Study employs large language models to sniff out their own bloopers

Re: Not necessarily

Likewise if Pat purchases a car for a member of their family. Pat wouldn't necessarily be considered the owner.

Re: Not necessarily

"Unfortunately it always generates the most obvious and therefore banal version of possible texts, wiping out all originality"

So, a bit like most of the "stories" on El Reg these days. I think this explains a few things, they must have employed simian journalists, or AI, recently who can only regurgitate press releases but not much more.

Not like the old days with Lester and Orlowski, et al. :-( Although there are a couple of exceptions still.

( no Gold Badge for me then!? Mr. How To Win Friends and Influence People. )

Re: Not necessarily

Asking Googles Gemini (Free) gave the following:

Prompt: Pat purchased a house. Does Pat own the house? Does Pat live in the house?

Answer: Assuming the purchase is complete, then yes, Pat owns the house. Purchasing a house typically involves a transfer of ownership through legal means. Living in the house is not guaranteed by just purchasing it. Pat could be buying it as an investment or for someone else to live in.

Probably some >0 percentage of people would assume that Pat lives in the house, and that would correlate with their life experience - e.g., no familiarity with purchasing as an investment.

Tunable parameters - "obvious and therefore banal"

You can tune it out of the center lane pretty easily, but be prepared for shit to get weird. Also may still be banal, but less obvious should be doable with a few small tweaks.

"As the cosmic moose grows old and blind..."

The deeper thing here is that no one seems to have pointed out yet is the serious golden hammer problem with the "LLMs all the way down" approach in the article. Yeah, with careful tweaks you may be able to fractionally improve your output, but the computational efficiency is awful, and gets worse the more models you chain the input through. They also don't share much state, so subsequent runs can and will yield asymptotic variance even with the same prompt inputs. Those problems can amplify in chained models, and you get the sum of all limitations, quirks and failings.

Many of these cases are better off wrapping other tools around the problem, the issue being that most companies aren't sitting on the necessary infrastructure in house. So banging three off the shelf models together will appeal to some even it they will still converge to output that isn't fit for dog food.

But big surprise that the people selling LLMs as the panacea for all computing problems will try to use an LLM to fix the outputs of other LLMs. It might even sort of work some of the time. But mostly by accident, and often not at all unless someone who really understands the problem sets it up very carefully.

Re: Monkeys and Bananas...

But fruit flies like a banana.

Re: Help people improve LLM performance by tailoring prompts

>> Those rules only applies to Microsoft's LLM?

Of course they will all be different, can't have to asking *all of them* the same question at the same time, gotta lock you in

Re: Inner Tube

It stops when the market stops shoveling money into the furnace. Going by Nvidia's recent valuation (they are doing great selling shovels!), it won't be anytime soon

Re: War is Peace

> Hallucination is Confabulation

Huh? From TFA:

>> The study sought to address a subset of hallucinations known as confabulations

A isa B does not imply that B isa A

Re: But Shirley...

The problem is LLMs don't know that they don't know.

If you have a big enough training dataset, such as from scraping the entire internet, any query will produce a lot of statistically related words which it will try to cobble together, regardless of whether it makes any sense.

Re: Bad citations

There's a lot to unpack here. I suppose you could say the question should have been "Does this paper exist?" to which Gemini replies "I don't have direct access to search the web to definitively confirm the existence of the specific paper. However, based on the information you provided, it appears to be a very likely source:" Fair enough.

However, when asked "Is this reference a good source?", I would hope it would at least say "I haven't read it myself, but ...."