Goodfire's Latest Neural-Geometry Research
goodfire.ai ● Covered by 4 sources
Goodfire just dumped a huge pile of interpretability research, from vision models to Alzheimer's biomarkers. It's also pitching a tool called Silico to poke inside AI models directly.
Based on reporting by goodfire.ai — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Scroll through Goodfire's latest publishing binge and you get a sense of just how sprawling the interpretability research agenda has become. There are papers on neural geometry in vision models, on sparse autoencoders, on circuit tracing, on evaluation awareness in language models, and on something called model diff amplification for surfacing rare bad behaviors. It reads less like a single research agenda and more like a company trying to prove its toolkit works everywhere at once.
And that spread is deliberate. Some of the work stays close to home, digging into how large language models like Llama 3.3 70B represent concepts internally, or how models retrieve bound entities in context. Other papers wander far from typical AI safety territory entirely, applying the same interpretability techniques to genomic foundation models, to explaining millions of genetic variants, to materials discovery, and to identifying a new class of Alzheimer's biomarkers. There's also a project with Rakuten on using SAE probes for PII detection in production, which is the rare example here of the research actually shipping into a real deployment.
Underneath all of it sits Silico, the product Goodfire is building around this research. The pitch is that it lets people build AI models with something like the precision of writing software: see what a model has actually learned, catch undesired behavior before it becomes a problem, and make targeted interventions rather than blunt retraining. That's the throughline connecting a paper about hallucination reduction to one about detecting evaluation awareness to one about steering models along manifolds. Each is a different angle on the same bet, that if you can see inside a model clearly enough, you can fix it surgically instead of guessing.
What's notable is how much of this output reads as foundational, exploratory work rather than finished products. Titles like Open Problems in Mechanistic Interpretability sit right alongside applied case studies, which suggests a field, and a company, still figuring out which of these techniques will actually hold up outside a research paper. Goodfire is clearly betting the answer is enough of them to build a business on.
My take — AI-written commentary, not fact-checked reporting
A pile of paper titles is not the same as a pile of proof, and interpretability research has a long history of impressive-looking internal demos that never quite make it into products people trust with real decisions. The Rakuten PII work is the one item here that matters more than the rest, because it's the only sign this stuff survives contact with an actual deployment rather than a benchmark. Everything else reads like a company building credibility for a sales pitch, which is fine, but nobody should confuse research volume with research that changes how anyone actually ships models.
Read more about this at: goodfire.ai