PRAXIST open-sourced: 49 golds on MLE-bench at 1/12th the cost of Claude Code
SignalCompetitive582 · reddit · 2026-08-28
Sapient Intelligence (the HRM architecture team) released PRAXIST, a lineage-centered generational system for autonomous R&D agents. It converts reproducible artifacts and evaluator outcomes into a typed evidence graph, separating local artifact construction from cohort-level evidence synthesis so later attempts inherit validated mechanisms, unresolved claims and constraints—stopping long campaigns from relearning the same lessons.
Key results: On the 75-task MLE-bench with official graders, PRAXIST earns 60 medals (80.0%, 49 gold) versus 55 medals (73.3%, 34 gold) for a Claude Code baseline on Claude Opus 4.8—at US$3,054 model spend versus US$38,370, roughly a twelfth of the cost. Four open-ended case studies (quant trading, LiDAR-inertial-visual SLAM, tokamak magnetic control, rocket landing) all beat task-native baselines with the discovery path on record.
Licensing & signal: PRAXIST is open source, but companies above US$1M annual revenue need a commercial license. The poster reads it as proof that harnesses dramatically amplify LLMs—and possibly a hint that HRM itself didn't scale well.
More from coding & agent
- Engineer: my job went from silently thinking at a computer to constantly talking to it — KevinNaughtonJr · 2026-10-16
- Vercel Goes Full Backend: Python GA, Containers, and Daemons — cramforce · 2026-10-16
- WorldMemArena Accepted at NeurIPS: Memory Storage ≠ Memory Use — xwang_lk · 2026-10-16
- Vercel Ship Demos Container Deploys for Go, Ruby, Java and More — edwin · 2026-10-16
- TRL adds day-0 support for fine-tuning Muse Spark Open's 816B MoE with LoRA on 64 H100s — QGallouedec · 2026-10-16
- Databricks engineers land impactful PRs on day 2 with AI dev tools, vs week 3 before — Yuchenj_UW · 2026-10-16