Goodfire, a startup that builds interpretability tools for reading what is happening inside AI models, opened a private beta for Silico, a system it describes as a standing team of AI researchers available on demand. The company posted a demo thread on X on July 15 showing Silico running a series of research tasks end to end, from replicating published techniques to probing biology models, without a human executing the steps. If the results hold up outside a product-launch thread, Silico is a concrete example of an agent handling the mechanics of machine learning research itself, not just assisting a human who runs it.
According to Goodfire’s thread, Silico replicated a technique called J-space on Zhipu AI’s GLM-5.2, the Chinese lab’s large language model, overnight. It then extended the model’s context window to roughly 256,000 tokens and reproduced the associated multi-hop question-answering results on its own. In a separate run, Silico reproduced RLFR, a method Goodfire says its own team spent months building, in two days. RLFR places probes inside a model and uses their readings as reward signals for reinforcement learning; Goodfire says the automated reproduction cut hallucinations in Alibaba’s open-weight Qwen3-8B by 37 percent without a corresponding drop in capability.
The thread describes two further examples. Applying a sparse-feature probing method it calls BSFs to protein language models, Silico surfaced internal subspaces whose activations track known protein structures, without any supervision pointing it there. Separately, it replicated PICASSO, an interpretability method for digital pathology models, on a system called Midnight-12k in a single attempt, breaking the model’s view of tissue samples into readable concepts, ranking which ones drove its cancer predictions, and simulating how altering the tissue would shift those predictions.
Every result described so far is Goodfire’s own account of its own product replicating its own prior research. The company has not published an independent benchmark, a third-party reproduction, or a timed comparison against a human team working the same problems outside its own past effort. The “months versus two days” contrast for RLFR measures Silico against Goodfire’s internal history, not against a rival lab or an external clock, which is a different and weaker claim than it first appears.
Silico enters a small but growing set of systems built to automate the research process itself rather than the coding that supports it, alongside earlier efforts like Sakana AI’s automated-scientist projects. A tool that edits code under a human’s direction is a productivity aid. A tool that designs an experiment, runs it, and reports back what it found is closer to a research collaborator, and the distinction matters for how much oversight a lab can safely remove from its pipeline.
Access sits behind a private beta requested through Goodfire’s contact page, so outside researchers cannot yet test Silico’s claims against their own workloads. Interpretability and safety teams evaluating automated research tools should treat this thread as a pitch rather than a verified result until Goodfire opens access widely enough, or publishes methodology detailed enough, for a third party to rerun the same experiments and check the numbers independently.
Based on a thread posted by Goodfire (@GoodfireAI) on X on July 15, 2026.
