agentic

openai's chief brain Jakub says ai’s bout to go from prompt puppy to full-on phd

currently: ai needs hand-holding—“please write this code,” “please analyze this chart,” like babysitting a genius
but in 5y: it’ll be doing full-blown research on its own, no babysitter
deep research tool already crawls and synthesizes info in mins—early prototype vibes
next step: give it more compute and let it tackle open problems solo

key sauce? reinforcement learning
pre-train = world model from data
RL = teach it how to think, trial/error + human feedback
they’re pushing RL hard—models now solve gnarly stuff like global remote dev scheduling with zero human hand-holding
open question: should pre-train and RL stay separate or merge into one big learning loop?

can ai actually reason? nah, not like humans
it doesn’t remember what it learned first or last
asks “why?” and it just shrugs in tokens
but Jakub’s bullish—future models might not mimic human thought, but they’ll find their own alien logic to think for real

tldr: ai’s not just eating jobs, it’s gunning for tenure