Solveit learning platform

I recently completed a course hosted by Jeremy Howard, John Whitaker, and Eric Ries. The Answer.ai team built Solveit, a code-first learning platform designed to teach in small steps with quick iterations and immediate feedback. The course taught both the Solveit method and the platform itself.

The Philosophy Behind Solveit

Solveit draws its name and inspiration from mathematician George Pólya's classic book How to Solve It. Pólya breaks problem-solving into four steps:

1. Understand the problem
2. Develop a plan
3. Execute the plan
4. Reflect on the results

What the Course Covers

The course spans 10 lectures covering a surprisingly broad range of topics. Here's a taste:

• Agents from scratch — We built the tool loop on top of liteLLM. (Example dialogue)
• Web development with FastHTML — I built an AI evaluation tool for talking like a fighter pilot
• Algorithms — Using Advent of Code as a playground
• Infrastructure — SSH tunneling, spinning up VMs, and VPNs
• Paper and book reading — We got a first look at Eric's new book and walked through LeJEPA, Meta/Yann LeCun's new paper

It's hard to say exactly what the course is because it covers so much ground, but the goal is simple: teach you how to do stuff that's commonly needed.

How the Platform Works

Solveit gives you access to a virtual machine where you develop dialogues. These work like Jupyter notebooks, but with prompts as a cell input option. The LLM (Claude) can see every cell above it in the notebook, giving it full context for better answers.

Prompt Modes

There are three output modes for the built-in agent:

• Learning mode — This is powerful. It asks clarifying questions to calibrate answers to your level, then leads you toward the solution like a good tutor would. It doesn't just hand you the answer.
• Concise mode — Short, direct output. Great for things you already know or quick tasks. I used it recently to create a flowchart in Graphviz.
• Standard mode — Direct LLM output from Claude. (I never use it.)

For hard problems, there's a thinking mode that can be engaged with any of the prompt options.

Standout Features

Web Development

You can iteratively build web apps and render results directly in the dialogue, or add routes to view in the browser. There's an option to get a public URL so you can spin up a live website in minutes.

Code Learning

Tools let the agent read a new repo's source code into memory. With learning mode enabled, it can teach you how to use the repo. This lets you go from noob to proficient in a staggeringly short time.

General Learning

A screen-sharing tool lets you share screenshots with the agent, even from applications on other machines. In class, Jeremy demoed learning video editing software this way. I used it to set up SSH tunneling between my desktop and the platform's GPU. The agent helped me debug and set proper permissions.

Final Thoughts

Solveit gets you into a flow state for development because of the quick feedback cycle.

This is the second iteration of the course, and both the platform and tools have come a long way in the past year. But so have I. I committed to doing the homework and keeping up with lectures, and it's been incredibly rewarding. It's like having a personal tutor calibrated to your skill level.

What really blew me away was when Johno used SSH tunnelling to connect to a Raspberry Pi camera and his 3D printer. He then made a web app to control the 3D printer which moved slides on a microscope and displayed the microscope image via the Raspberry Pi camera. Ten out of ten. It's already paying dividends at work and in personal learning.

If you're reading this and want to try Solveit, email me and I'll send you a referral code.

---