Translucent: open-source AI plans hardware experiments, designs battery charge in $1.99 of model spend
YashasGunderia · x · 2026-11-05
Builder Yashas Gunderia open-sourced Translucent, a physical-test infrastructure for hardware teams built for Meta's Global AI Developer Hackathon, with replayable recorded investigations in the browser.
- Workflow: records + failing baseline + objective/constraints; a planner proposes experiments, a human approves, a solver runs them, a comparator judges results
- Headline demo (LG M50 cells): Muse Wonder wrote its own importer, fitted the lab fixture's thermal model (0.15 K RMSE), and designed a 35.875-min charging scheme matching a reference controller given the hidden lab condition — 13 turns, 12 experiments, 15.7 min, $1.99 model spend
- Other cases include fast-charge optimization beating a 28.2-min 1.5C baseline at 37.2 min under raised plating limits
He's seeking feedback from battery, robotics, electronics, materials and hardware-testing practitioners.
More from Research
- PhD researcher launches 'paper nostalgia' series with Real NVP coupling-layer explainer — LucaAmb · 2026-11-05
- LeCun vs. researcher: LLM weights store no discrete facts — JEPA bets on 'understanding' instead — solyarisoftware · 2026-11-05
- New theory says curricula help with unlearning simple rules, not learning — summerfieldlab · 2026-11-05
- uni-robot-agent gets Qwen working in sim and on real robots, launching OpenRUA — teortaxesTex · 2026-11-05
- Fabricated references in 500K-paper sample surge from 6.3 to 23.0 per 10,000 since 2023 — skdh · 2026-11-05
- Nature paper warns recursive training on AI-generated data causes irreversible model collapse — solyarisoftware · 2026-11-05