TorchLens 3.0 ships: one-call inspection and ablation of any DNN's internals
burny_tech · x · 2026-11-07
JohnMark Taylor released TorchLens 3.0, aiming to eliminate the "plumbing" of inspecting and intervening on DNN internals for mech interp and NeuroAI research, on any model:
- tl.trace(model, input) records every op's tensors and params in one forward pass.
- Facets expose internals like attention patterns (e.g., ViT layer 5, shape 1×12×197×197).
- Interventions are trivial: zero-ablating one attention head flips a "tabby cat" classification to "Egyptian cat" (prob 0.489), with the modified graph visualized.
- Includes outputtable(k=3) and draw visualizations, usable by agents too.
More from coding & agent
- Problem #4 of running 100 AI agents: cascading failures and runaway retries — goyalshaliniuk · 2026-11-07
- Problem #5 of running 100 AI agents: observability, security and per-agent cost tracking — goyalshaliniuk · 2026-11-07
- Problem #3 of running 100 AI agents: tool contention, rate limits and cost blowups — goyalshaliniuk · 2026-11-07
- Problem #2 of running 100 AI agents: shared context drift, loss and token bloat — goyalshaliniuk · 2026-11-07
- Problem #1 of running 100 AI agents: coordination, conflicts and dependency chaos — goyalshaliniuk · 2026-11-07
- 5 engineering problems that emerge when you deploy 100 AI agents in production — goyalshaliniuk · 2026-11-07