@seelffff: > reads papers on arXiv autonomously > finds and checks datasets on HF Hub > writes the training script itself > genera…

reads papers on arXiv autonomously finds and checks datasets on HF Hub writes the training script itself generates synthetic data if datasets aren’t good enough monitors reward curves, fixes failures uploads weights automatically

6 steps. 10 hours. 0 humans in the loop. took a 1.7B model from 10% to 32% on GPQA.

a junior researcher costs $60k/year and needs onboarding.

this costs compute.

one prompt. three agents. all running at the same time.

i gave Claude Code a single instruction: analyze my ML learning repo from 3 different angles in parallel.

→ Agent 1: map the full topic structure → Agent 2: inventory every code example and library → Agent 3: evaluate it as a learning resource - what’s missing, what’s assumed

it doesn’t run them one by one. it spawns all 3 simultaneously and compiles when they’re done.

most people are still using AI like a search engine. one question → one answer → next question.

this is a different thing entirely.

the actual potential:

→ agent 1 writes tests while agent 2 writes the feature → agent 1 reviews PRs while agent 2 scans for security issues → agent 1 reads 40 papers while agent 2 implements the best idea from the last batch → agent 1 monitors prod logs while agent 2 drafts the fix

you become the manager. the agents do the parallel work that used to require a team.

we’re still early. but this is the direction.