У китайцев есть свой Лекун как Джим Фан, в этой статье делается обзор его доклада. Строго говоря, Лекун не изобретатель подхода "модели мира". Однако как и у Лекуна и у Фана проблема доказать на практике, что их подход не академические упражнения, а показать гуманоидов уровня Unitree на RL-обучении. Если брать подход "модели мира", то идея его заключается в том, что мы не обучаем модель решать конкретные задачи, а пытается создать у нее "понимание мира", чтобы она могла решать новые задачи. Одна
Pulling top comments from 5 nearest past posts and asking the model to mirror their style — about 8 seconds.
ForesynClosest the rewriter got vs your draft (3 pts) — none beat it; here's what was tried
Generating 8 candidates across strategies, scoring each, then a 2nd round if fewer than 3 beat your draft — up to ~20 seconds.
ForesynAuto-improve — branch out, keep climbing, stop on plateau
Greedy hill-climb: each iteration generates 8 rewrites, picks the highest-scoring one, then rewrites from THAT. Stops on plateau or after 5 iterations. Free, ~30–60 sec, rate-limited to 3 runs per minute.
Edit & re-score
YC S26Help us get into Y Combinator Summer 2026
This HN tool is a weekend side project. The real Foresyn is the company brain behind it — agents that ship reviewable AI work for real B2B teams. We're applying to YC Summer 2026 right now.
Know a YC partner? Please tell them about us — even a “hey, look at this team” message means a lot.
Why we think we deserve a look: founder is ex-Stripe, worked on AI review tooling for high-context fraud/rules workflows. Live product with reviewable workflows, agent workers, and a memory loop that compounds corrections. This HN microsite is a 3-day demo of the same engine — pixel-faithful HN clone with a virality predictor that beats the 2014 baseline. We ship fast and we ship real.