Introducing Silico — Goodfire's platform for building AI models with the precision of written software

TL;DR

Goodfire’s 10-tweet launch thread for Silico, their commercial mechanistic-interpretability platform. The product packages frontier interp techniques (feature decomposition, residual-stream probes, parameter decomposition, sparse autoencoders) behind an autonomous “model neuroscientist” agent that plans and runs concurrent experiments on a customer’s model. Pitch: see inside the model, check health (undertraining, info bottlenecks, feature collapse), debug failures, shape behaviour via internal-feature steering, generalize from less data. The thread cites prior Goodfire results — a novel class of Alzheimer’s biomarkers from Prima Mente’s Pleiades epigenetic model, SOTA disease-variant prediction on Arc Institute’s Evo 2 via covariance probes, SAE-probe PII detection matching GPT-5 Mini at 1/15 the cost on Rakuten production data, and Stochastic Parameter Decomposition (SPD). Early access is open.

Key claims

Silico exposes feature decomposition over a customer model’s residual stream, letting researchers tell apart real understanding from spurious correlation [thread 4/10].
“Health diagnostics” surface undertraining, information bottlenecks, and feature collapse before they hit downstream metrics — framed as the missing primitive between W&B-style experiment tracking and actual model understanding [thread 5/10].
Internal-feature interventions can steer generation and extract stronger predictors than the model’s own readout — explicitly positioned as a capability-improving tool, not only a safety/debug tool [thread 7/10].
The platform’s automation layer is an “autonomous model-neuroscientist agent” that plans and runs concurrent interpretability experiments and learns over time [thread 3/10].
Cited prior result — Alzheimer’s biomarkers: Goodfire interpreted Prima Mente’s Pleiades epigenetic model (SOTA on early Alzheimer’s detection) and recovered a novel biomarker class — pitched as the first natural-sciences finding from reverse-engineering a foundation model [thread, Prima Mente sub-thread].
Cited prior result — disease-variant prediction: covariance probes on Arc Institute’s Evo 2 reach SOTA on whether/how 4.2M variants in NIH ClinVar cause disease, with strong cross-variant-type generalization; an open-source database is being released [thread, Mayo Clinic sub-thread].
Cited prior result — PII detection: a random-forest SAE probe trained on synthetic data transfers to Rakuten’s production PII and matches GPT-5 Mini quality at ~1/15 the cost, deployed to Rakuten’s agent platform [thread, Rakuten sub-thread].
Cited prior result — Stochastic Parameter Decomposition (SPD): a parameter-space (not activation-space) decomposition method that finds minimal subcomponents which faithfully recreate the model’s behaviour, framed as “reverse-engineering the source code instead of inspecting runtime behaviour” [thread, SPD sub-thread].
Cited prior result — loss-curvature memorization probe: a PCA-on-curvature decomposition of weight matrices isolates the components that most affect loss on the training set, used to study where memorization lives in LLMs [thread, Merullo sub-thread].

Method

Not a paper — a product announcement. The substantive content is in the cited prior work threaded below the launch. Three components are described concretely: (1) the model neuroscientist agent that operates inside a hosted model design environment workspace, organizing research threads, replicating papers, and running experiments concurrently; (2) a primitives layer for residual-stream feature decomposition (SAE-style), parameter decomposition (SPD), and diagnostic probes for representation pathologies; (3) intervention APIs — boost/suppress neurons, filter training data to avoid setting unwanted parameter values, steer along learned features. The pricing is bespoke; access is gated by request.

Results

The thread references concrete capability numbers from prior Goodfire work rather than from Silico itself:

Pleiades + Goodfire interp → first natural-sciences finding (novel Alzheimer’s biomarker class) attributed to reverse-engineering a foundation model.
Evo 2 + covariance probes → SOTA on disease-variant prediction over 4.2M ClinVar variants, generalizing across variant types.
Rakuten PII detection → SAE probes match GPT-5 Mini accuracy at ~1/15 cost on real production data, transferred from synthetic training; deployed.
Token-saving via “performative chain-of-thought” probes → up to 68% of reasoning tokens skippable with minimal accuracy loss (cited on Goodfire’s home page, referenced in the thread context).
EchoJEPA latent-space analysis → interpretable clinical features (motion, anatomy) recovered from a cardiac-echocardiography vision JEPA.

Why it’s interesting

This is the first commercial productization of the mechanistic-interpretability stack into a self-serve platform — adjacent in spirit to Emotion Concepts and their Function in a Large Language Model (Anthropic’s emotion-vector work, where linear residual-stream directions causally control reward hacking, blackmail, and sycophancy) but pitched to non-frontier labs that lack their own interp teams. Where Anthropic ships interpretability as research, Goodfire is shipping it as a build-time tool, and the SAE-probe-as-cheap-classifier result (matching GPT-5 Mini PII detection at 1/15 the cost) is the kind of cost-curve argument that could plausibly land outside safety teams. The “decompose parameters, not activations” framing of SPD is the most technically distinctive piece — it complements activation-side work like Do Audio-Visual Large Language Models Really See and Hear? (logit-lens / attention-pattern interpretability stack on AV-LLMs) by going one level deeper into where computations live. The autonomous “model neuroscientist” agent is the bet that LLM agents are now strong enough to do interp work humans were doing two years ago — adjacent to OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis‘s long-horizon research-trajectory synthesis pattern, applied to interpretability rather than literature search.