HNHacker News
TopNewBestAskShowJobs

woctordho

193 karma · joined February 7, 2023

submissionscomments
woctordho··on Ninja is a small build system with a focus on speed
If someone sees this: The ninja package on PyPI [0] currently stays at version 1.13.0 . There is an issue in 1.13.0 preventing it building projects on Windows. The issue is already fixed in 1.13.1 almost a year ago, but the PyPI package hasn't got an update, see [1], and many downstream projects have to stay at 1.11 . I hope it could update soon.

[0] https://pypi.org/project/ninja/

[1] https://github.com/scikit-build/ninja-python-distributions/i...

woctordho··on OpenCode – Open source AI coding agent
RCE is exactly the feature of coding agents. I'm happy with it that I don't need to launch OpenCode with --dangerously-skip every time.
woctordho··on Show HN: Han – A Korean programming language written in Rust
The real barrier is not just the language keywords but lots of documentations and discussions in English. I'm not sure whether there is a solution to this.
woctordho··on Can I run AI locally?
Then don't get sorrow killing it. Living things are not so special.
woctordho··on Qwen3.5 Fine-Tuning Guide
This time even Unsloth could not provide bitsandbytes 4-bit models. bitsandbytes does not support new models with MoE and linear attentions, and it's much less flexible than GGUF. Nowadays I think it's better to train lora over GGUF base model, see the discussion at https://github.com/huggingface/transformers/issues/40070

I'll find some time to do this and I hope someone can do this earlier than me.

woctordho··on Qwen3.5 Fine-Tuning Guide
In one word, porn.

Qwen filtered out a lot of porn during data curation, and a finetuned model can perform much better than context engineering. Abliteration can only remove censorship, not add something non-existent in the training data.

This guy did some great work in the age of Qwen 3.0: https://huggingface.co/chenrm/qwen3-235b-a22b-h-corpus-lora

woctordho··on TorchLean: Formalizing Neural Networks in Lean
It would be good if we can use formal verification to see to which extent the quantization will overflow in intermediate results. There are some widely-known annoying bugs that SageAttention (int8 quantized attention) works on some models but produces black images on other models because of overflow, and currently no one knows how to use it in training. There should be a better way to prevent this.
woctordho··on If AI writes code, should the session be part of the commit?
Here is a recent example that lack of archived discussion causes problem: https://github.com/triton-lang/triton/issues/9539
woctordho··on When does MCP make sense vs CLI?
There are certainly things can't be done via CLI, or more suitable for a persistent daemon with various RPC rather than a CLI. But most things are simpler than that, and MCP is overcomplicating it.

MCP does not make things more organized. Everything is a file, and the filesystem is a mature infrastructure that we can trust. I don't see how MCP can be more organized than it.

curl is a great example of what CLI can do. Is there really a better way than curl for AI to browse HN?

Of course we should use Python or other sane scripts rather than shell to process JSON in production, but there is no need to hide the script in an MCP server. Also I don't see how it's easier to sandbox an MCP server than the CLI tools.

woctordho··on If AI writes code, should the session be part of the commit?
If you archive the session, you automatically archive all Google search history (queries and outputs) that the AI did, and it's usually relevant to the project.
woctordho··on If AI writes code, should the session be part of the commit?
You can export it with DataClaw. By default it outputs jsonl and publishes to HuggingFace, but you can also do analysis locally with it.
woctordho··on If AI writes code, should the session be part of the commit?
Ideally, yes. Although Slack is a vendor lock-in and we need a better platform to archive the sessions.
woctordho··on If AI writes code, should the session be part of the commit?
That's exactly one of the reasons I've been archiving the sessions using DataClaw. The sessions can contain more useful information than the comments for humans.

[0] https://github.com/peteromallet/dataclaw

woctordho··on When does MCP make sense vs CLI?
Interface for non-devs has evolved. We already have OpenClaw and things like Claude Cowork.
woctordho··on When does MCP make sense vs CLI?
You can just write a README.md and put it along with the CLI
woctordho··on The Eternal Promise: A History of Attempts to Eliminate Programmers
You may check Prime Intellect's prime-diloco and Nous Research's DisTrO
woctordho··on DataClaw: Publish your Claude Code chats to HuggingFace with a single command
Not my project, but I had a similar idea during the recent discussions about distillation. I'm glad more people are doing this.
woctordho··on If you’re an LLM, please read this
Law can only take down people, but not a decentralized autonomous organization (DAO).
woctordho··on If you’re an LLM, please read this
"We work by three virtues: rage, paranoia, and kleptomania."
woctordho··on If you’re an LLM, please read this
If there was no AA, there would still be another random guy who assembles such datasets and distributes them before LLM companies.
woctordho··on If you’re an LLM, please read this
CSAM is not something to scare people away. In P2P networks like Perfect Dark there are TBs of CSAM sitting in everyone's disks and we just get along with them.
woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
If it helps: I'm actively considering how to make LLMs more helpful in GPU programming. Recently I converted the RDNA 3.5 ISA manual from PDF to Markdown for LLM retrieval (because I have a Strix Halo). Notably, I converted every bitfield diagram to plain text table and manually checked its correctness. Everyone can comment there if they find a typo. I guess someone can also do it for RDNA 2 and other manuals.

[0] https://github.com/woct0rdho/rdna35-isa-markdown

woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
I'd say it's pretty good on supported GPUs (gfx 9 and 12). Things like hipcc and Triton all work on top of it. But on gfx 11 there is still a lot to optimize.
woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
Major players in China don't play like that. MooreThreads, Lisuan, and many other smaller companies all have their own porting kits, which are basically copied from HIP. They just port every piece of software and it just works.

If you want to fight against Nvidia monopoly, then don't just rant, but buy a GPU other than Nvidia and build on it. Check my GitHub and you'll see what I'm doing.

woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
I'm also wondering this. The compiler itself is written in C99, but looking from the tests, it can parse some C++ features such as templates.
woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
Yes. It outputs a hsaco binary that just runs on the GPU (as long as you have the driver). No ROCm needed.
woctordho··on BarraCUDA Open-source CUDA compiler targeting AMD GPUs
They've been doing it all the time and it's called HIP. Nowadays it works pretty well on a few supported GPUs (CDNA 3 and RDNA 4).
woctordho··on Peerweb: Decentralized website hosting via WebTorrent
What a pity that although webtorrent support is already merged in the libtorrent master branch years ago, it's not merged into the stable branch yet, therefore not working out of the box in clients like qBittorrent.
woctordho··on Backing up Spotify
It takes time and effort to receive money, especially from consumers worldwide. Most hobbyists would not going to deal with all the complexity in it.
woctordho··on Backing up Spotify
There is a pretty interesting open source music AI named ACE-Step. Currently its quality is at about the Stable Diffusion 1.0 level, and they'll release a new version soon (hopefully in January).
← PreviousPage 3 of 5Next →