HN Explorer

killerstorm

↗ HN profile [ 39.7 ms ] full profile
Karma
0
Created
()
Submissions
0
  1. A connection between capability-based security and lambda calculus (killerstorm.github.io)
    2 points 1
  2. Show HN: Skill capsules" for LLMs, a "poor man's continual learning" (github.com)

    "Continual learning" is considered one of the "blockers" for LLMs: they can't learn on the job, don't improve over time, etc. In particular, Dwarkesh Patel describes it as a number of problem which has to be solved to…

    1 points 3
  3. METR's time-horizon of coding tasks does not mean what you think it means (killerstorm.github.io)
    1 points 1
  4. The future of LLMs: cognitive core and cartridges? (killerstorm.github.io)
    2 points 0
  5. Ask HN: Why isn't capability-based security more common?

    Recent ["self-propagating NPM malware"](https://news.ycombinator.com/item?id=45260741) reminds us that the predominant security model is basically whack-a-mole: you gotta trust _every_ piece of software you run…

    12 points 13
  6. Automated ML research, AI drives the process from idea to report (killerstorm.github.io)

    LLMs are now sufficiently smart to do at least some basic ML research tasks autonomously (going from idea to report). And quite likely by the end of the year they might be capable enough to reproduce ML papers with some…

    3 points 0
  7. Show HN: Run automated ML experiments using Claude Code (github.com)

    I made a template which can be used to conduct (basic) ML experiments in a fully automated mode: Claude Code will write the code, you only need to provide a working environment and the idea. The goal was largely to…

    1 points 1
  8. Show HN: I vibe-coded some unusual transformer models (github.com)

    Goals: * demonstrate that LLMs are smart enough to conduct ML experiments pretty much on their own * specifically, vibe-coding isn't just for web stuff * encourage people to conduct these small experiments * in…

    1 points 1
  9. Can LLMs represent meaning? / Getting over the impenetrable wall (killerstorm.github.io)
    2 points 2
  10. Show HN: A Better LLM UI in Emacs (x.com)

    All LLM user interfaces I've seen so far are somewhat frustrating: * ChatGPT web requires a lot of copy-paste, it rewrites whole document even if you need to update a part of it, etc. * Github Copilot completions are…

    8 points 0
  11. Rebuild Software Initiative (rebuild.software)
    3 points 1
  12. Show HN: I built a GPT-4 bot which builds software incrementally

    The ability to synthesize a relatively short snipped of code was already demonstrated. But I thought it would be interesting to test whether GPT-4 can replace a programmer completely. To do that, AI needs to plan its…

    5 points 4
  13. A note to my former self: You're not supposed to take care of everything (blog.killerstorm.dev)
    3 points 1
  14. Ask HN: Are there immediate practical uses of “Bitcoin 2.0”?

    There is a plenty of so-called "Bitcoin 2.0" projects which try to build something on top of Bitcoin, or, maybe, just inspired by Bitcoin. But many of them seem to be either outright impractical (i.e. trying to…

    3 points 0
  15. Colored coins paint sophisticated future for Bitcoin (coindesk.com)
    1 points 0
  16. Please review: DoReserve.com, an appointment booking system

    We have just created/launched this site. The idea is that small/medium businesses can outsource online booking system to this service. We have noticed that a lot of small companies such as beauty salons and restaurants…

    5 points 7