I think for everyone interested in this project checking your OG twitter thread is also somewhat mandatory:

https://x.com/_chenglou/status/2110557373651931267

It took me a bit to wrap my head around what you did.

Clever, I wonder what would happen if you did this in reverse? Like, if we froze the weights then added some layers trained on a bunch of weird edge cases then tried to convert those weights back into branches.

QA through entropy lol

It seems interesting but I am confused. What is the purpose of that landing page? It is not clear since it showcases several (not related) ideas:

- turning js code into something else, arguably high-level code for training neural networks

- a mini OS inside the browser

- a gallery

- document search

I think this is basically a compiler that turns arbitrary JS code into weights of a neural net that then behaves exactly like the JS program. And the stuff at the bottom are demo projects that are now running as a neural net.

the "DEMOS" label is rather small, I understand the confusion

As far as I understand, this doesn't turn "regular" JavaScript into a NN, it's rather "write your logic in a DSL that creates a neural-net-shaped circuit"

the source code for the OS demo is not "regular" code, it's full of `emiRelu`... https://github.com/chenglou/handwired/blob/main/src/handwire...

The main site & readme seem to overclaim, while the docs say things like "With Handwired you write a numeric program" and "No tool converts existing JavaScript source into a program" - https://github.com/chenglou/handwired/tree/main/docs

happy to be proven wrong, to have misunderstood this, etc.

What is stopping Anthropic, Google etc from essentially building models that completely replace the concept of code?

You point a model to an API and get a fully working UI right away. I don’t mean an agent CODED the UI - I mean the model inferred it from its training data and renders it to you “live”

Or the other way around? You serve it a UI- and it acts as the backend etc? State and persistence becomes an issue, but the model could freeze things into deterministic structures (specific parts become -just in time-frozen, for example when data needs to be saved)

Maybe this is old news already- but the consequences are kind of mind blowing

We are ~18 months away from the entire UI being replaced.

Novelty. The models still cannot produce new solutions. They are getting really good at solving tough problems but only if the problem is defined specifically enough in advance to warrant a narrow solution. Otherwise these models are good at replacing the people who aren’t good at this either.

google/tracr https://deepmind.google/research/publications/22295/

Website is missing the #1 thing I imagine most readers would want to know: why?

Lol, if you pass a = 4.5 to their homepage const result = (a > 4) ? 10 : 12 you get 11. Not everything is differentiable.