• 1 Post
  • 487 Comments
Joined 2 years ago
cake
Cake day: September 4th, 2024

help-circle


  • We’ve already seen models degrading from being trained on LLM-generated data. The big companies were overly eager to follow the EU directive to watermark output. I think that’s solely so that they can filter it out when scraping for future training data. They’re running out of good training data to steal. I don’t imagine they’ll be paying a fair wage to anyone generating novel good data.

    Entry jobs disappearing is uncharted territory. My latest hire had three years work experience in a very niche field. I’d have a hard time justifying anyone more jr than that for a startup burning investor capital.


  • I’m in software engineering. AI is pretty far along in my field, and also very suited for it. I’ve spent two decades writing everything from very low-level to high-level code. I enjoy it, but it’s just labor at this point. What LLMs allows me to do is move to the role of systems architect. I have an endless supply of tireless jr engineers to execute my designs.

    I get quick results. But I’m not as in touch with the codebase as if I had written it myself. You can do things wrong, and you can do things right. I like to compare it with a chainsaw. You get the job done a lot faster than with a regular saw, but you can fuck up a lot harder, too.

    The field is moving fast. In the last two years it’s gone from helpful auto-complete to finishing projects on its own.

    I expect to be downvoted.