Working with a thread

A thread is the home for one line of research. It keeps the conversation, context, plans, experiments, reports, and interpretation together so you can leave the work and return without reconstructing it from scratch.

The best threads feel less like chat history and more like a lab notebook with an active collaborator. They preserve not only what you asked, but which evidence changed the question and why the next experiment exists.

Start with the question

Choose New thread and describe what you want to learn or produce. You do not need to specify every implementation detail up front. Begin with the decision the work should inform, then add constraints that would materially change the result.

I want to understand how months of the year are represented in Gemma 2 2B.

That question is enough to begin, but it becomes more useful with a boundary:

Focus on whether the representation is consistent across prompt formats. Use the attached checkpoint, hold out two months from probe training, and produce an explorer we can inspect together.

Include any constraint that would meaningfully change the data, model, evaluation, safety boundary, compute budget, or output. Silico can work through the remaining choices with you.

Choose how Silico works

The model picker configures three roles:

• Planner develops plans and maintains the research direction.
• Worker runs experiments and produces reports.
• Critic reviews plans and evidence.

The defaults are a good place to begin. Change them only when the work calls for a different balance of speed, cost, or capability; model selection should not become a prerequisite for asking a good question.

Enable Auto-accept for routine work that can proceed without review. Keep it off for expensive, irreversible, or high-stakes work. You can change it from the thread composer.

What Auto-accept does

Auto-accept keeps a visible review window rather than launching work immediately:

• A ready plan waits three minutes before Silico launches it.
• An unanswered question waits 30 seconds before Silico returns to the question and proceeds using its best judgment.
• Interacting with a question gives you up to a minute from the latest input.

Choose Pause on a plan or question when you need more time. Resume starts a fresh review window without turning Auto-accept off for the rest of the thread.

If a launch fails or the organization cannot admit more billable work, the plan or question remains pending for you. Auto-accept never marks an unsuccessful launch as completed.

Attach context at the point of use

Open Add files, code, models in the composer. You can:

• upload files from your computer
• upload from a connected cluster
• browse existing resources
• browse files by location
• add a GitHub repository
• add a Hugging Face model or dataset
• add an arXiv paper

An attachment is most useful when its role is explicit. “Use this CSV as the held-out evaluation set” is better than attaching final_v2.csv and relying on its filename. Removing unrelated context helps Silico stay focused on the question you are asking now.

For repositories, name the branch, path, or subsystem when the whole repository is ambiguous. Connect GitHub under Personal settings → Integrations for private repositories.

Private or gated Hugging Face resources require a connected token. Files on compute require access to their location.

Attach an earlier thread or report when a new investigation depends on it. Say whether it should be treated as a baseline, a method to reproduce, or a claim to challenge. Preserve the model, data, checkpoint, and metric context—not only the conclusion.

Steer through the thread

Use the thread conversation as the main place to collaborate with Silico: refine the question, compare interpretations, request controls, and review plans. This keeps research direction with the planner rather than scattering it across individual worker logs.

While Silico is working, a new message queues behind the current action. Stop the current action when the direction is wrong and should not finish.

When Silico asks about a checkpoint, split, metric, scope, or output, answer as precisely as you can. Include the reason when it affects interpretation. Your decision stays in the research record for anyone returning to the work later.

Return to long-running work

Open Summary first when you return after a day or a month. It should tell you what the thread currently believes, what is still running, which reports matter, and what remains uncertain. Use Timeline when you need to reconstruct the path to that state.

For an older thread without a summary, choose Generate from past runs. This consumes RUs.

Use the code icon when the thread has a codebase and you need its files or history.

Keep one research direction together

Keep working in the same thread when the follow-up depends on prior evidence. Fork when you want to branch from a specific response or test a competing interpretation without disturbing the source. Start a new thread when the question can stand on its own.

A thread should have a coherent research direction, not a single eternal topic. “Everything about Gemma” is too broad; “How temporal concepts are represented in Gemma 2 2B” gives later experiments a shared reason to be together.

Folders, pinning, and search help you navigate threads without changing their research identity. See Organize threads.

Next step

When Silico proposes a plan, continue with Review and launch a plan.