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.

Sandboxing Python by disallowing certain functions was always a fool's errand. Here is a good discussion about it from 2024 with many more links: https://news.ycombinator.com/item?id=42244461

With a webcam, you usually have a fixed background, which is an easier problem to solve. Here is a project from 2020 that can do 4k at 30fps https://github.com/PeterL1n/BackgroundMattingV2

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?

    > Deepseek also came with heavily subsidized plans, at least at first.

Do you have a source for that?

I got a source straight from the hoses mouth, which claims that DeepSeek-R1 was highly profitable:

    > If all tokens were billed at DeepSeek-R1’s pricing (*), the total daily revenue would be $562,027, with a cost profit margin of 545%.

https://github.com/deepseek-ai/open-infra-index/blob/main/20...

With DeepSeek-V4.1-Flash, the memory requirements for the KV cache have been reduced by over 50x and FLOPS by over 5x, so it is even cheaper: https://arxiv.org/pdf/2609.19969

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

I could find are a bunch of Photopea repositories on GitHub, but the authors are all either from China or Russia, so getting damages for infringement will be difficult to enforce. Hiring a lawyer sounds like a waste of money to me.

Only 17%? I'd say driving down the price of AI helps at least 95% of the world's population. It certainly helps me.

I was wondering whether the model could help me with German bureaucracy. Unfortunately, the answer is "no".

More specifically, I asked the model about what I should put on my contact page, which can cost you in the order of 500 € in Germany if you don't write the right magic words.

Kolibri incorrectly referenced the "Telemediengesetz" ("telecommunication act"), which has been superseded by the "Digitale-Dienste-Gesetz" (DDG, "digital services act") since 2024. The model knows about the DDG, but does not reference it unless specifically instructed to do so.

If anyone of the developers reads this, you can fix this by introducing a recency bias during training. You can even control it by conditioning the model on a date provided with the system prompt or first prompt, so you can travel in time.

This article suggests a massive oversupply of GPUs for inference, but I don't see any mention of training. I was under the impression that training makes up a significant fraction of OpenAI's/Anthropic's compute budget. Or is that not the case anymore?

How did you measure energy usage?

Edit: I found a linked article that mentions the inference provider who does the measurements.

Called it three weeks ago: https://news.ycombinator.com/item?id=49595431

The full-response APIs of many providers have been inadequate for a while now because their timeout intervals do not account for lots of thinking. You can use the streaming API to avoid timeouts.

Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.

I don't know the tool used here, but DeepSeek-V4.1-Flash with bash can create a somewhat similar SVG: https://asdf10.com/diagram.svg

Disclaimer: I cut it off after a few minutes because I got impatient. It could have gotten closer if I waited longer.

I agree that ASCII is definitely cooler, for example when personalizing the hashed data as in your example.

    > 1024 bit keys with recordings of not too old data can probably be found

I think GitHub might turn into a scary vector of supply chain attacks in the foreseeable future. There is a five digit number of users still running around with 1024 bit RSA keys.

Here's a larger partial hash collision (108 trailing bits):

    echo 23ca73454a1b981fe51cad0dbd05f4e696795ba67abb28c61aea1a024e5bbeca | xxd -r -p | sha256sum
    echo a16a8141361ae9834ad171ec28961fc8a951ff1bfc3a9ce0dc2fcdbdfa2ccd35 | xxd -r -p | sha256sum

From this post from 6 years ago: https://www.reddit.com/r/crypto/comments/guctw4/finding_sha2...

When running this ssh command in qterminal, I can not click on the URL because it refreshes faster than I can right-click and click on "Open Link". Do I have to type the URL by hand or is there a workaround?

EDIT: Found a workaround. Double-click the URL and paste it with middle-click.

I thought the same when someone promoted their platform for selling image prompts here on HN, but now they have at least 100k sales. I guess there will always be people willing to pay for something if it takes even the slightest bit of effort.

For me, DeepSeek-V4.1-Flash works very well for CUDA kernel optimization. Access to ncu (NVIDIA Nsight Compute CLI) also helps.

    > Why would I pick GLM over Claude?

To support the company that makes their model weights available for download, while Anthropic lobbies to restrict access.

It wasn't a secret either. They blogged about it last month: https://z.ai/blog/glm-5.3-flash#:~:text=Serving%20at%20Scale...