Announcing the Public Launch of SolveIt
At Answer.AI, we built a tool that lets us learn, explore, and develop mastery with the help of AI. We call it SolveIt. As with everything we build, our goal was not to automate humans out of the process, but to help them grow. Over the past few years, it has helped our team flourish to such a great extent that Eric Ries (author of The Lean Startup and Incorruptible), our co-founder, said that we have a “moral imperative” to share it more widely. Today, we are doing that. It costs nothing to join and you pay only for what you use: solve.it.com
In SolveIt, humans, AI, code, notes, documents, and outputs live and work together. We use it in nearly all our work, including software development, research experiments, teaching, and learning. Over the past two years, we invited 3,000 preview users to test it out with us, and we share a bustling Discord server with them. We’ve all been spending a lot of time learning together, and now it’s time to invite others to become part of this community.
Many people are feeling overwhelmed or defeated by how AI use is often forced on us to maximize outputs at the expense of human understanding. One of the best ways to counter this pressure is to join a community of people who are moving in the opposite direction.
You Can’t Outsource Understanding
We have a key theory at Answer.AI: one thing you can never outsource is your own human understanding. We are not interested in creating a bunch of task artifacts (reports, analyses, computer programs) just for the sake of doing so. Our goal is to create deeper understanding and better tools and abstractions for ourselves as we do our work. Otherwise, we feel like we’d be wasting our time.
Great pioneers like Bret Victor, Donald Knuth, Ivan Sutherland, and Kenneth Iverson showed that people think best when they can work with their code, documentation and data in the same place, directly manipulating them and interactively seeing the results. We realized that AI needs to be in that place too.
Systems like Claude and ChatGPT have just one way of working: you type into a text box, and your message gets added to the bottom of the conversation. You and the AI don’t share a workspace in any meaningful sense. In SolveIt, you do. Your documents, your code and the AI live together on a dedicated server, and the AI runs code in the same interpreter you use.
Case Study: Understanding Color Theory
As an example, let’s look at a recent SolveIt dialog created by our colleague Jack Hogan, a researcher at Answer.AI who did his PhD on multi-modal machine learning. As part of a new GUI framework we’re building, he wanted to create a system to automatically generate a full color palette from any chosen hue.
He shared a fascinating dialog with the rest of the team showing his explorations. You can check it out for yourself here.
Jack starts with some simple demos, illustrating the benefit of having a constant brightness:

Jack walks through his design rationale, necessary utilities, creating gamut boundaries, ladder design, ladder optimization, and CSS generation. While he had a rough idea of what he wanted to cover in advance, it was only through his explorations that he fully mapped it out, leading to some interesting detours along the way (including Fourier transforms).
The work is built up iteratively, in small, well-organized steps. Markdown notes explain the properties he wanted to hold for the palette and for the generator itself. Code implements the needed calculations. LaTeX renders the formulas, such as for relative luminance and perceptual color-difference metric.
Using SolveIt, I imported a copy of Jack’s dialog with just two clicks to try it for myself. When I saw a graph or code snippet I didn’t understand, I inserted a prompt and asked the LLM of my choosing a question directly beneath the part that confused me. Below, I wanted more explanation of what his graph meant.

Later in the dialog, Jack outputs cards showing the color ladders he had designed and optimized:

How is this different from other AI coding assistants? Jack had in fact first attempted this exploration in Claude Code. During a lengthy back-and-forth, he discussed a wide range of color theory resources, looked at existing frameworks, and got Claude to run experiments, eventually reaching a set of generated palettes. He told us that the discussion felt collaborative at first, but when he looked back at the exported project, he couldn’t stand behind it. The palette generator appeared to work but it was complex, and because it was the culmination of many millions of tokens worth of experiments that zig-zagged down different avenues, he couldn’t explain the rationale for its approach or why simpler approaches weren’t enough. Jack tried a second attempt: in a fresh Claude Code session, he set about fixing the unreadable prose and clarifying unexplained assumptions. However, while attempting to establish a logical progression to the ideas, he quickly realized he was only post-rationalizing the spurious artifact that Claude had conjured up.
On his third attempt, Jack started from scratch in SolveIt. He set up the dialog with only a statement of objectives and requirements at the top, and from there it was much easier and more satisfying to build in small steps that each followed from the last. Every choice comes with its reason, including why the simpler options fell short. The work still took turns he hadn’t planned, including the need for Fourier series, but he could explain every one of them, and this time he could stand behind the result.
SolveIt is a Power Tool
Jack’s dialog above is like a lab notebook of his research, and the rest of us could read it and see his thought process. When we want to deeply understand the work we are doing (which is more often than not), we turn to SolveIt. We use it for reading papers, studying mathematics, reviewing contracts, analyzing data, writing books, and developing software. We’ve been able to start writing SolveIt in SolveIt, which has allowed us to go much faster.
The reason we can do so much in SolveIt is because it is a power tool, giving us lots of features to customize, control, and drive the direction of our work. Work in SolveIt takes place in a dialog: a place where code and notes, and heading hierarchies, and exploration and AI discussions are all happening and interspersed. As seen in the color gamut theory example, you can edit the dialog. You can move things around. You can delete things. You can change things, and then you can add a prompt, a question to the AI anywhere inside the dialog. It doesn’t have to be at the bottom. The AI only sees everything up to that point, so you’re really in control of what the AI sees.
You can directly edit the AI’s responses. It can be quite infuriating in other products to get into an argument with an AI where you’re going back and forth and then you’ve really polluted your context with this whole argument about how you don’t want it to answer prompts.

