jinay

Timestamped link to the YouTube video: https://youtu.be/T08wc4xD3KA?t=124

Ooh I haven't, but this is exactly the kind of follow-up I was looking for. Thanks for sharing!

I recently did a deep dive on open-endedness, and my favorite example of its power is Picbreeder from 2008 [1]. It was a simple website where users could somewhat arbitrarily combine pictures created by super simple NNs. Most images were garbage, but a few resembled real objects. The best part is that attempts to replicate these by a traditional hill-climbing method would result in drastically more complicated solutions or even no solution at all.

It's a helpful analogy to understand the contrast between today's gradient descent vs open-ended exploration.

[1] First half of https://www.youtube.com/watch?v=T08wc4xD3KA

More notes from my deep dive: https://x.com/jinaycodes/status/1932078206166749392

Soon we'll start having a humans.txt format to account for this

https://xkcd.com/927/

Please add an llms.txt file! https://llmstxt.org/

I'd love to see how far I can take this by giving it to an LLM and asking it to format for me with Quarkdown.

https://github.com/JinayJain/dictator

Built one for myself. It's context-aware and promptable.

Tested well on Linux, not so much on other platforms but in theory should support them.

It's a bit meta but I wrote it mostly using Claude Code. Once I had an MVP, I was able to prompt much faster by just speaking out what I wanted it to change.

I used to think that typing speed was not really that important, especially when now we have so many LLMs doing the typing for us. But honestly, now I think it's even more important because the specificity and detail in your prompts are paramount to getting a good response, and something like a dictation tool (which is what I'm using right now) is really good for generating very specific prompts.

In fact, I wrote all this out using a dictation tool in ~20 seconds (258 WPM).

Went in an incognito window and searched "io" and this announcement was shown right above Google IO [1].

[1] https://i.imgur.com/xNKjFXa.png

I'm more referring to launch timing

Likely no coincidence that they announce their company, io, during Google I/O.

Search "io" on Google right now and see what comes up...

That's not the only concern though. The web is by far the most accessible way to distribute something, which is paramount in getting people to use it.

They really need to fix the kerning on some of their charts

I wonder if you could use the Chrome local AI API to build something like this: https://developer.chrome.com/docs/ai

For anyone who wants to listen, it's on this page: https://yummy-fir-7a4.notion.site/dia

Make sure you're using the "-it-qat" suffixed models like "gemma3:27b-it-qat"

Cursor is likely very tuned for Claude (prompts-wise and all) due to its dominance with 3.5 and now 3.7. Still, Gemini 2.5's tool calling has been pretty poor in my experience which Cursor heavily relies on.

Fun fact: the Codeium office is actually the Pied Piper office from the show [1], so yeah, they like Silicon Valley.

[1] https://x.com/codeiumdev/status/1851400416946462967

How does it stack up against GitHub's own search?

Hmm, I've never used /. so it wasn't me who shared it.

Yep, it does some naive rescaling of the clock to make it circular (since perspective would make it more like an ellipse in the image), but then assumes 12 is always at the top.

Thanks!

Thank you!

Woah generalized pose estimation seems really useful and it's cool that the example is for reading a clock. Thanks for sharing!

When I was first learning computer vision, I wrote a program that could tell the time from an image of a clock [1]. I had no purpose for it besides the fact that it seemed like a cool problem to try and solve.

Years later, I get an email from a stranger in Korea, asking me how to run my program. Why would he want to use my silly program? Turns out you can adapt the code to read analog pressure gauges which is really useful for chemical plants. Goes to show that there's often a use for most things.

[1] https://github.com/jinayjain/timekeeper

Thanks for taking the time to read! I tried to be brutally honest.

I'm in the process of building a similar project aimed towards beginners, so it's hopefully a bit less daunting than Cardinal/VCV Rack.

https://wavepen.jinay.dev/

I built a simple autodifferentiation engine [1] in JavaScript when I was first learning about backpropagation. It's a great way to learn how frameworks like PyTorch work.

[1] https://blog.jinay.dev/posts/backprop/

Did Shannon use it for something or just as an experiment?

It's almost guaranteed that there will be quantizations and adaptations to consumer hardware that should make it both versions of the model accessible to mortals on consumer GPUs (and even CPUs).

That seems really useful, thank you!