Find a minimal BF16 multiply-add circuit — real problems beat asymptotic exponent shaving
MikePFrank · x · 2026-10-09
MikePFrank argues AI efforts should target concrete constant-size problems: finding a Boolean circuit for BF16 multiply-add with the fewest two-input gates or fewest parallel levels. He dismisses reducing an exponent from 1 to 0.99999 in an astronomically impractical algorithm as 'mathematical slop' that only demonstrates poor judgment about what's worth spending time and tokens on.
Related event: Researcher calls AI-driven exponent optimization "math slop"(2 posts)→
More from Research
- Data firms need frontier-level training infra to prove data value, says Proximal — AccBalanced · 2026-10-10
- AI agent misses $14M in lease clauses: Surge AI's GDP.xlsx benchmarks agents on real spreadsheets — echen · 2026-10-10
- Calico open-sources Cerberus DNA model in PyTorch: 786kb input, 8361 human tracks — anshulkundaje · 2026-10-10
- New preprint uses state space models for long-sequence genomic prediction, open-sources Cerberus PyTorch models — anshulkundaje · 2026-10-10
- MaCVi maritime computer vision workshop lands at WACV 2027 with 3D sonar and underwater benchmarks — HildeKuehne · 2026-10-10
- OpenAI pushes 722 math manuscripts to GitHub from a model nobody can use — thursdai_pod · 2026-10-10