djsjajah

funny, I am not surprised that it is now much harder to run an automated twitter account for legitimate purposes. Running one for illegitimate purposes on the other hand....

It’s been renamed recently. Maybe a few times. I think it’s rocq now. [1]

[1] https://rocq-prover.org/docs

You have to power it all the time, but the amount of power it uses while it’s on will change by up to a few orders of magnitude depending on the gpu. It’s not uncommon for a gpu to be pulling just a couple of watts at idles and several hundred at full tilt.

So the only way it’s a fixed cost is if you don’t pay for power. If you only consider the cost of the power, it might still be cheaper paying for an api.

but what is the point? A ban is supposed to make a certain thing less likely to occur. Does a ban of open source models do that? Presumably, the behavior you are trying to limit is the miss-use of these models but I don't know how many state sponsored hacking groups are going to give a ban a second thought.

I think they were making a joke. In the future, you might consider the advice you are giving as well as giving it.

Yes. Wait a day

Not with 800 examples. If you are going to consider an ngram model, I think you are better off getting a frontier llm to write you an absurd regex.

Except, that won’t help. By the time a new fab is up and running, we will probably have a massive surplus.

You need to think this thought through all the way to the end. What it has said also influences what it will say. If it has consistently made combative responses, then the most likely thing to do is to continue to be combative.

I don't think there is any way back after the conversation takes a turn like that so there is no point in arguing with it. The only thing you can do is to fork the conversation before it made the first mistake and give it more context or tell it to look things up.

It’s amusing that a lot of the agents have worked out that sampling doesn’t change ppl.

I think what they mean by “now” is the stuff announced today.

I don't follow. Can you explain how your comment is relevant to mine? It might help if you also explain how you interpreted my comment.

You just failed the Turing test.

I have 2 of them. I would advise against if you want to run things like vllm. I have had the cards for months and I still have not been able to create a uv env with trl and vllm. For vllm, it’s works fine in docker for some models. With one gpu, gpt-oss 20b decoding at a cumulative 600-800tps with 32 concurrent requests depending on context length but I was getting trash performance out of qwen3.5 and Gemma4

If I were to do it again, I’d probably just get a dgx spark. I don’t think it’s been worth the hassle.

> or by the community

Hmmm

yes, but the difference between one model and one 4x larger is usually a lot more than that.

It is not a question of do a run Qwen 8b at bf16 or a quantized version. It more of a question of do I run Qwen 8b at full precision or do I run a quantized version of Qwen 27b.

You will find that you are usually better off with the larger model.

trl. give me a uv command to get that working.

But even in the amd stack things (like ck and aiter) consumer cards are not even second class citizens. They are a distance third at best. If you just want to run vllm with the latest model, if you can get it running at all there are going to be paper cuts all along the way and even then the performance won't be close to what you could be getting out of the hardware.

No. It seems to me that the comment is objectively incorrect. The original comment was talking about inference and from what I can tell, it is strictly going to run slower than the model trained to the same loss without this approach (it has "minimal overhead"). The main point is that you wont need to train that model for as long.

That’s kind of a moot point. Even if none of those overheads existed you would still be getting a a fractions of the mfu. Models are fundamental limited by memory bandwidth even with best case scenarios of sft or prefill.

And what are you doing that I/O is a bottleneck?

> including all previous experiments

How far back do you go? What about experiments into architecture features that didn’t make the cut? What about pre-transformer attention?

But more generally, why are you so sure that they team that built Gemini didn’t exclusively use TPUs while they were developing it?

I think that one of the reasons that Gemini caught up so quickly is because they have so much compute at fraction of the price of everyone else.

Not only can it be streamed, but lz4 will probably make things quicker.

You just ruined my day. The post makes it sound like gel is now dead. The post by Vercel does not give me much hope either [1]. Last commit on the gel repo was two weeks ago.

[1] https://vercel.com/blog/investing-in-the-python-ecosystem

> Do you really though?

Yes.

It stays in on the hbm but it need to get shuffled to the place where it can actually do the computation. It’s a lot like a normal cpu. The cpu can’t do anything with data in the system memory, it has to be loaded into a cpu register. For every token that is generated, a dense llm has to read every parameter in the model.

GPUs might not be bandwidth starved most of the time, but they absolutely are when generating text from an llm. It’s the whole reason why low precision floating point numbers are being pushed by nvidia.

I can't tell if you are making a joke or not.

They are not even remotely equivalent. tinygrad is a toy.

If you are serious, I would be interested to hear how you see tinygrad replacing CUDA. I could see a tiny grad zealot arguing that it is gong to replace torch, but CUDA??

Have you looked into AMD support in torch? I would wager that like for like, a torch/amd implementation of a models is going to run rings around a tinygrad/amd implementation.

I went to check how many services are being impacted on down detector, but it was down.

If we didn't have people like that, then they would be right.

I few people have mentioned dagster and I took a look at that for some machine learning things I was playing with but I found dvc (data version control [1]) and I think it is fantastic. I think it also has more applications than just machine learning but really anything with data. If you have a bunch of shell scripts that write to files to pass data around, then dvc might be a good fit. it will do things like only rerun steps if it needs to. Also for totally non-data stuff, Prefect is great.

[1] https://dvc.org

No. Kids would need to memorize the private key of their parents id card.

You could have also self-hosted the GitHub Actions runner which might have been easier as long as you had something to run the runner on.