Meta AI - can we really trust AI???

[Alex321]

It may well be "learning." As I understand it AI is far from intelligent but simply scrapes around the web for information and presents a summary of the content to the user. This isn't "intelligence," any of us could do this but more slowly.

While it probably shouldn't really be described as intelligence, it is nowhere near as simple as you suggest.

I was simply bored, I wasn't looking to trick the AI. The issue for Meta is my view AI cannot be trusted is confirmed and will take a lot to change. No one at Meta will fact check Rovers. This is more Facebook misinformation generated by the company itself.

It will get more accurate over time. It is improving by leaps and bounds, and while it certainly doesn't get it right all the time, it is getting better at sifting the false information out.

But certainly you should double check anything it returns that matters to you.

 

[markemark]

I’m not convinced by the great advancements. A couple of years ago it leapt forward every month. Since then less so. There’s virtually no big leaps in the last year.

As for getting smarter, I don’t think that’s true either. The data it uses is often wrong, often manipulated or often ai generated. It’s not getting smarter.

What’s happening is the businesses are incorporating it. But not as much as the hype suggested. Nowhere near as much.

It’ll end up being a tool people use like computers or the internet.
Some job types will go. New ones will appear. As for the majority. You won’t lose your job to ai. But you will lose your job to someone who knows how to use ai.

 

[IanSmithCSE]

Good morning,

Google's Ai recently tole me that I probably didn't mean to be searching for what I had entered as search criteria.

Apart from being sort of funny it is also a worrying sign that the AI is "just" a text pattern recognition program rather than something that "understands" what it is reading,

I know that there are other search engines but by the time you have added a few adverts, paid for links and of course Google's AI, real search results, the real useful results, the ones that Google made its name on are getting harder and harder to find.

Although Google's AI summary is clearly identified, how it affects the results isn't, okay Google has always kept its search methodology private so in one way this is just anothrt upgrade.

I am too far out of loop on AI to know what what is being touted as AI actually is, at the moment is seems like a fairly specialised form of text pattern recognition without any form of education or thought. What I mean is that you can pass to a text pattern recognition program a deeply respected and industry wide accepted statement that an 18650 Lithium Ion battery has a maximum capcity of say 3,500mWh but does it understand that this is the truith and then query the advert from DodgyBatteriesAreUs that claim 10,000mWh for £1.99?

Given that misinformation way outweighs the truth on the web how does the document reader know how to weight it's source?

For me, over the last few years searching the web in general has show a massive increase in sites that use autmatically generated content being returned as best matches and as these are returned first they become "The Source Of True Information".

Bye

Ian

 

[sungod]

in 'ai' systems based on large language models, input is encoded an passed through various stages of processing to produce an output

the models are trained at the cost of vast energy demand and blatant pillaging of original creative works, probably also draining investment from more beneficial areas

at no point does the system 'understand', the output is essentially probabilistic, it produces it's best match for the input tokens, which may or may not be correct

more recent 'reasoning' models do not actually reason in human terms, but they can break things down into smaller steps to reach the final result, rather than the operator needing to do it manually

but in spite of their huge compute capacity and ocean-boiling energy consumption, they still get soundly thrashed by a decades-old atari chess game

in specific domains, these models can be trained on high quality data and provide useful results, but the general ones are increasingly vulnerable to ingesting their own slop as it proliferates across the internet, with content owners blocking or poisoning them in the absence of licence agreements

can you 'trust' ai? no, it can certainly be useful, but it's no substitute for understanding what you're doing and being able to assess and cross-check/correct as necessary

 

[Regular.Cyclist]

We recently asked Copilot ‘AI‘ to write me a draft Medical Devices Policy and draft Terms of Reference for a Medical Devices Committee.

I was impressed with what it produced and, with an hour or so working on it, had it ready to go to the board for approval. I know for previous experience it would take many times longer, running into weeks or months, to write one from scratch.

It has it‘s place but must be used with caution.