Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Hill investigates after F-35 parts diverted to Hong Kong(politico.com)
    discuss
  2. Resend · Email for Developers(resend.com)
    1comments
  3. White House launches Trump TV following press bans(thehill.com)
    discuss
  4. Anthropic Principle(wikipedia.org)
    discuss
  5. Et Tu, Brute? Economic Misalignment in Personal AI Agents(arxiv.org)
    discuss
  6. I said no and Apple said yes(dbushell.com)
    discuss
  7. Robot bunnies deployed in Florida to fight invasive pythons(popsci.com)
    1comments
  8. Police hugging AI so tight, soon we'll all be pre-crime suspects(responsiblestatecraft.org)
    discuss
  9. The Sun Doesn't Shine on Me (2006)(fullhoffman.com)
    discuss
  10. KeiroLabs – Web research infrastructure for AI agents(keirolabs.cloud)
    1comments
  11. EncryptedLLM: Privacy-Preserving Large Language Model Inference(icml.cc)
    discuss
  12. The agents never stop. A council can(flurdy.com)
    1comments
  13. How to cut back on social media addiction(shuaibkhan.com)
    discuss
  14. Show HN: Z8Log – Structured logging your AI coding agent can query(z8log.com)
    discuss
  15. Trusted directories for listing your Mac app(reddit.com)
    discuss
  16. RippleNav – A web instrument provides real-time visual feedback on thought focus(ripplenav.org)
    discuss
  17. Saudi's Ceer unveils first EVs(reuters.com)
    discuss
  18. Welcome Silona Bonewald, New MetaBrainz Foundation Executive Director(metabrainz.org)
    1comments
  19. Show HN: Easytris – Tetris that wants you to win(medv.io)
    discuss
  20. A Tribute to Donald Knuth (2024)(paramrathour.github.io)
    discuss
  21. Create AI Image Free – No Sign-Up Needed(createaiimage.net)
    discuss
  22. Scanner says the HTML attachment is clean. What happens when you open it?(mohitkhare.com)
    discuss
  23. Chewbacca's Ancestor Is Real(maxhawtdawg.com)
    discuss
  24. The Set of Math-Sterpieces(paramrathour.github.io)
    discuss
  25. Reverse-Engineering Jane Street's ASIC(mooofin.github.io)
    discuss
  26. How people are testing AI agents?
    discuss
  27. Frontier Overhangs(stratechery.com)
    discuss
  28. Some Old Expectations for the New Apple(taoofmac.com)
    discuss
  29. Physics Problem Finder(beneke7.github.io)
    discuss
  30. The Armored Sarcophagus Saving Ukrainian Soldiers(theatlantic.com)
    discuss

Ask HN: Is Astra's writing starting to bother anyone else?

2 pointsby 1h ago
2 comments
As someone who pretty much only uses GPT models (due to how generous the limits are), I always felt like I’d dodged a bullet when people complained about Claude and its load-bearing Claude-isms. GPT models, up to 5.6 Sol, certainly had their own AI-isms, but they always wrote in a way that was easy for me to understand well.

But then Astra came in. I feel like I have to focus ten times as hard to actually understand anything it’s trying to say. The way it writes is just so… dense, as if it’s trying to pack as much information in as few tokens as possible, at the expense of readability.

I wonder if this is due to token-saving optimizations that both OpenAI and Anthropic are doing? Or if it’s just a natural consequence of models getting “smarter”?

Either way, I think I’m switching back to Sol.

1h agoHN ↗

You're absolutely right, and I should have caught that.

1h agoHN ↗

You use models raw as they come?

You can tell the models how to write. Describe style, writing level, intended audience, and response format.