2025 - 04 April
AI 2027
What 2026 looks like This
2023:
The hype is insane now. Everyone is talking about how these things have common sense understanding (Or do they? Lots of bitter think pieces arguing the opposite) and how AI assistants and companions are just around the corner.
2024
We don’t see anything substantially bigger.
(See articles like GPT 4.5 is no GPT-5 - by Gary Marcus)
If all this AI tech is accelerating GDP, the effect size is too small to detect, at least for now.
(Link: Why no “societal impact” two years after ChatGPT?)
2025
It turns out that with some tweaks to the architecture, you can take a giant pre-trained multimodal transformer and then use it as a component in a larger system, a bureaucracy but with lots of learned neural net components instead of pure prompt programming, and then fine-tune the whole system via RL to get good at tasks in a sort of agentic way.
Making models bigger is not what’s cool anymore. They are trillions of parameters big already. What’s cool is making them run longer, in bureaucracies of various designs, before giving their answers
Well this feels just eerie to me… Predicting that the research agenda would flip from scaling to fine-tuning and increasing inference, four years before it happened, seems incredibly hard. That is, it’s easy to make guesses about what good topics of research are (that’s half the job of every graduate school advisor). What is hard is to pick the dominant, most successful topics in a research area that literally millions of people are working on. He managed to do this by nailing what the DeepSeek breakthroughs would be (reinforcement learning on the same size models + mixture of experts) and what the “Reasoning” models would do (run for a long time at inference).
Thus, what would have cost a billion dollars in 2020 now only costs ten million.
See all the yelling about DeepSeek getting “the same level model” for “only $6 million”.
The generous take on this paper is that it digs into the question “what does it mean to care about truth?”, along the lines of the anthropic alignment research, and whether the models are hiding/faking intent.
But… from their “What is ChatGPT?” section, I can’t tell if they’re even aware of the reinforcement learning from human feedback portion of training. Since this is the part where, ideally, the models align with human preferences (which includes truth), the paper kinda looks like a long winded version of the familiar Greedy Reductionism fallacy, aka “they are just next token predictors so they have zero understanding of anything”.
One of the best Scott Alexander essays. A response to 2025 - 03 March#^4bc6a2 with much more depth. It has his humor at his best with snippets like
I sometimes imagine Heaven as a place of green wine, crimson seas, and golden mountains. Everyone goes there, good and bad alike. And if you still have enough innocence in your soul to enjoy things, then great, you’re in Heaven and presumably you have a good time. And if instead you’re one of those people who constitutionally hates everything, then you spend eternity writing thinkpieces with titles like “Can We As A Society Finally Shut Up About Golden Mountains?”
But it’s also a pretty thorough exploration of the goods and bads that may come from Ghiblification — that is, the ability to mass produce something 90% as good as what used to take a human a lifetime of training to create — as well as the ways that individuals can deal with this seismic shift:
You will see wonders beyond your imagination, nod, think “that’s a cool wonder”, and become inured to it. In the process, everything else that matters will wither away. If you get meaning from your job, the AI will take your job. If you get meaning from helping others, the AI will end poverty and cure cancer without your help.
https://www.experimental-history.com/p/the-bluetooth-test-and-other-keyholes