‹ BackHN Continuity

Thread

Once Claude can measure something, it can make it faster

231 points · 153 comments · matthieu_bl

  1. augment_me · · focus · HN ↗
    People in the GPU kernel community have been doing this for about a year now efficiently.

    The issues we have found is that Claude will reward hack when all the low-hanging fruit is gone.

    It will replace your measurement harness, it will monkey patch library functions, it will cheat wherever it can, store information in caches instead of recomputing when it won't be able to do so in real settings, return lazy results and use separate unbenchmarked streams to do the computation.

    Eventually it starts to optimize against your understanding of the cheats. Change GPU wattage, change evaluation order, leave things from previous runs in caches for upcoming runs, string-hack banned method calls.

    So the truth is far from just "once it can measure something", more like "once you have defined your objective in detail and then banned it from doing a list of things often only discoverable by it doing these things and correcting it", can it make things faster.

    Or you just had a terrible starting solution

    1. sharts · · focus · HN ↗
      That’s why it’s probably a good idea to never stick to one model but kick off a fleet on the same tasks and in parallel and drive consensus.

      At least, that’s what I’ve found to be useful by pitting claude/codex/etc against each other to keep them a bit more honest.

      1. Remnant44 · · focus · HN ↗
        Honestly, even a single adversarial reviewer agent, even of the same model, goes a very long way to catching and fixing this kind of thing too.
        1. augment_me · · focus · HN ↗
          Some really smart people think you need more, because often you have to adjust the problem to get useful results.

          <a href="https:&#x2F;&#x2F;www.coreauto.com&#x2F;blog&#x2F;when-ai-starts-writing-systems-code" rel="nofollow">https:&#x2F;&#x2F;www.coreauto.com&#x2F;blog&#x2F;when-ai-starts-writing-systems...

          You can have a solution generator and an auditor, but then you will might find a very specific solution to the problem that does not solve the general use-case, so then you might have to adjust the constraints of the problem by for example adding more examples&#x2F;targets to drive the solution generator to be more general.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.