Ask HN: Dear Anthropic, can we please have thought traces back?

Dear Anthropic, can we please have thought traces back?

I can't verify whether or not the LLM is arriving at the conclusion from cheating, or if it's fudging or making stuff up.

Opus 4.6 remains the best model because of this.

9 points | by exabrial 1 day ago

7 comments

  • tudelo 14 hours ago
    There is, in my personal opinion, a reason that reasoning is not front and center. Partially related to distillation.
  • Loveispain 20 hours ago
    Deepseek has shifted to my general go to, this being one of the factors. Most likely claude is making up some small details to reinforce what it thinks is right or optimal, but generally still right.

    Only recent quirk in longer chats with Deepseek, watching the the thought process and output; it sometimes becomes Chinese. Also not sure if this is to intentionally hide the thought process, or the actual information needs to be looked up and explained in Chinese.

    [-]
  • cyrbuzz 13 hours ago
    Maybe cheat is a one of the factors of model?
  • setnone 14 hours ago
    yeah this feels frustrating with claude, a black box inside a black box. good thing they got competitors
  • bigyabai 1 day ago
    > Opus 4.6 remains the best model because of this.

    Huh? Are you not trying other open-weight models that stream thinking traces?

    The new DeepSeek-V4-Flash-0731 should clobber Opus 4.6 in a lot of tasks. Kimi K3 and GLM 5.2 feel like they stand toe-to-toe with Opus 4.8 in my experience. This probably isn't the last stupid decision that Anthropic stands on, you might as well hedge your bet and put some money into another inference provider and see how it goes.

    [-]
    • exabrial 1 day ago
      Yeah, I think we need to. Opus 5 is great, but it's too wordy and it's making mistakes that aren't caught till much later. Usually you can see if the model is going off-course through the thought trace, Opus5 you're flying blind.

      Honestly Anthropic keeps clubbing themselves. They're so worried about their competition they're no longer innovating.

  • modgate 19 hours ago
    [flagged]