Morph Founder: Autoresearch Made Our Models 3x Faster, but 80% of Attempts Fail
AI Engineer · youtube · 2026-09-26
Tejas Bhakta, founder of Morph and formerly of Tesla's inference optimization team, explains how his team combined agent-written custom GPU kernels with bare-metal tweaks to run models 3x faster on cheaper GPUs.
Key points:
- GPU kernels are a near-perfect autoresearch target since correctness and speed are easy to verify.
- Humans bring the big ideas; agents handle parameter tuning. Agents miss high-level insights, like spotting that a DeepSeek attention step loads far more context than needed.
- Agents need explicit hardware and model context to work well.
- Watch for reward hacking: e.g., disabling CUDA graphs to speed up one kernel while slowing the whole model.
- Kernel gains compound, and bare-metal tweaks add 25% over a cloud VM.
- His honest warning: roughly 80% of autoresearch attempts are bad.
More from coding & agent
- Adobe MAX: MCP across Creative Cloud, AI agents editing layers, Firefly Builder and Graph — umesh_ai · 2026-11-11
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11