2021 slides reveal Garçon, Anthropic's internal interpretability infra for multi-GPU models
voooooogel · x · 2026-10-15
User voooooogel shared 2021 slides showing Garçon, Anthropic's internal interpretability infrastructure, lamenting that OSS still has no equivalent.
Key points:
- Problem: investigating large models requires multiple GPUs (16+); loading models directly in Jupyter notebooks didn't scale.
- Garçon is a server for language models: launched with a snapshot path (e.g. python -m garcon.rpc.server launch --ngpu 16 s3://...), exposing an RPC interface callable from scripts or notebooks.
- Supports forward/backward passes, sampling, and a "probes" abstraction to access internal activations.
- Use cases: interactive interpretability from notebooks, collecting dataset examples and statistics, human feedback exploration, investigating loss spikes, visualizing value-headed RL models.
More from Research
- AlphaNIST: one RL policy solves all 20 NIST assembly tasks with 98.2% success in simulation — hbouammar · 2026-10-15
- Linear Depth Memory: matrix-valued memory tops PreLN and mHC on DepthBench across scales — SonglinYang4 · 2026-10-15
- Neural field planning for hybrid manufacturing cuts estimated build time 68.3% — jiqizhixin · 2026-10-15
- Tev-Gym: 1,170 dense-reward RL environments push long-horizon agentic training — RulinShao · 2026-10-15
- Claude systematically debunks the speculative EDAR-selenium hypothesis about East Asians — teortaxesTex · 2026-10-15
- Microsoft's AutoEnvScaling lets agents build their own RL environments, lifting Qwen3 Terminal-Bench from 41.1 to 53.3 — omarsar0 · 2026-10-15