Silico - Devoured - Daily Dev News

Back to Devoured - July 16, 2026

DEVOURED

Silico

Silico

AI research Thread Reader

Goodfire's platform, Silico, enables engineers to perform automated interpretability research and model debugging through an autonomous agent.

What: Silico uses autonomous agents to run concurrent experiments on model internals. It has successfully replicated research papers like J-space and PICASSO, identified Alzheimer’s biomarkers, and reduced hallucinations in Qwen3-8B by 37% using internal activation probing.

Why it matters: This represents a shift toward treating AI models as deterministic software systems that can be inspected and debugged at the weight and activation level, rather than as opaque black boxes.

Takeaway: Request private beta access at goodfire.ai/contact to use their autonomous model-neuroscientist for your own model experiments.

Deep dive

Automated Experimentation: Uses a model-neuroscientist agent to plan and execute research threads.
Internal Probing: Uses probes on neural network activations as reward signals for reinforcement learning.
Capability Removal: Demonstrated model surgery by fine-tuning on as few as 4 tokens to remove specific capabilities like language knowledge.
Predictive Debugging: Interprets training datasets by passing them through models to predict emergent behaviors before full training begins.
Geometric Analysis: Uses neural geometry to map internal model states across modalities like protein language models and pathology models.

Decoder

Neural Geometry: The study of how LLMs represent concepts through curved paths and manifolds within their high-dimensional activation spaces.
Activation: The numerical output values of neurons in a network layer after processing input, which reflect what the model 'sees' or 'thinks'.
Probing: A technique where a secondary, simpler model is trained to extract specific information from the internal states of a large model to determine what it has learned.

Original article

replicate J-space on GLM 5.2

train a reward model and run RL to reduce hallucinations

show me how this model makes cancer predictions

Using our platform Silico is like having a team of AI researchers ready to run experiments like these.

Private beta is open now. (1/6)

Silico replicated J-space on GLM-5.2 overnight.

It then extended context to ~256k tokens, replicating the key results on multi-hop question answering. (2/6)

https://twitter.com/1353836358901501952/status/2074185348142280912

Our team spent months developing RLFR, our method which uses probes on a model's internals as reward signals for RL.

Silico reproduced it in 2 days, reducing hallucinations in Qwen3-8B by 37% without capability loss. (3/6)

Silico lets us look inside models to see what they’ve learned.

Using BSFs on protein language models, it found - without supervision - subspaces in the model whose activations correlate with known protein structures. (4/6)

Here, Silico replicated PICASSO on Midnight-12k in one-shot.

PICASSO interprets digital pathology models. It breaks what the model sees into readable concepts, shows which drive its cancer predictions, and simulates how changes to the tissue would alter those predictions. (5/6)

https://twitter.com/1429874793914851328/status/2071319010185224496

These are just a few examples of what you can do with Silico.

Request access to the private beta here: goodfire.ai/contact

Explore all 56 articles in today's digest