every six months the internet decides we’re all about to be replaced.
first it was “AI can write essays now, school is over.” school is fine. then it was “AI can write code, developers are done.” developers are fine. then it was “AI passed the bar exam, lawyers are finished.” lawyers are billing more than ever. now it’s “AGI is around the corner, everyone is done.” everyone is fine.
nothing ever happens.
i don’t mean nothing changes. stuff changes. customer support got gutted. some writing gigs dried up. junior roles are tighter because a senior with copilot does the work of two. that’s real. but the gap between “some jobs shifted” and “civilization as we know it is ending” is so enormous that i can’t take anyone seriously who jumps from one to the other with a straight face.
the trick
large language models predict the next token. that’s it. that’s the whole thing.
you give it a sequence of text, it outputs probabilities for what comes next, it picks one, repeat. it does this absurdly well. well enough to fool people into thinking there’s something thinking behind the curtain. there isn’t. it’s a very sophisticated autocomplete trained on the entire internet.
i use these tools every day. i build with them. i know exactly how good they are and exactly where they fall apart. ask one to do something its training data covered and it’s magic. ask it to reason about something genuinely new and it either hallucinates confidently or just rephrases your question back at you. that’s not a bug. that’s the architecture. you can’t patch your way out of “this thing fundamentally does not understand anything.”
but try saying that on twitter. you’ll get 200 replies from people whose entire personality is being early to the AI revolution telling you you just don’t get it.
the “slow down” theater
so here’s the funniest thing that happened recently.
dario amodei, sam altman, and elon musk all publicly agreed they should “slow the frontier.” three guys who spent four years in a full-sprint arms race burning billions on compute, suddenly holding hands and pumping the brakes.
the media ate it up. “even the builders are scared.” responsible leadership. existential caution. very serious stuff.
come on.
they hit a wall.
scaling laws worked for a while. more data, more compute, more parameters, benchmarks go up. gpt-3 to gpt-4 was a real jump. but since then? you get slightly better formatting. slightly fewer hallucinations. slightly better at following instructions. the core limitations haven’t moved an inch. it still can’t reason. it still can’t verify its own output. it still doesn’t know what it doesn’t know.
the best models right now are already near-perfect at what LLMs can do. there’s barely any room left. the problem is that what LLMs can do and what these companies promised are two completely different things.
so what do you do when you raised $100 billion on a promise of superintelligence and the curve is flattening? you can’t go to investors and say “hey, the architecture we bet everything on might not get us there.” you’d lose half your valuation before the sentence is finished.
instead you say “we’re so powerful we need to slow down for safety.” genius move. reframes a technical ceiling as a moral choice. “we could go further, we’re just choosing not to.” keeps the mystique alive. keeps the money flowing. buys time to figure out if there’s actually a next step or if this is it.
“slow the frontier” isn’t responsibility. it’s a press release for a product that stopped improving.
same movie, different decade
the AI doom cycle isn’t new. it’s not even creative.
the printing press was going to destroy memory and make people stupid. radio was going to end reading. television was going to rot brains. the internet was going to collapse society. social media was going to end democracy. crypto was going to replace money. every single time: panic, breathless predictions, think pieces about how nothing will ever be the same.
and every single time: the technology found its level. some things changed. most things didn’t. the people screaming loudest moved on to the next thing to scream about.
AI is a tool. a really good one. calculators didn’t replace mathematicians. photoshop didn’t replace photographers. excel didn’t replace accountants. the good ones adapted and got faster. the ones producing commodity-grade work got squeezed. that’s what’s happening now and it’s exactly as dramatic as it sounds, which is: not very.
the tell
watch what these companies actually ship versus what they promise.
they promised autonomous agents that run your business. they shipped chatbots. they promised AGI. they shipped autocomplete that sometimes writes a decent function. they promised a revolution in science. they shipped a better search bar.
every product they deliver is useful. not one of them is what was promised. and the gap between “useful tool” and “superintelligent system that makes humans obsolete” isn’t closing with another training run.
the doom narrative is marketing. “we need regulation because AI is too dangerous” sounds a lot better than “we need more runway because we don’t know what to build next.” saying your product might end civilization is the most effective advertising campaign ever created. it implies a power that doesn’t exist.
nobody who actually works with these things daily is panicking. it’s the consultants, the thought leaders, the people who need you scared so they can sell you the solution. the ones who changed their linkedin title to “AI strategist” eight months ago and now need you to believe the robots are coming so you’ll buy their course.
the robots aren’t coming. the autocomplete got really good. that’s the whole story.
nothing ever happens.
— alex
the loudest predictions come from people selling tickets to the future.