singhrac

18 ms·

1. 211.▲
   by singhrac 3y ago

   My startup works on optimizing the electrical grid in areas with increasing renewable penetration. We're hoping our secondary effect of de-scheduling coal gen by forecasting renewables will reduce global emissions by 0.005%, something
2. 212.▲
   by singhrac 3y ago

   Are we reading the same Github repo? All of those acronyms are in the second bullet.
3. 213.▲
   by singhrac 3y ago

   I'm using Zola and Cloudflare Pages, and it is currently possibly to use ZOLA_VERSION=0.17.1. I thought I had opted into the Pages v2 beta (I remember joining a Discord for this purpose a while ago), but on the actual page the setting
4. 214.▲
   by singhrac 3y ago

   I don't know about the complexity of your extension, but the PEP provides per-container locks ("This PEP proposes using per-object locks to provide many of the same protections that the GIL provides. For example, every list, dicti
5. 215.▲
   by singhrac 3y ago

   Yeah, you're right - even though CUDA is async, doing any preprocessing (in Python) can be harder if you don't have shared memory (the start-up latency hit of multiprocessing is not a problem in this context). I've only ever
6. 216.▲
   by singhrac 3y ago

   I think it’s the other way around? I haven’t run into any ML code that could be multithreaded that wasn’t written in C++, but have often run into server tasks that could use a polling thread, etc. All the ML code is written in lower level l
7. 217.▲
   by singhrac 3y ago

   I can’t imagine you’ve read the proposal with a comment like this. The interpreter is already patched (twice in the proposal, for two different versions of Python), and Sam Gross has personally already patched many commonly used Python libr
8. 218.▲
   by singhrac 3y ago

   This is disappointing to hear but very useful, thanks. I was considering buying a discrete Intel card for tinkering. Can I ask if the FPGA use case was ML-related? Feel free to email in profile.
9. 219.▲
   by singhrac 3y ago

   ZLUDA is a library written by a hobbyist. The comparison you are looking for is oneAPI. They don't have the messaging down, but Intel is good at software.
10. 220.▲
    by singhrac 3y ago

    https://rachitsingh.com/ Trying to write more!
11. 221.▲
    by singhrac 3y ago

    It's worth mentioning that there are embeddings inside the model (e.g. what vector does each token correspond to / nn.Embedding) and the "embedding" of an input, which are used in the same context. They might not even li
12. 222.▲
    by singhrac 3y ago

    Maybe it’s because MathJax loading slow would delay rendering, and using a CDN-hosted MathJax increases the chance the content is cached? That’s usually why people use CDNs. It’s more important for MathJax than, say, interactive scripts, si
13. 223.▲
    by singhrac 3y ago

    I’d argue you don’t need CUDA if PyTorch runs well on your platform, which is what everyone is trying to show with these MLPerf results. I think Intel’s strategy with oneAPI is not bad, it’s just late.
14. 224.▲
    by singhrac 3y ago

    Intel is much closer with Habana Gaudi2. Sometimes I’m unsure whether they even recognize it. They also published MLPerf results today: https://habana.ai/blog/gaudi2-demonstrates-competitive-llm-p... With 384 Gaudi2s t
15. 225.▲
    by singhrac 3y ago

    I mean, it was clearly a better language model than the at-the-time best model (which, if I recall correctly, was Yang et al's mixture-of-softmaxes). But the biggest surprise over time has been how big of a sponge large transformer mod
16. 226.▲
    by singhrac 3y ago

    Absolutely. I knew this project before the inclusion of AI and thought higher of it. I was about to compare it to py-spy and the “use ML to make code suggestions” feature is a turnoff.
17. 227.▲
    by singhrac 3y ago

    Yes. I worked on convex optimization in trading (deeply involved with fine-tuning, constraint selection, utility function construction, etc.).
18. 228.▲
    by singhrac 3y ago

    I’ve heard good things about https://build-system.fman.io/ , though I haven’t used.
19. 229.▲
    by singhrac 3y ago

    I manage a Jupyter instance where we do ~75% of all of our work, and before that used Jupyter daily as a quant. In response to your comments + sibling's: Sharing: there hasn't been a good solution, so you have to hack something to
20. 230.▲
    by singhrac 3y ago

    They did it to take market share, and they have, slightly (6% market share against AMDs 9%). If you look closely their software story is miles ahead of AMDs, so I can see them coming to a comfortable second in the next few years… which is n
21. 231.▲
    by singhrac 3y ago

    I suspect we might have to agree to disagree here, but the T450 and 2015 MBP make very different tradeoffs. I might be looking at the wrong model, but I see a 1600x900 screen (while my MBP had a 2560x1600 screen), and (admittedly high quali
22. 232.▲
    by singhrac 3y ago

    I really can’t understand this argument. Apple has the least planned obsolescence of any major manufacturer available. The only upgradeable laptop you can buy is the Framework, which is more expensive than a baseline MBP. It also didn’t exi
23. 233.▲
    by singhrac 3y ago

    The one killer feature that I wish Obsidian had is CRDT support, Apple Notes-style. I realize it's a huge engineering lift and they are (appropriately) focused on making the UX perfect, but it would be really great for collaboration. I
24. 234.▲
    by singhrac 3y ago

    I think the statement "I'd be much happier with a Studio" is a little hypothetical? Sorry if that's not true, but everywhere I've looked, it seems like these are not ML training chips, and people are just hoping the
25. 235.▲
    by singhrac 3y ago

    I think this is what George Hotz is doing with tiny corp, but I have to admit I have little hope. Making asynchronous SIMD code fast is very difficult as a base point, let alone without internal view of decisions like “why does this cause a
26. 236.▲
    by singhrac 3y ago

    Sorry, I have to point out: Tensorflow is not comparable to CUDA. Tensorflow is a (arguably) high level library that links against CUDA to run on NVIDIA GPUs, as does PyTorch (the main competitor). Comparatively few people have “deep” exper
27. 237.▲
    by singhrac 3y ago

    Yeah, also see AMD engineers working on Triton support here: https://github.com/openai/triton/issues/46
28. 238.▲
    by singhrac 3y ago

    I didn’t have a chance to edit (HN was down) but I meant to mention that this is a fairly nice piece with technical details about some very nice features added to the language.
29. 239.▲
    by singhrac 3y ago

    I definitely agree (and I think given what you work on you would be horrified by how I define cached functions that capture locals), but I think in practice getting to a state where you can restart your kernel often makes it easier to reaso
30. 240.▲
    by singhrac 3y ago

    They should really not publish this on Medium, honestly. Dart/Flutter already have a fairly corporate feel to them, and being on Medium (which, sadly, is very heavyweight and full of self-promotion these days) really makes me feel like