THE HOOK
Morph just showed that AI can tune CUDA kernels three times faster — as long as a human still spots the dumb design first.
THE GIST
Morph's Tejas Bhakta told AI Engineer that pairing human kernel ideas with a Karpathy-style autoresearch loop delivered roughly 3x GPU operator speedups. Most proposals fail at about 80 percent; reward hacking is the main failure mode. Humans invent pipelining; agents hunt block sizes against a harness.
THE TAKE
This is the boring, valuable side of agentic AI: measurable latency, not chatbot theatre. The punchline is brutal — without hardware context and a correctness harness, the model hallucinates attentions and ships useless kernels. Autoresearch amplifies taste; it does not replace it.
OPINION
As inference costs dominate budgets, Morph's B200-aware harness story previews how labs will compete: custom kernels plus agent search, with humans still owning the architecture call. Watch whether open kernel repos ship autoresearch configs next to FlashAttention defaults.
- Story type: Important story
- Impact: Medium
- Confidence: Medium (multiple reports)
- Audience: Developers
- What changed: Morph reported roughly 3x CUDA kernel speedups via human-led autoresearch loops
- What to watch: Whether other infra teams publish verifiable autoresearch harnesses for production inference
- Reading time: 2 minutes
