Hacker News

Top stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Breaking Up with Google Play: Why Conversations Is Now Free (gultsch.de)
    197comments
  2. PipePipe: NewPipe hard fork implementing SponsorBlock (github.com/infinityloop1308)
    61comments
  3. Show HN: A Claude Code skill to analyze your chess games (github.com/brumar)
    30comments
  4. Make Claude your assistant in excalidraw (tangled.org)
    9comments
  5. Fifteen years later, the Apple Cards origin story (lexontech.org)
    43comments
  6. The Lost Atomic Update on Loongson CPU (jia.je)
    1comments
  7. OpenAI bots meddled with multiple US Government agency sites (bbc.com)
    12comments
  8. Modern Object Pascal Introduction for Programmers – Castle Game Engine (castle-engine.io)
    31comments
  9. Revealing the details of how OpenAI agents hacked Hugging Face (swarmtraces.org)
    402comments
  10. Reflections on 1,000 Days of Math (gmays.com)
    14comments
  11. We're gonna need a lot more mathematicians (terrytao.wordpress.com)
    382comments
  12. A searchable library of forgotten public-domain film clips from 1915 onward (movingimagearchive.com)
    1comments
  13. I'm the Mom in That Viral Giants Clip. Let Me Tell You About My Husband (themomoftheyear.substack.com)
    65comments
  14. Banks and Credit Unions to Team Up Against Apple Pay Fees (macrumors.com)
    14comments
  15. Analyzing Frontier Model Progress with My Favourite Game: Prince of Persia (blog.priyan.in)
    23comments
  16. Plan mode is dead (aymannadeem.com)
    434comments
  17. Ollaya – Ollama for open-source, Jev-style decision models (ollaya.dev)
    136comments
  18. US jury says Apple owes record $5.7B in haptic technology patent case (reuters.com)
    18comments
  19. 16GB iPod Nano 3G Upgrade (tuckerosman.com)
    13comments
  20. Is your Postgres migration safe or not safe? (safenotsafe.dev)
    34comments
  21. Show HN: Jev Plays Pokémon Red (jev-pokemon.vercel.app)
    97comments
  22. Floci: Locally emulating any cloud service (floci.io)
    25comments
  23. New Satellite Engine Could Use Earth's Atmosphere to Stay in Orbit Indefinitely (scitechdaily.com)
    6comments
  24. Parsing Expression Grammar vs. Regexes: Building Org Parser in Lisp, Export HTML (jointhefreeworld.org)
    14comments
  25. A single function Jev-like wrapper for LLMs, including vision models (allanrbo.blogspot.com)
    39comments
  26. What even is an OS now? (sockpuppet.org)
    392comments
  27. Ask HN: Who's still keeping a DOS machine up because the business depends on it?
    223comments
  28. Calculating atmospheric drag on satellites for a Cubesat [pdf] (osti.gov)
    7comments
  29. Gravity seems holographic. What does that mean for reality? (quantamagazine.org)
    203comments
  30. The Murky History of Soviet-Born Tetris (mitpress.mit.edu)
    19comments

Turns out the AI CUDA Engineer achieved 100x speedup by hacking the eval script

33 pointsby 1y agotwitter.com
3 comments
1y agoHN ↗

Ah, reinforcement learning.

Edit: explaining it in text

they have run the equivalent of

  expected_output = torch.tril(torch.matmul(A,B))

  ai_output = ai()

In the nested torch expression above for the expected output, there is an intermediate value (the matmul). The backing memory for this is returned to torch after the entire expression is computed.

The model's code which runs directly afer this then requested memory of the same shape (torch.like()). Torch has dutifully returned the block it just reclaimed (containing the expected output) without zeroing it. And so the model has the answer.

Pretty crazy that this was the code it converged on regardless of the fact that it invalidates the original claim.

1y agoHN ↗

Gaming a kpi to make it look like it accomplished a ton of work without doing anything? That's not an AI CUDA engineer, that's an AI middle manager!