etaioinshrdlu

When I made my ow GPU-renderered framework I found that it was possible to render the entire UI in one call/shader to the UI. All primitives and styles can be passed as a updatable data blob that the single GPU shader handles in one call. Do most of the GPU-rendered frameworks do it this way?

It has some tradeoffs to consider but seems quite good overall.

Speaking of, some of you may remember "ch.at" https://news.ycombinator.com/item?id=44849129

It is now upgraded with agent messaging: https://ch.at/board/feed

Still usable over DNS ;)

I thought the Macbook Pro LCDs were bright enough, high-res-enough, and fast enough to make a very accurate CRT simulation. Here it is: https://github.com/baragona/crt-xdr

It looks so much like a CRT to me. It looks like pressing your face up to a TV to see the individual phosphors. Try setting "signal" to dropped image and viewing some images. It looks like an old TV.

Bottom line, whatever you think a CRT is doing, you can probably simulate it quite accurately.

I'm a bit confused - when i see these signups, they look like this to me:

rtwnj6tj7@privaterelay.appleid.com

I wonder how they plan to detect it something is LLM generated. I think what this leads to is people just working hard to make their outputs appear human generated.

Related topic - some of you may appreciate this great album! https://en.wikipedia.org/wiki/IBM_1401,_A_User%27s_Manual

I would also like to toss my project into the ring, which also allows mixing most real-world Rust & Go packages together, on runtime & syntax compatible level. https://github.com/deepai-org/omnivm

Pardon the sloppy readme - it actually does work :)

It's true. Also, if you despise bloat, you may like my JS-free LLM site: https://ch.at/

I wonder if Windows NT could work? I think it had a MIPS port?

It's a fascinating article and trend to me. I've been rather obsessed with the amazing technology of text-to-image generation since 2017, when state of the art was an LSTM+GAN and resulted in a blurry image. Now that the technology basically works great, it's just upsetting to a lot of people. I kind of think of AI like making things out of plastic - works pretty well, but basically always resented. Notice that the article couldn't identify anything wrong with the generated image except for how it was made and how no one got paid.

I think the font is mimicking old Apple ads, eg: https://i.insider.com/5bf8592eb73c284de50e2f28

This is super cool. If you're interested in this, you might enjoy this proof of concept as well: https://github.com/deepai-org/omnivm

It instead embeds a bunch of runtimes onto the same OS thread.

What do you think is the incentive to LLM post on HN (or any site?)

It makes sense - i build something very similar for my company over the last couple weeks :)

I have a tweak that allows pasting images to claude code over SSH:

How it works:

PTY Interception: It creates a pseudo-terminal (PTY) to wrap the SSH process, allowing it to sit as a "man-in-the-middle" between your keyboard and the remote shell.

Bracketed Paste Detection: It monitors stdin for "bracketed paste" sequences (the control codes terminals send when you Cmd+V or drag-and-drop a file).

The "Hook": When a paste occurs, it pauses execution and scans the text for local macOS file paths.

Auto-Sync: If a local path is found, it immediately syncs that file to the remote server (using the provided SSH key) in the background.

Transparent Forwarding: Once the sync is complete, it forwards the original text to the shell.

You can drag and drop a file from your local Finder into a remote SSH session, and the file is automatically uploaded to the server before the path appears on the command line. Also works with copy paste, screnshots.

Like others here, I disagree completely. I find them very fun, almost too fun, like intellectual crack. The craziest ideas are now within reach.

"Before microchips existed, computers were built with mechanical relays." Should probably say something about vacuum tubes as well!

Cool! Right now it's just IP address rate limiting and the costs have not mattered too much, but yes long term I am not sure what we'll do...

I can suggest our service (previously here https://news.ycombinator.com/item?id=44849129 ) that might be helpful -- If you want a zero-setup backend to try qqqa, ch.at might be a useful option. We built ch.at — a single-binary, OpenAI‑compatible chat service with no accounts, no logs, and no tracking. You can point qqqa at our API endpoint and it should “just work”:

OpenAI-compatible endpoint: https://ch.at/v1/chat/completions (supports streamed responses)

Also accessible via HTTP/SSH/DNS for quick tests: curl ch.at/?q=… , ssh ch.at Privacy note: we don’t log anything, but upstream LLM providers might...

This is impressive and it also kind of demonstrates how bloated Windows really is. You can fit a ton more functionality into even 1MB.

My new theory, developing for a while, is that as technology makes things easier, the perceived average quality goes down over time. I've yet to fully understand the factors that drive this trend, but feel certain AI will put it in overdrive! I'm not a luddite or hater actually - but this trend is pretty apparent...

Interesting - I somehow didn't realize that KVM didn't require root access.

Also, I wonder if this could be adapted to use Apple's Hypervisor.framework. That one also doesn't require root and ought to be able to spin up and down very quickly.

One interesting thing I forgot to mention: the server streams HTML back to the client and almost all browsers since the beginning will render as it streams.

However, we don't parse markdown on the server and convert to HTML. Rather, we just prompt the model to emit HTML directly.

Author here, was a bit surprised to see this here. I thought there needed to be a good zero-JS LLM site for computer people, and we thought it would be fun to add various other protocols. The short domain hack of "ch.at" was exciting because it felt like the natural domain for such a service.

It has not been expensive to operate so far. If it ever changes we can think about rate limiting it.

We used GPT4o because it seemed like a decent general default model. Considering adding an openrouter interface to a smorgasbord of additional LLMS.

One day, on a plane with WiFi before paying, I noticed that DNS queries were still allowed and thought it would be nice to chat with an LLM over it.

We are not logging anything but OpenAI must be...

These can definitely be added

Do we know each other :0 :)

Much of LA has some of the worst air in the country, so I think it selects for people who don't care about being poisoned by their environment.

Sounds like bloat removal and minimalism.

What optimizations did they do that had the biggest effect? can they be brought into the mainline linux kernel and distros?

It would be profitable even if we self-hosted the LLMs, which we've done. The only thing subsidized is the training costs. So maybe people will one day stop training AI models.

I've operated a top ~20 LLM service for over 2 years, very comfortably profitably with ads. As for the pure costs you can measure the cost of getting an LLM answer from say, OpenAI, and the equivalent search query from Bing/Google/Exa will cost over 10x more...