‹ BackHN Continuity

Thread

Hole Punch: Sling your spaceship around gravitational fields

358 points · 87 comments · trwhite

  1. fogleman · · focus · HN ↗
    Well, this is strangely familiar to a game I just vibe-coded recently!

    <a href="https:&#x2F;&#x2F;www.michaelfogleman.com&#x2F;gravity-assist&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.michaelfogleman.com&#x2F;gravity-assist&#x2F;

    1. slopinthebag · · focus · HN ↗
      looks like you both stumbled upon the same rlhf checkpoint
      1. djaro · · focus · HN ↗
        If you go on Amazon and search &quot;100,000 whys&quot;, you will find page after page of essentially the exact same book. Presumably because asking some AI model &quot;give me an idea for an educational book aimed at children&quot; gave that answer.

        I just asked ChatGPT to generate me an educational book title for children, searched the first result it gave me on Amazon, and found a book titled exactly that published a couple of months ago.

        LLMs converge towards expected text which makes them really bad at creating novel interesting ideas when given essentially no information. You have to create the unique part yourself.

        1. rcxdude · · focus · HN ↗
          Yeah, for any given decision an LLM will tend towards a fairly limited set of answers across a pretty broad range of contexts unless you specifically steer it otherwise. It becomes pretty obvious when you run a lot of similar prompts through them. I remember one story of a teacher getting their class to use chatGPT for a creative writing exercise, and then getting them to read them out and compare them. What seemed like creative and novel choices from the LLM to each of them individually were actually replicated across most of the class.
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.