Sorry, unable to follow Yu Hai. Please try again.
Trained Cursor-level next-edit autocomplete in 6 months as cofounder / CTO at Sweep (YC S23). Now running running Datacurve's SWE RLVR environment pipeline used by tier-1 labs for training and eval. Blog on training (SFT & RL): https://blog.sweep.dev/posts/oss-next-edit Open-source weights: https://huggingface.co/sweepai/sweep-next-edit-v2-7B Background: Canadian Math Olympiad top 50, $10k scholarship to Waterloo CS (98% MAV, ~60% median), dropped out after two years to cofound Sweep. At Sweep we built one of the first end-to-end coding agents in 2023, when GPT-4 had just launched, shipping PRs at a 25% merge rate. Two years in we pivoted into next-edit autocomplete and a coding agent for JetBrains IDEs (IntelliJ, PyCharm, etc.). Became the #1 plugin on the JetBrains Marketplace at 1.7k DAUs, won bakeoffs against Windsurf and Augment at Ramp and Amplitude. On training the autocomplete model: bootstrapped with synthetic edit trajectories generated from GitHub commits, used RL to lift quality and suppress noisy completions. After deploy, trained on real production traffic to ground the model in users' actual future action trajectories, lifting acceptance rate by 2%. Inference ran at <100ms e2e on a custom stack deployed across three continents, serving 500k autocompletes/day. Model now powers other JetBrains AI plugins, including Firebender.