Fork a research notebook

Each notebook is a reusable research starting point: a question, method, supporting experiments, controls, and an interactive result.

Inspect the result, then continue it in Silico to change the model, dataset, intervention, or evaluation for your own question.

Intentional model design

Teach a language model to hallucinate less, then verify the change with an independent judge.

Language models · Advanced Train hallucinations out with a probe Teach a language model to hallucinate less, then verify on held-out prompts that it became more factual without losing general ability. 260–340 RU · estimate · 2 hr 10 min · 4× H100 Activation probesReinforcement learning →

Scientific discovery

Test whether a genomic model has learned the evolutionary relationships between species.

Life sciences · Introductory Finding the Tree of Life in Evo 2 Test whether a genomic model has learned evolutionary relationships across 2,395 species, then challenge the result with composition and shuffled-label controls. 90–140 RU · estimate · 55 min · 1× H100 Neural geometryEvaluation and benchmarking →

Autonomous training

Give Silico a training goal, then inspect the implementation, GPU run, recovery, and final evaluation.

Post-training · Advanced Train an agent with reinforcement learning in Silico Give Silico a training goal and let it implement, launch, recover, and verify a complete reinforcement-learning run. 640–780 RU · estimate · 3 hr 20 min · 4× H100 · Prime Reinforcement learningEvaluation and benchmarking →

Explore more notebooks

Put a guardrail around an LLM agent

Train a detector for unsafe tool calls, choose an operating threshold, and measure what the guardrail costs on legitimate security work.

AI security · Intermediate Build a tool-call safety guardrail from a model's activations Train an activation probe that catches unsafe tool calls before execution, then calibrate it against benign security work. 55–75 RU · measured · 42 min · 1× H100 Activation probesSafety evaluation →

Find and test an internal signal

Read a structured signal from model activations, then intervene on it to distinguish a useful mechanism from a merely decodable pattern.

Language models · Introductory Getting Started with Neural Geometry Find a structured concept in a model's activations, test whether the geometry is real, and intervene on it. 45–65 RU · estimate · 35 min · 1× H100 Representation analysisNeural geometry →

Life sciences · Intermediate Probe the copy-number axis in Evo 2 7B Decode tandem-repeat copy number from a genomic model's activations, then test whether the same axis can steer generation. 90–130 RU · estimate · 70 min · 1× H100 Linear probesSteering →

Test a multimodal intervention

Change where a vision-language model attends, then test the effect against a conflicting prompt and a matched random control.

Multimodal · Intermediate Steer a VLM's gaze over images Redirect a small set of attention heads and test whether visual steering can override a conflicting text prompt. 60–85 RU · estimate · 50 min · 1× H100 Attention analysisSteering →