mlajtos

I thought the name was intentional. :)

That part about passthrough strongly reminded me of Turing’s Unorganized Machines (randomly wired NAND-gate networks): https://weightagnostic.github.io/papers/turing1948.pdf (worth a read from page 9)

I wish I was thought math with APL.

Time is indeed a flat circle.

APL desperately needs its renaissance. Original vision was hand-written, consistent, and executable math notation. This was never accomplished.

If you are into this, read ahead: https://mlajtos.mu/posts/new-kind-of-paper

What was your motivation to do this?

No, GPUs are not needed for efficient inference.

https://arxiv.org/pdf/2411.04732

Extremely valid counter points, thank you.

We already interact with false reality through our "old school" computers – internet is full of bots arguing with each other and with us. But my proposition doesn't have to distort the interpreted content.

Neural nets (RNNs) are Turing-complete, so they can simulate web browser 1:1. In theory, of course. Let say we find a way to train a neural net to identically simulate web browser. The weights of this blob might at first seem like an opaque non-sense, but in reality it would/could contain a more efficient implementation than whatever we have came up with.

Alan Kay believed computer science should take its cues from biology. Rather than constructing software like static buildings, we ought to cultivate it like living, evolving organisms.

I know about Scamper, but it is dead.

I've been thinking about SmallTalk web browser 4 years ago, more in-depth here: https://www.reddit.com/r/smalltalk/comments/jnigzb/native_we...

Since then, a lot have changed. One dedicated SmallTalker with LLM-infused Squeak might do wonders.

I have never understood why nobody wrote a web browser on top of SmallTalk.

I propose a different lens to look at this problem.

A neural net can be defined with less than 100LoC. The knowledge is in the weights. What if we went from source code of the web (HTML, CSS, JS, WASM) directly to generated interactive simulation of the said web? https://gamengen.github.io

What if this blob of weights could interpret way more stuff, not just the web?

Without looking into the source – GoldenLayout, right? :)

Never used it. https://github.com/mlajtos/mosaic/blob/master/DesignDecision...

What kind of apps?

Instead of content, put there webview/iframe – https://github.com/mlajtos/mosaic

Time to rebuild Mosaic. https://github.com/mlajtos/mosaic

Safari can't handle the demo. https://dockview.dev/demo

Problem was that iPadOS 18.0 did not support alive graphs.

You are right, graphs are alive since 18.1. I fixed the article. Thank you.

You are right, graphs are alive in 18.1. I fixed the article. Thank you.

You can try out MyScript Calculator, which is doing exactly that – https://www.myscript.com/calculator

I prefer to keep my handwriting over the printed symbols. This might sound crazy, but when I see my handwriting transform into printed symbols, I feel as if the system stole it from me. :) Another crazy take – it is more difficult to read handwritten script, so you slow down your thinking, which might help you to think more deeply. I can't back these claims with any evidence other than "I like it that way"..

> But the bad feedback criticism is spot on, though the better solution is to replace symbols in-place, why would you prefer your "awful handwriting" to beautiful math symbols with excellent legibility?

You mean it should replace your handwriting with typed symbols?

I want to see more threads like this one – people using Math Notes for all kinds of needs. :)

Yes, the first version is already usable for everyday simple stuff. In the future, it will definitely gain more features. But you are wrong in one thing – it will become much-much more than graphing calculator.

You are not wrong. In 2021, I did bootstrap everything on my own – dataset, segmentation algorithm, neural net, UI, math engine, etc. And in the end, it worked properly only for me. But that was my aim – own complete stack to be able to try out everything that might be interesting without being burdened by 3rd party solutions. It was a prototype that I could show to people – spread the idea. Doing product of that magnitude on my own would never fly.

I don't think Apple did a "cool tech demo". In my view, they published first version of a very ambitious project. They worked on it for years with top talent. And they felt this is the minimum lovable product on which they will iterate.

And if you would like to build product-ready solution similar to Math Notes, you can leverage existing stuff. MyScript.com have really good handwriting recognition system with bells and whistles. You can take a lot of shortcuts.

It is true that in the WWDC24 keynote, Jenny Chen demoed a graph that could be changed by dragging a slider of one of the variables [0], but as you can see in demo in the article, graphs are dead – they do not respond to changing variables. Or maybe I am just doing something wrong...

[0]: https://youtu.be/RXeOiIDNNek?t=2900

Yes, math engine works differently when you are in Notes and Math Notes. There is a setting ("Insert results" vs. "Suggest results") that controls this behavior.

Yes, the handwriting recognition system needs to communicate results as soon as possible to instill trust in the user. I need to be able to spot and fix misinterpretations as quickly as possible, otherwise errors accumulate real fast.

I thought the same few years ago – putting recognized symbol above is simple and effective. But you can get the same effect (confidence in recognized input) with the Smart Script technique when done right.

I briefly experimented with that in 2022 (nothing published though) and it worked pretty well. However, the implementation has to be good – you can't just beautify every symbol because it becomes distracting really fast. The less confidence system has about some symbol, the more it should fix it. High-confidence misclassifications will still be a theoretical problem, but I didn't experienced them that often.

Fair point. Interactivity is a spectrum that should match the user preference. Interactivity helps when you are learning the tool, because you can do & fix mistakes faster. But when you want to think deeply, it should stand away and be there only when you explicitly request it.