dionb wrote on 2025-03-18, 13:35:

So should you use AI?

The most dangerous aspect is when LLMs boldly deliver a blatantly wrong answer with the confidence of 10 gurus combined, and this answer is used and passed as the truth for that point on, yes people actually do this.

My take on the current models basically agrees with yours. Sure use them, but always double check and see if you end up with something useful. Some times validating defies the purpose, but there are cases where you might save time and effort.

Never blindly trust AI. It's even less credible than Wikipedia. But if you can "tame it", you can profit from it: solving math problems, finding bugs, generating images, and other things like that.
For web search related things Perplexity is one of the best.

I'm worried about the ability to think outside the box of future generations. They'll likely not question AI and accept its answers as gospel truth, with all that it entails.

Also, when will everyone on the internet become chatbots?

One good thing about Vogons is that there's no point system that some AI can hijack. So far, It feels real. I heard that on reddit that's currently an issue.

You can tell that AI is not self-aware because it has no epistemic humility. Never does it say "I'm not sure... I think it might be X, Y, or Z... let me do some research"

It just blurts out whatever is the "most likely answer" based on 10,000 Reddit posts, which is probably wrong.

I was looking for examples of the first dx9 games the other day, and Google’s AI replied with:

It clearly got confused and pulled one of the first directx games.

it's interesting that you found co-pilot most accurate, from a little experience i found that one seems 'better' at all kinds of tech questions, not sure if its luck or that its training has been reinforced or supervised a bit more smartly perhaps

in any case that was an interesting set of questions as the know how is relatively sparse online - it would need to be in forums like this, remnants of old articles, scanned books and magazine (assuming ai 'reads' them) - and even then in comparison with many topics the actual details wont be widely repeated.

Given that, AI would either tend to repeat specific sections of found text or generate answers that veer away from verifiable facts as its trail goes cold (in regard to the very specific answers searched for)

ask about something very well known, like the details of a famous motor and i'd guess the answers are better just because there is more text to "learn" from. Not sure though

tauro wrote on 2025-03-18, 16:28:

I'm worried about the ability to think outside the box of future generations. They'll likely not question AI and accept its answers as gospel truth, with all that it entails.

Also, when will everyone on the internet become chatbots?

I think that's already true with "I googled it" and with the circular and/or poor citations behind some wiki articles. People tend to believe it as if it has an inherent authority like a news program. People also tend to believe recounted experiences and reasoning given in person too, within their own social network. AI will just eclipse search results by combining everything in a seemingly authoritative "voice"

In fact AI is likely to become the interactive tool in place of searching or "googling" something, and with more content online being LLM generated there is a risk that AI will train itself on itself, as it were. There is also a risk that, if people themselves stop generating content for AI to learn from, others may maliciously create "content" totally unseen except for AI, or otherwise tamper with AI, such that results skew in a particular political direction, or simply misinform so widely that actual truths become buried under a lie repeated so often it is perceived as true

taken in extreme the contents online as of the early 2020's may be the final stage in human knowledge (well, it won't - but you can see that if nothing new is generated because everyone relies on AI then what is there to learn from anew?)

gerry wrote on 2025-03-18, 20:59:

it's interesting that you found co-pilot most accurate, from a little experience i found that one seems 'better' at all kinds of tech questions, not sure if its luck or that its training has been reinforced or supervised a bit more smartly perhaps

Not sure exactly why, but it clearly uses better sources for this stuff - and quotes what they were. Kind of ironic given how useless Bing usually seems for the same thing.

in any case that was an interesting set of questions as the know how is relatively sparse online - it would need to be in forums like this, remnants of old articles, scanned books and magazine (assuming ai 'reads' them) - and even then in comparison with many topics the actual details wont be widely repeated.

That's why I chose this example - multiple layers of complexity, info is sparse but very much in existence (and I knew it well enough myself to judge the responses).

Given that, AI would either tend to repeat specific sections of found text or generate answers that veer away from verifiable facts as its trail goes cold (in regard to the very specific answers searched for)

ask about something very well known, like the details of a famous motor and i'd guess the answers are better just because there is more text to "learn" from. Not sure though

It would be better, but even there, it's all about guessing the next word, not being accurate as such.

tauro wrote on 2025-03-18, 16:28:

I'm worried about the ability to think outside the box of future generations. They'll likely not question AI and accept its answers as gospel truth, with all that it entails.

Also, when will everyone on the internet become chatbots?

I wouldn't worry too much about that. Looking at my own children, they are very much aware how much of what AI offers is completely made up - more so than many adults I know of. Its's the older generation, the people who make decisions (and vote) that I worry about.

taken in extreme the contents online as of the early 2020's may be the final stage in human knowledge (well, it won't - but you can see that if nothing new is generated because everyone relies on AI then what is there to learn from anew?)

Don't underestimate humans.Yes, Einstein's apocryphal quote about common sense not being common is very much true, but on the other hand the very limitations of AI ensure that for the forseeable future us old-fashioned wet-brains who still know how many fingers a hand should have, how text should look and what the arrow of causation looks like will still be in the driving seat, not some computers basing their content on ours in the past.

Great job ! I just tried a few those AI's to see what the results were on a Elan Musk quote from an interview 2 months ago. Copilot was the closest, the rest were so far off it was stupid 🤣

added:

dionb wrote on 2025-03-19, 00:29:

Looking at my own children, they are very much aware how much of what AI offers is completely made up - more so than many adults I know of. Its's the older generation, the people who make decisions (and vote) that I worry about.

Agree the vast majority of older adults may soon think AI answers as "gospel" and cannot be wrong, only because they won't mentally challenge what some computer tells them but would if someone they do not know saying same thing 🙁

Reminds me of this - https://www.youtube.com/watch?v=yI8u_WSFfMs