gpugreg

Sorting would certainly be great. I'd especially welcome sorting by some kind of quality metric. The "thumbs down" count is not available via API anymore, but you could use thumbs up per views as a proxy, although that might require some reweighing, since videos with many views usually have fewer thumbs up.

You could also add another layer of voting on hackertube.org, so you'd get a more hackernews-focused voting signal.

How do the categories work? I'd download the video transcripts and embed the text (maybe summarized first). This way you could support custom categories via embedding similarity.

The syntax was specifically chosen to avoid confusion with other tag formats. XML could show up in any user- or LLM-generated text and would mess up the message format. Some variations even used unicode symbols like "｜" or "〈" to decrease the chance of confusion.

Also, back then, some people believed that it could prevent prompt injection, but history has shown that it was a futile attempt.

Context is just one long string. You can do whatever you want with it. Training helps to conform to specific syntactic idiosyncrasies, but LLMs are large enough these days that they can follow pretty much any instructions most of the time.

For example, if you want to use an LLM API that does not natively support tool calls, you can tell the LLM to emit the tool calls for your preferred tools in a markdown code block and parse them yourself.

Amusingly, the agent also spoke German with an English accent. I guess it takes a bit to switch contexts.

I'll believe in AGI when Google can figure out how to leverage their model to make their user experience not suck.

    > LLMs are super lazy about caring about wall clock time.

My bandaid for problem is to prefix every bash command with "timeout 60 " at the harness level and remind the LLM to hurry up. Of course, this is not a perfect solution, since LLMs might try to run jobs in the background or split up work into smaller pieces instead of running smaller experiments. And of course this does not help with the multi-million dollar training runs, but it has served me well most of the time.

llama.cpp is already extremely fast for single-token responses (<5 ms). I can't see Jev being faster when taking network latency into account, except maybe for multimodal inputs.

    > I was training custom ML models back in 2017.

Maybe you are a good person to ask my question then. I have not looked into Jev much, but is it much different from using a regular LLM and constraining its token output to the action space? (e.g. like using llama.cpp's GBNF grammars). Is it just that Jev's "confidence scores" are significantly better than the softmaxed logits? Or is there something else I am missing?

LLMs can decompile assembly code quite well these days, so I doubt that WebAssembly will be a hurdle.