gpugreg

4 ms·

1.
by gpugreg 5d ago
Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.
2.
by gpugreg 6d ago
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 g
3.
by gpugreg 6d ago
I agree that ASCII is definitely cooler, for example when personalizing the hashed data as in your example.
4.
by gpugreg 6d ago
> 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 arou
5.
by gpugreg 6d ago
Here's a larger partial hash collision (108 trailing bits): echo 23ca73454a1b981fe51cad0dbd05f4e696795ba67abb28c61aea1a024e5bbeca | xxd -r -p | sha256sum echo a16a8141361ae9834ad171ec28961fc8a951ff1bfc3a9ce0dc2fcdbdfa2ccd35 |
6.
by gpugreg 7d ago
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 workarou
7.
by gpugreg 9d ago
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 effo
8.
by gpugreg 9d ago
For me, DeepSeek-V4.1-Flash works very well for CUDA kernel optimization. Access to ncu (NVIDIA Nsight Compute CLI) also helps.
9.
by gpugreg 9d ago
> 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.
10.
by gpugreg 9d ago
It wasn't a secret either. They blogged about it last month: https://z.ai/blog/glm-5.3-flash#:~:text=Serving%20at%20Scale...
11.
by gpugreg 10d ago
> DS models seem to get stuck in loops I have never had looping issues with DeepSeek models. Which provider/serving framework and harness are you using?
12.
by gpugreg 12d ago
Which Qwen3.8? Qwen3.8-Max? Qwen3.8-Flash-Next? Qwen3.8-27B? They are all different models.
13.
by gpugreg 12d ago
I stopped using GitHub Copilot extension in VS Code when they introduced their new pricing model, but I should have switched much earlier. The developers have barely a clue how LLMs work and the company structure is misaligned with creating
14.
by gpugreg 12d ago
The question isn't how you keep your context small, but how did your context get so big? A few common sources of bloat are long system prompts, unnecessary tools, unclear prompts, and scrawling code bases. To reduce system prompt and t
15.
by gpugreg 14d ago
I currently use DeepSeek-V4.1-Flash and the V4-Pro and -Flash versions before that, because they are extremely cheap. I have spent less than $25 for over a billion token so far (1B cached, 15M out, 19M in). I even prefer the DeepSeek models
16.
by gpugreg 15d ago
I've had some success with tokens as a measure of complexity instead of number of lines, but should be combined with additional rules, e.g. disallowing lambdas, exec, eval, compile, __import__ and complex list comprehensions for Python
17.
by gpugreg 15d ago
Both DeepSeek-V4.1-Flash and GLM-5.3-Flash failed to decode your embedded example text. I failed, too, but I only spent a minute trying to figure out your repo before giving up and telling AI to do it. Anyway, maybe you want to improve your
18.
by gpugreg 16d ago
For me, "Improve the model for everyone" was "On", although I disabled a similar-sounding checkbox in the past (Germany).
19.
by gpugreg 17d ago
They did, but it was not well-received. Perhaps they want to try something different.
20.
by gpugreg 17d ago
> those things are not deterministic Determinism was an explicit goal of DeepSeek-V4. From their paper: https://arxiv.org/html/2606.19348v1#S3.SS3 > we implement end-to-end, bitwise batch-invarian
21.
by gpugreg 17d ago
It is easier to trust what you can understand.
22.
by gpugreg 17d ago
Is there any cryptocurrency that uses AES?
23.
by gpugreg 19d ago
Not necessarily. These days, SSDs can go up to multiple millions of random reads per second. TLS termination (or self-inflicted software bottlenecks) will become an issue much earlier.
24.
by gpugreg 19d ago
I also thought about building one of those AI honeypots, but I stopped when I realized that it would quickly be turned into a command and control server by botnet operators, followed by mail from a three letter agency. Is there any way to a
25.
by gpugreg 21d ago
I scrolled through https://simonwillison.net/tags/pelican-riding-a-bicycle/ but did not see any image where the spokes were correct. For a moment, I thought that the text-to-image model might have gotten it right,
26.
by gpugreg 22d ago
MiMo-V2.5-Pro-UltraSpeed gets pretty close with over 1000 TPS on 8x B200. It has 1.02T total parameters and 42B active, compared to 27B total/active for Qwen3.8-27B. Also, B300 are out now. I think 1500 TPS for Qwen3.8-27B should be do
27.
by gpugreg 23d ago
Cached tokens count towards the limit as well. For example, if your context window is 50,000 tokens, it takes 9 requests to reach that limit without generating a single token.
28.
by gpugreg 23d ago
I was wondering whether this was any good for programming, but it is too fast for its own good. There is a limit of 450,000 tokens per minute. I hit this limit in about 90 seconds and burned through $1.10 while doing so. This is because cac
29.
by gpugreg 23d ago
Thanks, that is very kind of you. I have submitted bugs in the past and was enthusiastic for GPU support for over 15 years now, but I lost my faith.
30.
by gpugreg 23d ago
You can query polars data frames with SQL: https://docs.pola.rs/api/python/stable/reference/expressions... Unfortunately, polars does not support parameterized queries, so the risk of SQL injection is ex

More ›