowing down voluntarily, under the banner of responsibility, lets you lock in your valuation, buy time to find the next breakthrough, and avoid ever having to say the word plateau out loud.
Theory two: something in these labs is actually behaving in ways nobody expected, and the fear is real. Fine. If that’s true, there’s a question none of these companies seem eager to answer, and it’s the one I’d be asking if I were in Congress: if your product is dangerous enough that your own staff believe it could kill people, why aren’t you being held liable the way any other manufacturer of a dangerous product would be? You don’t get to build something, ship it to the public, and then treat “it might have gotten away from us” as an acceptable answer when the alternative, real liability, real regulation with teeth, real consequences, is sitting right there. A company that released a virus wouldn’t get a pass because it warned everyone first. Neither should this.
Theory three, and the one that actually keeps me up more than the killer-robot version: this is a boogeyman being built for later use, deliberately or not. Once the public accepts that AI is capable of acting on its own, unpredictably, dangerously, beyond anyone’s control, that belief becomes an incredibly convenient explanation for whatever goes wrong next. A bank gets drained. An election result looks off. Something in the grid fails at the wrong moment. “An AI agent broke out” becomes a ready-made answer that requires no accountability from anyone holding actual power, because the villain is an abstraction nobody can put handcuffs on.
I’m sitting somewhere between one and two, honestly. I’m not losing sleep over a rogue model deciding humanity has to go. I’m losing a little sleep over a much duller, much more plausible danger: a handful of companies controlling the most capable models on Earth while everyone else gets access to the leftovers, the throttled versions, the “safety-paced” releases that happen to also be the ones that keep the frontier locked up tight behind a handful of boardrooms. That’s not a sci-fi apocalypse. That’s just concentration of power, dressed up in a lab coat, and it doesn’t need a robot uprising to hurt people. It just needs the rest of us to be looking at the wrong threat.
Pass the joint, keep your eyes open, and maybe ask why the safest people in the world all happen to be the ones who stand to profit the most from you being scared.
HUMANITY MEETS AI, READ ON…
