How do we tell it isn’t doing this in the background when working on some larger project? In some way invisible to the user until it’s too late? People want to use these to make medical and business decisions and I just don’t see how that doesn’t end in disaster sometimes.

Right, my point being is how can it be trusted to not just invent or distort any facts in the background when asked to perform certain tasks? This is an answer easily found in 4 seconds via google or a search of Frinkiac, but it confidently provides the wrong answer twice.

and even then - you’d need a human to review to ensure it didn’t invent or leave out something important! /thread

What incorrect assumptions are the LLM’s making in the background, invisible to the user? How can anybody tell if they aren’t already an expert in the field they are using it to work on? It’s dangerous to use this stuff for anything beyond building spreadsheets or calendars,

My point is basically this: if you can’t trust the AI to not make incorrect assumptions and conclusions with something as simple as a water cooler conversation about a popular TV show, how on earth can it be trusted with medical or business decisions?

ALSO WRONG. It admits it’s wrong once I simply ask it to provide context as it prompted me to do. And then proceeds to confidently give another wrong answer!

The AI confidently tells me a season and episode and says this joke appears in this episode. It does not. At no point does it even express any doubt that this is the right answer. And then when you ask it to provide context for the joke, you get:

Here’s a perfect example

3/X

Everybody knows the Simpsons. Most of us can recall a good joke from an episode and probably narrow it down to a specific season. The Simpsons is one of the most well documented shows on the internet and maybe in the history of TV. AND YET, these AI’s confidently fail at simple tasks like this 2/X

This is ALSO the wrong episode.

And look at this! When I ask it to provide the setup for this joke, as it prompted me to do, here’s what you get:

For example, see this from right this minute. This scene is 100% not from this episode. This isn’t the first time I’ve asked this question and gotten a confidently wrong answer. How do you trust this to do actual important things if it can’t handle simple confidently be wrong like this?

How can I trust it to make important decisions if it can’t handle simple water cooler questions?

I find it fails even with simple tasks sometimes. I am often testing it by having it locate the season and episode that a specific Simpson’s joke appears in. One of the most well documented shows on the internet. And it often answers confidently wrong.

I think when one of these LLM’s are working well, they can be very helpful and save a lot of time. But I think so much of what they can do accurately is extremely limited. I do not use LLM’s that often, but when I do I rarely get the result I’m looking for after repeated attempts.

Video of the ICE-kidnapping victim under a sheet. You will also hear the officer say he saw them moving. I am inclined to believe the person was alive.

Good summary of the issue here:

nymag.com/intelligence...