Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. A 1996 dial-up chat room simulator (lolchat.rip)(lolchat.rip)
    discuss
  2. The ruble markup on AI tokens(infertrail.com)
    discuss
  3. Temporal Straightening for Latent Planning(arxiv.org)
    discuss
  4. Sovereign Substrate Appliance: Layer-1 Edge Node on ARM64(richardkugel.substack.com)
    discuss
  5. Anyone using a Wispr Flow alternative that is non-cloud?
    discuss
  6. MacPad: Touch Mac in iPad(github.com/dcmmc)
    1comments
  7. Ant Group releases finance-focused Ling-3.0-flash-Fin(artificialanalysis.ai)
    discuss
  8. New method to minimize immunosuppression after organ transplantation(nature.com)
    1comments
  9. The 'AI Safety' Movement Is Making AI Less Safe(reason.com)
    1comments
  10. Show HN: Livenerf – a benchmark for whether Opus 5.5 gets nerfed(github.com/ninjahawk)
    discuss
  11. Adding side drawer with workspace concepts to rio (rust terminal) like cmux has(github.com/raphamorim)
    discuss
  12. The Jagged Frontier of Jev 1.13(typesafe.ai)
    discuss
  13. Mercury Is Shrinking, and Faster Than We Thought(nytimes.com)
    2comments
  14. Swarm Scaling(tobyord.com)
    discuss
  15. A WIP Lego Sorter: Maybe a million pounds is possible [video](youtube.com)
    discuss
  16. Artificial Intelligence in Research(tmodrzyk.github.io)
    discuss
  17. CNN Mood 'Like a Funeral' as Paramount Acquires Warner Bros(variety.com)
    7comments
  18. Show HN: Palette, a Figma plugin that turns one color into a whole UI system(figma.com)
    discuss
  19. AUGUSTOSFACES(augusto.at)
    discuss
  20. What is the World's Loveliest Language(economist.com)
    discuss
  21. Xounds Redux and FruitMenu Redux(totallynotmalicious.com)
    1comments
  22. Design your programming languages right (2024)(veritates.love)
    discuss
  23. U.S. Government seizes 54-pound Martian meteorite sold at auction last year(nytimes.com)
    3comments
  24. A Taxonomy for Navigating the Global Landscape of AI Regulation (2025)(acm.org)
    discuss
  25. The 38 most influential Bluesky posters(bsky38.com)
    discuss
  26. UN General Assembly: Trump renames AI to Super Intelligence(washingtonpost.com)
    4comments
  27. The Mac Mini M6 is the fastest way to.. emulate a Windows 98 Pentium II machine(nyaa.sh)
    discuss
  28. Show HN: A fast, offline Markdown viewer for macOS built in Swift(github.com/temir-dev)
    discuss
  29. About Impersonation Risk Detection(support.apple.com)
    discuss
  30. Trump orders AI rebrand as "super intelligence"(axios.com)
    3comments

Your computer is more powerful than you think

5 pointsby 1d agofsf.org
3 comments
1d agoHN ↗

Aye.

Before AI good good enough for people to argue it was a "scam" because it was only taking other people's jobs and not their own*, people were arguing computers were a scam because Shakuntala Devi could find the 23rd root of (*checks notes*):

   916,748,676,920,039,158,098,660,927,585,380,162,483,106,680,144,308,622,407,126,516,427,934,657,040,867,096,593,279,205,767,480,806,790,022,783,016,354,924,852,380,335,745,316,935,111,903,596,577,547,340,075,681,688,305,620,821,016,129,132,845,564,805,780, 158,806,771

in just 50 seconds, while the supercomputer she was up against took over a minute.

Of course, this happened in 1977 with a computer that was already obsolete at the time, but stories have a way of sticking.

My 2013 MacBook Air could train a small neural network from scratch on the MNIST database to read handwritten digits, then read all those digits, and then also calculate the 23rd root of that number, all in the (literal, I timed it) blink of an eye.

* why yes, this is a recent thing I've experienced here

1d agoHN ↗

Pretty awesome. I fed it to Astra Max and it took over 2 minutes ( https://chatgpt.com/share/6ab1bd98-7974-83e8-9734-8e10b5b3ef... ) so that's interesting in itself. If we knew how she did that, I wonder if the insight would be transferable to transformer-based ANNs.

Interesting character even aside from her mental-arithmetic abilities.

Edit: Deepseek 4.1 Flash running locally got it in 134 seconds. Beaten by Astra but not by much. Neither model seems to recognize the problem (at least it doesn't appear in DS4.1's chain of thought.)

4h agoHN ↗

I saw a video at one point about how she probably did it, but I can't find it. I feel it was Numberphile or Computerphile or StandUpMaths, something in that general circle.

5.6 Luna on a free account took 56 seconds, and has some stuff I remember from that video: https://chatgpt.com/s/t_6ab2d40e145c81918a3abb288b4b99a8

(The extra stuff at the header is due to my custom prompt).

I suspect Astra Max took so long because it's the big model and doing the most thinking.