Team open-sources Tahuna infra to power Hillclimb, an autonomous ML experimentation loop
Monaim101 · reddit · 2026-09-14
The team open-sourced Tahuna, their ML workload infrastructure, and started Hillclimb, an autonomous experimentation loop where an agent proposes and iteratively runs improvements.
Execution layer design:
- Workflow: init → sync → computeSession → train / serve / hillclimb
- Content-addressed code/data sync and pinned manifests give each experiment a concrete snapshot
- GPU sessions stay warm between runs; one active workload per machine
- Warden, a Go execution agent, handles workspace materialization, uv dependency installs, the Python workload, logs, and artifact sync
Currently supports RunPod GPUs and R2 storage, with Docker self-hosting for the control plane; examples include SFT, RL agentic search, and MNIST. AGPL-3.0 licensed. The authors stress it's an early public preview and ask for feedback on inspectability and control of the execution layer.
More from coding & agent
- NVIDIA open-sources FlashREINFORCE: critic-free single-rollout async RL with 6,000+ stable updates — inductionheads · 2026-09-14
- TradingAgents: a free open-source AI hedge fund that runs on your laptop — mdancho84 · 2026-09-14
- User runs his life on an Instinct agent: deliveries, trips, and support emails via WhatsApp — sagartheamruth · 2026-09-14
- Think locally, tool-call globally: a pithy take on agent architecture — charles_irl · 2026-09-14
- Codex Remote blurred bar on iOS to be fixed with iOS 27 SDK — Dimillian · 2026-09-14
- Tip: Claude skills can also bundle runnable scripts — JeremyNguyenPhD · 2026-09-14