brrrrrm

Tree decoding exists and is being worked on by multiple groups. At the end of the day tree based methods are still just injecting some kind of interpretable prior.

Training is 3x the memory used by inference, and usually run at a much larger batch size

A provable way to recover convergence is to calculate the hessian. It’s computationally expensive but there are approximation methods.

Stuff like this always makes me think: what super structure do our brains form? Always end up concluding it’s the economy

The wings are doing a lot more work than the tail in that situation. For quick rotations in 3D, gyroscopic forces would probably be better to utilize than a tail

Cranking up the batch size kills convergence.

I predict knowing or caring about talking to a human via online anonymous chat is going to be a dated idea within five years.

Might seem crazy today, but I think it’ll boil down to the conceptual equivalent of “is this text spell checked?”

Not really. It’s just a really simple hype-tech mashup and should be called out for it.

Are there any Reddit alternatives these days? Something like HN but more slightly more casual (image/video embeds, subreddits)

It looks like it's pretty resistant to quantization. ollama 4bit 7B doesn't work very well, but the 16bit 2B does

whoa, lots of negative reactions to a pretty mundane comment I made about the science of it all. It's overwhelmingly likely, from other studies, that brains differ, sure. But, the AI derived signal mentioned in this study might not be picking up on that? In fact, it's hard to conclude as such from the experiment run simply because it was run on adults. Run it on infants to control for that confounding factor.

article about the paper: https://med.stanford.edu/news/all-news/2024/02/men-women-bra...

"When the researchers tested the model on around 1,500 brain scans, it could almost always tell if the scan came from a woman or a man."

It's pretty cool to see blackbox models derive reliable signals from basically raw data. However, the results don't really imply sex differences are biological in nature if I'm reading correctly. Being raised male/female probably has some profound impacts on the brain (these were adult brains). I bet a similar methodology would be able to tell who played American football growing up...

Recycled commentary on product changes made by an app trying to grow its ad revenue. Honestly this article could be AI generated

Not to slight the author, but Rust seems like quite a hard language to learn on the job.

OMP_NUM_THREADS=n is pretty standard in HPC workloads

Whoa thanks!

PeachPy is awesome - it's really a great way to play with assembly-level performance programming and not have to deal with assembler toolchains

I haven't found a library that makes me feel comfortable with using WebRTC for random little projects. Setting up TURN/STUN is annoying and I've actually never successfully setup a server/client model (it seems like it'd be useful to have the fast UDP-like stuff, despite WebRTC mostly being about client/client communication)

I work professionally in C++, but I wouldn’t tout it either.

I find Python to be the most pleasant personally.

Weird to think typescript is similar to a language that requires a class definition for “hello world” lol. You can’t even fit it on a single 80 character line…

yea, I guess it's the lack of human-friendly default assumptions mixed with prescriptive design patterns. definitely an anachronistic design intent :(

it feels like the characters-typed to functionality ratio is quite low. at the same time, the code isn't much cleaner and there's a lot of mental overhead in deciphering the abstractions (in this case Collector).

Isn’t it run multiple times and the slowest is dropped? Why would you expect the file to be evicted from ram so quickly?

looking at the sample code I’m quite glad I’ve never really touched Java. What a clunky language

matmul is old and useful -- there's a lot of hardware on a chip that makes it run much faster (prefetch, vectorization, instruction parallelism) and some of these languages have optimizations to expose those things automatically.

doesn't look that easy... zend_parse_parameters? pre-baked configure + make scripts?

check out how a modern language deals with this stuff https://bun.sh/docs/api/ffi#usage

whoops, I forgot to include the URL. dang would you mind deleting this?

great line: "cynicism masquerades as wisdom"

as documented it looks different (async vs sync + the necessity of `text:` annotations in the Node.js version?)

updated my comment to paste in what's written in the docs

why on earth did they design the Node.js and Web APIs to be slightly different and incompatible? (edit: this might just be a bug/oversight on the landing page?)

Node.js:

   const model = genAI.getGenerativeModel({ model: "gemini-pro-vision"});

   const result = model.generateContent({ contents: [{parts: [
      {text: "What’s in this photo?"},
      {inlineData: {data: imgBase64, mimeType: 'image/png'}}
    ] }] })

Web:

    const model = genAI.getGenerativeModel({ model: "gemini-pro-vision"});

    const result = await model.generateContent([
      "What’s in this photo?",
      {inlineData: {data: imgDataInBase64, mimeType: 'image/png'}}
    ]);