Chicago Sun-Times caught publishing likely ChatGPT generated list of Summer Reading recommendations. 10 of 15 books don't exist.
Wait, so chat GPT can't even compile a fucking list of books without making up 2/3rds of it's response out of thin air?
I don't really see the appeal of using AI if it's going to take more time and effort to fact check the responses it gives me because it has a massively high failure rate.
Now you're getting it
That's because despite what AI companies keep trying to ram down people's throats, it's not built to compile facts
I don’t really see the appeal of using AI if it’s going to take more time and effort to fact check the responses it gives me because it has a massively high failure rate.
You just run the output back through and ask it to fact check for you. Problem solved!
My company paid for some people to go to one of these "accelerate your company with AI" seminars - the recommendation that the "AI Expert" gave was to ask the LLM to include a percentage of how confident it was in its answers. I'm technical enough to understand that that isn't how LLMs work, but it was pretty scary how people thought that was a reasonable, sensible idea.
Yep, it's sold as "artificial intelligence" not "large language models" on purpose. They want you to think that it's intelligent and actually putting thought into its outout, rather than just outputting the most likely thing based on the input. It isn't intelligent in the slightest. It's just a fancy algorithm.
To be fair, I think it's really easy to fall into that sort of viewpoint. The way most people interact with them is inherently anthropomorphic, and I think that plus the fact that AI as a concept is almost as memed as flying cars in various media makes it really hard not to end up relating that way.
I have a technical background and understand LLMs enough to know that's bad, but I also used it like LCARS when it was new and thought it was effing amazing for a time. It's super easy to fall under that spell, IMO.
Treating it anthropomorphically is a sign of respect, similar to how a sailor would bond with their ship. It's not necessarily BAD or dumber or wrong to talk with it like its human - that's clearly what every single interface is telling you to do by representing it like a texting partner. You can't interact with a machine that speaks english non-anthropomorphically.
I don't disagree! But my point was that it will inherently present challenges to interacting with it objectively and fully rationally, IMO.
Totally - "scary" as in "this is going to cause so many issues and get people into real trouble" more than "man people are stupid"
LLM with a memory now: Yes, these books all exist and are highly recommended. I hear the Chicago Sun-Times is considering putting all of them on their summer reading list.
"Writer": (stopped reading at the word exist) print it!
I’m a newspaper editor. The people who are/were most excited about this tech, also happened to be the folks who did none of the actual writing to begin with.
We had sales folks gleefully hand us texts for advertisers that they’d ‘written’ with ChatGPT. Those texts contained so much wrong info, it wasn’t even funny. It made up things, had wrong information about websites, contact info, that sort of thing.
But since the sales monkeys weren’t actual writers, they didn’t catch on to that. Meanwhile, we were spending more time fact checking and unfucking their texts than if we’d written it ourselves in the first place.
It CAN be helpful to shorten or rearrange already written things, but if you ask it to write from scratch, it’s usually not going to be good.
I don't see how it fucked this up so badly. One of the few things I use AI for is book recommendations, and I have yet to be recommended a non existent book.
I treat LLM responses like I do random internet advice. Trust, but verify. Pretty light on the trust part lol.
I treat it more as distrust but verify. Sometimes it's right, but it has proven enough times to make shit up that it doesn't get my trust by default. Sometimes it can lead to me searching for the right thing though, so it is sometimes remotely useful. I rarely use it though and run it locally.
I always interpreted "trust but verify" to be an interpersonal thing, so I don't see a problem interpreting it as "distrust and verify" with machines.
Trust but verify as a concept is irrelevant to the majority of people. It specifically refers to how intel orgs' staff should handle their long term sources for information. It is applicable specifically when they have a high degree of trustworthiness already, but you still need to be a bit more sure than that.
If that's not your situation, you have no use for it.
You wouldn't take tips from a off-road rally driver during city traffic, would you?
I don't get the metaphor. I take tips from anyone, but I don't blindly execute them. That is to say, "but verify".
I think the colloquial usage of the phrase has differed from its original meaning. I've never heard it in the context you're referring to.
I consider LLMs to be "bullshit generators"
If the situation calls for only BS, an LLM is great for it. Anything else, not so much.