DuoDiscover AI agents top NVIDIA SOL-ExecBench with 2.68x TPU kernel speedup
CShorten30 · x · 2026-10-15
DuoDiscover, a collaboration between Google Cloud, Berkeley Sky, and Stanford AI Lab, deploys AI agents that ship real merged PRs and beat human-expert hardware baselines. Highlights: first superhuman CPU prefetcher (+3.93% / +6.31% overall, outperforming AlphaEvolve), 2.68x TPU kernel speedup, 1.92x faster vLLM serving, 1.44x on FreeToken, #1 on NVIDIA SOL-ExecBench (B200), and 2x experts on SWE-fficiency. The key idea: evaluation cost—not idea generation—is the bottleneck in engineering, so agents co-discover fast evaluation proxies ("yardsticks") to slash slow eval loops and compound real gains.
Related event: DuoDiscover Tops NVIDIA SOL-ExecBench, Speeds Up TPU Kernels 2.68x(5 posts)→
More from Infra
- Epic's Lore vs Anthropic's Scion: two post-Git, centralized VCS architectures compared — ssh4net · 2026-10-15
- Listed quantum computing companies top $50B market cap on just $132M in quarterly revenue — hardimanjames · 2026-10-15
- Liquid Inference exchange hits 28 providers and 1000+ models with 3-hour onboarding — JosephJacks_ · 2026-10-15
- Quote of OCP Summit: 'Optical has been 5 years away for the past 15 years' — BenBajarin · 2026-10-15
- Jailbroken PS5 runs Qwen 3.8 27B on its GPU at 20 tok/s with PS5LM Plus — 98Saman · 2026-10-15
- llama.cpp tuning alone makes Qwen3-4B 1.34x faster on a 2019 phone, offline — hbouammar · 2026-10-15