Each SolveIt instance is a persistent server, with a terminal, a file system, git, and anything else you choose to install. You can create as many instances as you like: one for a personal project, another for collaboration, a third for experiments. You can clone a repo, explore it with the AI, and push your changes, all in the same place where you do your thinking.
SolveIt is also a web app that hosts web apps. Any app you run on your instance can be served at its public URL, or on your own domain. You can write an app in a dialog, change a route, and see the change live, with no deploy step. Put a dialog in your AUTORUN folder and it starts with your instance and keeps running. That’s how we run production services. The AI Discord bot that our course students use lives in a SolveIt dialog, and when we edit the dialog, the bot updates. We even built, and now run, the servers that host SolveIt from inside SolveIt.
You can extend SolveIt from inside itself. Any Python function you write becomes a tool that the AI can call. The AI can read and edit the dialog it’s working in, run prompts in other dialogs, and, with our Chrome extension, operate your browser. A CRAFT file in a folder gives every dialog in that folder shared AI context, setup code, and even its own CSS and JavaScript, so the interface itself can change to suit the project.
You can share an instance with colleagues and edit the same dialog together in real time. And when the work is done, a dialog can become a document. Virgil, our sibling law firm, uses SolveIt’s markdown system (which is almost certainly the most powerful in the world) to draft client contracts, complete with templating and live cross-references.

None of this gets in your way when you’re starting out. A new user can open a dialog and start asking questions. But the ceiling is far higher than for any dumbed-down interface, and the more you invest in learning SolveIt, the more it can do for you.
Using powerful software is not something that should be restricted to geniuses. It is respectful of people to recognize that they are capable of using sophisticated tools. Jeremy had many critics early on with FastMail (an email service that Jeremy founded in 1999 and which is still operational today) saying that normal people won’t be able to use something so powerful with IMAP and search. He disagreed, saying that FastMail was designed for people who want to invest time because email is important and it’s valuable to learn how to use it properly. And indeed today, every email client people use, e.g. Gmail, is sophisticated and complex, and that’s the norm.
We took the same view at fast.ai. A core part of our work was that we respected that people would be capable and willing to work hard and learn nitty gritty details of deep learning if they put the time and effort in. Many doubted us, claiming only Stanford PhDs could train neural networks. We had our students create an AWS server from scratch and administer it themselves, because this was necessary to use deep learning in 2016. Many of them had never used Unix before, but they were motivated to learn. Fast.ai had a greater impact than the drag-and-drop tools of the time that claimed to “democratize AI”.
The Human-Centered Mission Behind fast.ai & Answer.AI
While SolveIt was built in the last two years, it is the continuation of the values and goals that fueled our work on fast.ai for the last decade. In 2016, we founded fast.ai with the mission to democratize access to AI. We knew deep learning could help people from varied domains tackle problems that they cared about, but at the time, only people who had studied at a small handful of academic labs were able to use it. We were troubled by the idea that a homogeneous, exclusive group would be the only ones able to use AI, thus giving them control of it. If AI ended up as powerful as we anticipated, this would mean a small group controlling huge amounts of money and power.
Our students proved the doubters wrong. They worked incredibly hard, achieved world-class results, and conducted cutting edge research. We set world records in the DawnBench competition, and fast.ai alumni went on to join top AI foundation labs and found their own AI companies. AI did become democratized, and part of that was thanks to fast.ai.
Fast.ai became Answer.AI, and the mission grew. Eric Ries joined as co-founder of Answer.AI, and we brought on a small team of people, nearly all fast.ai alumni. The problem was much the same as before, in a new form. It was no longer that only a small group of people knew how to create models, but that a small group of people controlled the models that mattered, and also controlled the narrative around that.
Answer.AI’s mission is to help society flourish with the help of AI, and we believe that human flourishing has to be human-centered. Neither fast.ai nor Answer.ai has ever focused on trying to create ASI or AGI. Instead, we focus on what humans do and how we can help them do it better with AI. This has become quite distinctive. All the big labs focus on automating away the human and creating a kind of a society that doesn’t need humans.
Many from the fast.ai community have now joined the SolveIt community. Over the past 2 years, we’ve amassed a thriving community of 3,000 beta users collaborating on a bustling discord server. SolveIt members share their projects. Many of us love to use SolveIt for close reading articles or books in more depth and for learning new topics. We are united by a shared sense of valuing human growth, human learning, human creativity, and human autonomy. Our common need is understanding of our work.
Come and Learn With Us
We’ve found SolveIt incredibly useful for our team, both for our own learning and for R&D. In fact, we’ve noticed that SolveIt lends itself to a way of working where learning exercises turn into R&D. For example, when I (Jeremy) read the Recursive Language Models paper in SolveIt, I re-implemented it interspersed into the paper as I went, and ended up with something that was a more concise and effective implementation compared to the original paper.
That’s the kind of work we built SolveIt for, where you come out of a session understanding more than when you went in, and sometimes with something new to show for it. If you’d like to work this way too, the best place to start is the first lesson of our How to Solve It with Code course, where we walk through SolveIt and show how our team uses it:
And if you want to know why we work this way, Jeremy’s keynote from the AI Engineer Conference, Growing on Purpose, covers the decades of research on human flourishing behind our approach:
You can try SolveIt for yourself by signing up here now: solve.it.com
We hope you find the tool we’ve created to support genuine human flourishing useful. But we would be happy even just to have you join our mission of putting humans front and center, and of valuing your own growth and learning.