darknoon

I frequently run into scenarios where it won't let me generate the email within 1password on a website, and I have to go to Fastmail and then manually do it. Is this something you have bene able to work around?

the `claude` binary is essentially a packed copy of bun + the js code, so this will replace the native runtime part of claude code.

This relies on knowledge of the distribution, just querying in the middle of A = [1, 2, 4, 8, 16, ..., 2^(n-1)] is slower than binary search

I haven't seen one that worked properly—can you list a couple examples? Some of the ones that say they're "AI" are just VTracer / Potrace and don't give nice control points.

I think you'd find that it's far from "any human" who can do this without looking anything up. I have 15y of dev exp and couldn't do this from memory on the cli. Maybe in c, but less helpful to getting stuff done!

really weird graph where they're comparing to 3x H100 PCI-E which is a config I don't think anyone is using.

they're trying to compare at iso-power? I just want to see their box vs a box of 8 h100s b/c that's what people would buy instead, and they can divide tokens and watts if that's the pitch.

Here's the problem, you're still going to get scraped and the LLM will understand it anyway. Maybe at best you'll get filtered out of the dataset b/c it's high perplexity text?

The developers also gave a talk about Helion on GPU Mode: https://www.youtube.com/watch?v=1zKvCLuvUYc

Here's the thing, they've completely given up and started making their (inferior to AMD) CPUs on TSMC. For example, Arrow Lake is on TSMC N3B. So it's not getting amortized over anything at all and their valuation is going to 0.

It's ok, somewhere between a qwen 2.5 VL and the frontier models (o3 / opus 4) on visual reasoning

In ML, often it does work to a degree even if it's not 100% correct. So getting it working at all is all about hacking b/c most ideas are bad and don't work. Then you'll find wins by incrementally correcting issues with the math / data / floating point precision / etc.

If you were doing a lot of scraping, you could just solve this on a GPU in 1/10 or less of the time it takes a human's phone to do it. Generally you need a decent computer to render a webpage while scraping it these days, so I don't see what this is solving.

anyone know why they mix in the 3 previous tokens? could have just as easily done 5 or 2 right?

One vote for image inputs here. I would love a fine-tuned qwen-2-vl-72b on demand, but most of the solutions are "talk to us" level expensive. I'm assuming you beat the price or convenience of a replicate / modal solution?

I think just reading the code wouldn't make you a good programmer, you'd need to "read" the anti-code, ie what doesn't work, by trial and error. Models overconfidence that their code will work often leads them to fail in practice.

this is somewhat similar, but diffusion transformers typically use a pre-trained text model as the text conditioning whereas, in this case it's integrated and trained together multimodally.

I tried this w/ AM5, but realized that despite there theoretically being enough lanes for dual x16 PCI-e 4.0 GPUs, I couldn't find any motherboards that are actually configured this way, since dual-GPU is dead in consumer for gaming.

Relevant video about some of the history of superconducting computers: https://www.youtube.com/watch?v=14r2oMsAaE8

Why does this webpage have auto-playing audio?

No, it's connected to AirTrain, which is slow and unpredictable, which is then connected to either the A or the LIRR.

I think Vision Pro isn't really a product, it's for developers / early adopters to make apps that will then be available once a consumer version is ready.

in the video, it seems like you were designing something more complicated than the logos I was expecting-more of a vector illustration. It seems roughly in line with stable diffusion, and other models which struggle to make the clean, symbolic logos I would love to have.

It's striking vs the France chart, though I wonder how this reflects electricity import / exports https://www.energy-charts.info/charts/energy/chart.htm?l=en&...

Interesting context I would like to know

Would be more interesting if Pytorch with MPS backend was also included.

Please reword this clickbait headline. It's just disk offloading.

Whisper v3, just a couple weeks ago https://huggingface.co/openai/whisper-large-v3

Stable Diffusion + fine-tuning is quite powerful for creating specific art styles that aren't easily described to DALL-E / MJ

It's a completion model, which can work a lot better for certain use cases.

Eg I can ask it to generate a huge chunk of code and it doesn't try to give an "example", it generates a realistically long bit of code.

This is great particularly when you want to generate an entire webpage, vs having a chat with an agent that's going to tell you how to build a webpage yourself with small snippets of example code.

They did a price drop on it, seems like it will stick around until they can make the Quest 3 a more consumer-friendly price.