Post

It's October of 2026 and I'm still seeing people say that AI is useless because it's just a next token predictor. Have all the criticism you want. But if you haven't noticed what kind of things a "next token predictor" is capable of given the right harness, then you're not paying attention.

26/10/26, 2:48 pm4 reposts73 likes

Roland Dreier

@rbd.bsky.social

Also it's not “just a next token predictor” because a few years ago people realized you could have a model generate a string of “next tokens” and then *feed that string back into the model*! And iterate that lots of times! (Not to mention “tool calling”!!)

Banda Bassotti

@orpach.neocities.org

I still find it remarkable that the creator of AI Dungeon of all things was the first to realise you could just stop GPT2's output midway, type a little more yourself, and have it continue. (plus some hidden context clues)

Citizen Platano 🇵🇷

@daniloc.xyz

motor is useless because it just spins in circle switch is useless because it’s just a gate transistor is useless because it’s just a switch microprocessor is usele

Derridon't

@derridont.bsky.social

When I was studying cognitive science and ML in the early 2010s the state of the field for what task proficiencues people thought would constitute "real" AI was... Pretty much all stuff LLMs can already do right now Correct next symbol prediction of novel sequences + arbitrary data is a *big* deal!

Derridon't

@derridont.bsky.social

It has been genuinely extraordinary to watch the vast chasm forming between the Motte where the goalposts for "real AI" started a decade ago, and the desperately retreating Bailey they have since been moved to in order to exclude each new LLM capability from counting as having reached them