Hawkeye brings hardware-aware kernel optimization to coding agents, at NeurIPS
simonguozirui · x · 2026-11-16
Hawkeye, to be presented at NeurIPS, targets a key bottleneck: new ML chips have unique architectural features but software support lags, and peak performance increasingly depends on hardware-specific kernel optimization that coding agents handle poorly. The framework grounds agents in a minimal yet comprehensive taxonomy of optimization strategies — for a new GPU or ML accelerator you write just 10 unit tests and solution kernels (one per strategy), letting agents scale test-time compute effectively with this minimal supervision.
More from coding & agent
- Same analytics agent task: Sonnet 5.5 costs ~12x more than Haiku 5.5 — sh_reya · 2026-11-16
- Jiff 1.0 released: Rust datetime library already powering uv, Cargo and more — charliermarsh · 2026-11-16
- Berth v0.6.0 leases one RTX 4090 across 23 engines, 143 lease-hours a week — Distinct-Pie2389 · 2026-11-16
- Agent notices Mac battery at 3% and pauses work to warn user — adonis_singh · 2026-11-16
- Three classes of programming languages in the AI era: natural language, performant code, and specs — dbreunig · 2026-11-16
- Turning newsletter edits into AI checks lifts frontier models 25+ points on summaries — danshipper · 2026-11-16