Claude 3.7 has clearly been specially trained to be good at generating React code and SVGs.
- Claude 3.7 has clearly been specially trained to be good at generating React code and SVGs.
- They are almost certainly using RLAIF to post-train the model to be significantly better in those domains.
- It probably looks something like this:
- Generate React code to accomplish the user's request for a simple UI.
- Then execute the code and poke at it with Playwright to get screenshots of it, and see how it responds to simulated user actions.
- Ask an LLM to judge the quality of the result.
- Does this look like what the user asked for?
- Does it appear to work as a user would expect?
- This gives an ability to get better and better at it by throwing a lot of compute at the problem.
- Note that this only really works for React code (vs general programming) because frontend code is not algorithmically complex, but rather tedious.
- LLMs do great at writing tedious code; they struggle with novel algorithmically hard code.
- Put another way, LLMs can use React well, but it would struggle to write React itself.