ukuina

16 ms·

1. 331.▲
   by ukuina 2y ago

   Agreed, the dehyping of article titles is one of the main reasons I built hackyournews.com, and the avoidance of clickbait via proactive summarization is consistently rewarding.
2. 332.▲
   by ukuina 2y ago

   Apple must be paying OpenAI a pretty penny to give the best LLM tokens away for free.
3. 333.▲
   by ukuina 2y ago

   Are there any text editors that use LSPs for non-coding tasks?
4. 334.▲
   Llama-zip: a command-line utility for lossless text compression(github.com)

   3 points by ukuina 2y ago|2 comments
5. 335.▲
   by ukuina 2y ago

   I recommend slowing down time while a button is being held to preview where the insertion will happen upon button release.
6. 336.▲
   by ukuina 2y ago

   This is the right answer. When it's time to make a product, you either use a managed VectorDB or ditch it for a more traditional datastore with careful indexing.
7. 337.▲
   by ukuina 2y ago

   QuickWit also sends telemetry by default: https://quickwit.io/docs/telemetry
8. 338.▲
   by ukuina 2y ago

   Here's a version that doesn't need a server. Just save as an HTML file and open it in a browser (either localhost or on a fileserver somewhere). <!DOCTYPE html> <html lang="en"> <head>
9. 339.▲
   by ukuina 2y ago

   Why does this even need a server?
10. 340.▲
    by ukuina 2y ago

    Update: Product name is Windows Recall.
11. 341.▲
    by ukuina 2y ago

    How is this different in effect from teacher resources like Classroom.com? Many websites provide teacher packs; this offering is a hyper-customized, always up-to-date version of the same.
12. 342.▲
    by ukuina 2y ago

    Llama3-8B is adequate for non-technical summarization or simple categorization.
13. 343.▲
    by ukuina 2y ago

    Sounds like a good RAG use-case unless all 1k documents need to be comprehended simultaneously. Look at H2O.ai: https://github.com/h2oai/h2ogpt
14. 344.▲
    by ukuina 2y ago

    Yeah, this is an INSTALL.md masquerading as a cookbook.
15. 345.▲
    by ukuina 2y ago

    Interestingly, I am moving more towards plain text notes because they are easier to ingest for LLMs.
16. 346.▲
    by ukuina 2y ago

    I'm grateful this is a simple blog post rather than a 20-page arXiv paper with dozens of meaningless graphs. Or worse, a 20-deep Twitter thread.
17. 347.▲
    by ukuina 2y ago

    FAIR really swung for the fences with Llama3. It's a very impressive model, but the 8K context size is quite limiting for most use-cases.
18. 348.▲
    by ukuina 2y ago

    So many high-performing, yet poorly-named OpenAI models in that list.
19. 349.▲
    by ukuina 2y ago

    OpenAI desperately needs a marketing consult. "GPT4o"? Seriously? Even "GPT4 Omni" is easier in conversation, and that's what the "o" stands for! They severely underestimate the number of casual users they
20. 350.▲
    by ukuina 2y ago

    Spotlight search can be slowed down by the number of apps exposing their content to the search index.
21. 351.▲
    by ukuina 2y ago

    "Apologies for the oversight. You're correct in that... <pointless repetition of your instructions, followed by incorrect result>"
22. 352.▲
    by ukuina 2y ago

    They tried that with Stage Manager and it's a mess.
23. 353.▲
    by ukuina 2y ago

    Yes, this is Windows AI Explorer.
24. 354.▲
    by ukuina 2y ago

    > To what extent this may be driven by the new project shaped in a copilot "friendly" way? Not OP, but this is the vast majority of the improvement I've seen.
25. 355.▲
    by ukuina 2y ago

    > Most of my time is spent thinking about the domain space, constraints and processing flows. This is the ideal place to add the most value, since a generic LLM will be weakest here.
26. 356.▲
    by ukuina 2y ago

    Remember when they claimed Yi had 200k context length despite it having 16k of usable context? I remember, because I spent non-trivial effort trying to make it work for long-form technical summarization. My lackluster findings were validate
27. 357.▲
    by ukuina 2y ago

    Full title is: "Fishing for Magikarp: Automatically Detecting Under-trained Tokens in Large Language Models"
28. 358.▲
    by ukuina 2y ago

    What an amazing look into one of the greatest minds in programming! Thank you for this treasure. The relevant bits: > I am now allowing the client to guess at the results of the users movement until the authoritative response from the se
29. 359.▲
    by ukuina 2y ago

    I would suggest going all-in on SQLite and even encrypting it on disk: https://sqlite.org/com/see.html An encrypted DB will be a differentiator compared to every other offline editor (Joplin et al). It would also appea
30. 360.▲
    by ukuina 2y ago

    > Palm devices were a pleasure to use. RIP Graffiti.