News for hackers. Stuff that matters.
Re:Google edits Super Bowl ad for AI that featured false inf... by crazygringo on Sunday February 09, 2025 @08:38AM (#42990597)
That's quite a leap there from "sometimes wrong" to "basically useless"...
Re:Google edits Super Bowl ad for AI that featured false inf... by rdtsc on Sunday February 09, 2025 @10:40AM (#42991307)
> quite a leap there

A healthy hop I would say, keeping in mind that this was showing off the best side of it in a commercial. So it just did a Google search, got a wrong factual result, the LLM couldn’t verify it was bogus, regurgitated it as is, and the executive piped in with “acshually, it’s just a plain old search result” somehow not realizing that just makes them look even worse.

Re:Google edits Super Bowl ad for AI that featured false inf... by WhyNotHugo on Sunday February 09, 2025 @09:11AM (#42990799)
If a model often produces information that is blatantly wrong, then you need to check ALL of its outputs. If you're going to have to double-check all information that it provides, you might as well skip using it entirely and search for the information directly.
Re:Google edits Super Bowl ad for AI that featured false inf... by crazygringo on Sunday February 09, 2025 @09:59AM (#42991058)
You're missing the part where searching for the information directly might take hours or even days.

You're looking for black-and-white truth, while the real world is actually more interested in efficiency.

I can spend 2 days scouring documentation and forum posts and experimenting to get ffmpeg or Matplotlib to produce exactly the results I want. Or I can just ask ChatGPT and check if its code works, and if it doesn't maybe spend 10 minutes correcting it so it does, or refining the prompt so it does.

And so you're also missing the part where verifying the correctness of output is very often many orders of magnitude faster than coming up with the output in the first place.

Re:Google edits Super Bowl ad for AI that featured false inf... by nolist_policy on Sunday February 09, 2025 @09:49AM (#42991003)
LLMs produce wrong results. News at 11.