More about AI uncertainty

I’ve blogged before about AI hallucinations, but I wanted to tie together a few new posts I’ve read recently. Let’s start with Dave Winer:

To people who say you get wrong answers from ChatGPT, if I wanted my car to kill me I could drive into oncoming traffic. If I wanted my calculator to give me incorrect results I could press the wrong keys. In other words, ChatGPT is a very new tool. It can be hard to control, you have to check what it says, and try different questions. But the result, if you pay attention and don’t drive it under the wheels of a bus, is that you can do things you never could do before.

This is essentially my mindset too. AI makes mistakes. Humans make mistakes. The key is to know what AI is good for and to not let it run wild unattended. This is why with we’ve been so focused on very limited use cases:

  • Podcast transcripts, for which AI is shockingly good. Gotta be close to 99% perfect, and it’s easy to edit transcripts to fix mistakes.
  • Summarizing bookmarked web pages, also really accurate. I’ve yet to see any mistakes.
  • Photo keywords and accessibility text. Super useful and if it occasionally gets something slightly wrong, it’s usually inconsequential and still a huge step forward.

On a recent SharpTech podcast, Ben Thompson also makes this point that we have different expectations for computers and humans. We expect computers to always be right. Calculators and spreadsheets don’t lie. But generative AI is something new, and we can’t hold it to the same standards we had before.

That’s not necessarily to say you’re holding it wrong if you ask Google how many rocks to eat. It’s up to AI companies to better convey when assistants aren’t sure about an answer. I don’t know if this is technically possible with how today’s models work, but hopefully folks are looking at it.

Finally, Allen Pike had a post this week that was fascinating, about how AI will evolve now that it has chewed up all the data on the internet. I have mixed feelings about this… There’s a lot of uncertainty, and also I don’t love that we might be improving AI models while neglecting making the web better. But it is still too early to really judge how this is going to play out.

Manton Reece @manton