menaerus

1st commit of the project was in August 2023 so not quite few months old. And the performance problem according to https://github.com/ClickHouse/ClickBench/pull/1159 has not been yet solved.

So, I think my curiosity still stands.

On 22nd of July

> It was loading the data for a week already, and the speed of loading data has dropped to 4 kilobytes per second, and it will take years to load the dataset.

And then on 2nd of August

> It loaded maybe 1% of the data so far.

And then one year after

> I didn't load the data after a year. > I tried it with the fsync removed, but even then it does not work

Ok, I find this more entertaining than I should and almost unbelievable. I know that this codebase was heavily written by LLMs but the execution can't be this bad?

What am I missing and why would Supabase buy the product which can't even load the dataset properly?

Do you make your professional career by building formally verified systems? I ask because I don't think that the reason comes down to "because most devs don't want to do formal verified system development". It's much more complicated of course.

I had it seen on servers that were typically reserved for the team but there was no official booking system for those machines. When you start using the machine you would typically put some note to make sure somebody else does not overrun your long-running tests or performance measurements.

Too simplistic view which I also used to believe in maybe ~10 years ago. There's a class of problems where such approach doesn't apply.

Also a speculation but I'm almost certain that physical optimizations are first done through simulators running on a computer.

> LLMs are not suitable for that kind of work.

I wonder why not or you meant not suitable yet?

I wanted to understand your background first because what you initially said is a very oversimplified view of LLM mechanics, and generally not quite the way how software is in practice written. Since you didn't answer that question, I will assume that you're not a SWE by a call. To give you an example of what I am trying to convey is: imagine a data-intensive workload hitting your storage/database/kernel implementation, and it's painfully slow, your customers are not happy. Then you as engineer sit down, spend days profiling and understanding the code, researching about existing algorithmic solutions to the same or similar issues found in the wild, you read some open-source implementations of viable approaches, you ditch some, some you take, you also read books, articles, other peoples experiences etc. And finally you end up, let's put it bluntly, with some sharded data structure by which you solve the bottleneck. It's not novel, the technique is so common and is already implemented across many many different products in slightly different flavors so I am wondering why do you think this is not a copyright breach but the LLM, which does more or less the same thing, is?

How do you know this is not true with other vendors? I'm not defending them but I wouldn't believe anyone in this business unconditionally. Anthropic agent fwiw is not open source, gemini and codex are.

I guess you're not writing code much or haven't done much so as your professional career?

They do invent code solution for the problem that exists in your codebase. Latter implies that the code solution LLM synthesizes is usually unique of a kind, so, it's not a copy-paste neither it is a simple extract from "another codebase" and adopted.

IMO they operate pretty similarly to humans - we synthesize our solutions, and therefore build-up our knowledge, by collecting knowledge from multiple other sources, including technical books and blogs, open-source code repositories, and our past experiences.

Benchmarks will be interesting to see but I doubt the cores alone will be able to reach beyond ~500GB/s of BW, and even if they match that figure it will be great.

Do you have a link?

Look into the history clases will be enough. Why would I need to prove anything to a random stranger on the internet?

Unless you have the source for the latter it's nothing more than propaganda. Also, DeepSeek report is 2-3 years old, and at that time there were no huawei chips, at least not known to the public. This is a different model, in different age where huawei chips are already delivered into the production as we see

Why do you think the Chinese are not using their ai chips, which afaics are from huawei, to train their models? If that is true, which I think it is, we don't have to wait until 2030 to see if they're gonna match the performance of nvidia GPU. The evidence is already here - glm is among the most competitive models out there.

Anthropic: 17 USD (pro), 100 USD (max)

GLM: 80 USD (pro), 168 USD (max) -> with "limited-time event" discount this becomes 56 USD and 117.6 USD

I also don't understand why are they so much costlier, and I would also like to give it a try.

It was evident that this will happen.

> Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale.

Does Rust not have SIMD intrinsics that you could use directly instead of relying on auto-vectorization?

Companies are heavily incentivized to do so so my guess would be yes. I also don't write the code by hand almost at all.

Sometimes the winters may be cold, some other times they may be exceptionally warm, and sometimes they're a mix of both. The law remains the same regardless what the winter was or is like. Countries that have exceptionally warm winters do not really have a winter conditions so I'd guess their law wouldn't have a requirement for winter tires or chains

If warranted by the law then yes, you have to abide to it.

I started using gemini with caution given the "much worse" benchmarking points it has gotten and still does but in practice there's very little evidence I found in comparison to claude models. It performs really well on non trivial tasks.

What kind of comment is that? No, I have not bought the h100 obviously, and, yes, you can buy the new one for 20k. Used one for even less.

You're correct, the difference isn't nil - H100 is data-center grade GPU with loads of bandwidth and compute while RTX PRO is a consumer grade GPU. The difference between the microarchitectures would make sense to point out if we had been comparing apples to apples and not apples to pits.

Worktrees are significantly faster to use and checkout than clones. They also use much less disk space, and obviously commits made in one worktree do not require git plumbing to get them visible in another etc.

For that price you get the H100. The machine is overpriced, and I'd personally rather go for a second-hand workstation with server-grade components.

I wonder what the target audience for this price range is? At that price point I would rather go for a single H100.

In certain, more complex, software engineering domains this almost became true as of today.

It doesn't appear to me so