Answer.AI Opens SolveIt to Everyone: Code, Notes and AI in One Dialog After a 3,000-User Beta
In a typical AI chat, context keeps growing and useful turns get buried under irrelevant ones. After a two-year closed beta with three thousand participants, Answer.AI has opened SolveIt to the public. In it, code, Markdown notes and model replies sit in one editable document. You can insert a prompt anywhere, and the model sees only what is above it. That makes it easier for a developer to keep the context clean and understand why the code looks the way it does.
AI-processed from Habr AI; edited by Hamidun News
Anyone who has tried writing code with an AI assistant knows the feeling: two hours in, you have a working program, but you don't quite understand why it is built the way it is. Answer.AI, the research organization behind fast.ai, proposes a different approach. It has opened SolveIt to everyone, an interactive environment where a human and a language model work in the same document. Before this, the platform spent two years in closed beta, with about three thousand users from the community on Discord.
Sign-up is free, and you pay for the compute resources you actually use. You can try it at solve.it.com.
Not a chat feed, but a living document
In an ordinary chatbot, messages are simply appended to the end of the conversation. SolveIt works differently: executable code, Markdown notes, LaTeX formulas and model replies all share one space. Think of a Jupyter notebook in which the AI is a full participant.
The key practical detail is context management. You can insert a prompt to the model at any point in the dialog, and the model takes into account only what is above it. Blocks can be moved, edited and deleted, including the AI's own replies. If the model goes off track, you don't have to live with the bad reply for the rest of the session: you just delete it, and it no longer clutters the context. The model also runs code in the same interpreter as the human, so both sides see the same state.
The team offers an example of its own: researcher Jack Hogan was building a color palette generator for an interface. He first tried it with Claude Code. After experiments that consumed millions of tokens, the assistant produced a working but overly complex tool, and Hogan could not explain why that approach had been chosen.
In SolveIt he moved in small steps: he recorded requirements in notes, calculated color gamut boundaries directly in code, wrote the relative luminance formula in LaTeX, tried different approaches, including a Fourier transform, and finally generated the CSS. In his words, being able to justify every decision greatly increased his confidence in the result. This is, of course, testimony from the authors themselves, not an independent test.
Server, hosting and extensions
Each session runs on a persistent server with its own file system and terminal, and git is already installed. You can create as many such instances as you like, for separate tasks or for collaboration. Several people can edit the same dialog at the same time, and the finished result can be exported to Markdown with templates and cross-references.
The platform also doubles as a host: any program running in an instance gets a public address or your own domain, with no separate deployment step. If you put a dialog in the AUTORUN directory, its scripts start together with the server and run continuously. The authors say this is how their own services run, including the SolveIt backend and a chatbot for students.
You can extend the system from within. Any Python function becomes a tool the AI can call. The model can read and edit the current document and trigger generation in neighboring dialogs. A Chrome extension lets you give it control over actions in web interfaces. A CRAFT file placed in a folder sets shared instructions for the AI and starter code for all nested documents, as well as its own CSS styles and JavaScript scripts.
Who it's for and what to keep in mind
The Answer.AI team uses SolveIt for data analysis, reading scientific papers, working through mathematical models, reviewing contracts and programming. The law firm Virgil prepares contracts in it. While going through a paper on recursive language models, Jeremy Howard wrote his own implementation of the algorithm straight from the text of the publication.
Ideologically, the project draws on the ideas of Bret Victor, Donald Knuth, Ivan Sutherland and Kenneth Iverson about combining code, data and documentation. The authors deliberately rejected the idea of full automation: the goal is for the human to understand the problem more deeply, not simply to receive a finished result.
The takeaway for practitioners: if you are used to handing an assistant large tasks and accepting the output, SolveIt demands a different rhythm of small steps, manual context editing and notes along the way. That takes longer than a single prompt, but it leaves a trail of reasoning that you can show a colleague or come back to a month later. How it will behave on large team codebases is not yet known: the examples in the announcement are research and analytical tasks, and the user assessments come from the developers themselves. To get started, the authors recommend the open introductory lesson of the course "How to Solve It with Code".
- Nov 18, 2026rust-anything teaches AI agents to rewrite apps in Rust module by module
- Nov 18, 2026One Chinese Comma Switched DeepSeek to Chinese in 94% of Responses
- Nov 18, 2026Hugging Face routes LLM requests through OpenRouter, cutting capacity errors by 46%
- Nov 18, 2026Willison Releases Three External Sandboxes So Agent-Run Code Stays Off Your Laptop
Key Questions Answered
What are the key details on Not a chat feed, but a living document?
In an ordinary chatbot, messages are simply appended to the end of the conversation. SolveIt works differently: executable code, Markdown notes, LaTeX formulas and model replies all share one space. Think of a Jupyter notebook in which the AI is a full participant. The key practical detail is context management. You can insert a prompt to the model at any point in the dialog, and the model takes into account only what is above it. Blocks can be moved, edited and deleted, including the AI's own replies.
What are the key details on Server, hosting and extensions?
Each session runs on a persistent server with its own file system and terminal, and git is already installed. You can create as many such instances as you like, for separate tasks or for collaboration. Several people can edit the same dialog at the same time, and the finished result can be exported to Markdown with templates and cross-references. The platform also doubles as a host: any program running in an instance gets a public address or your own domain, with no separate deployment step.
Who it's for and what to keep in mind?
The Answer.AI team uses SolveIt for data analysis, reading scientific papers, working through mathematical models, reviewing contracts and programming. The law firm Virgil prepares contracts in it. While going through a paper on recursive language models, Jeremy Howard wrote his own implementation of the algorithm straight from the text of the publication. Ideologically, the project draws on the ideas of Bret Victor, Donald Knuth, Ivan Sutherland and Kenneth Iverson about combining code, data and documentation.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.