vvbee wrote on 2025-12-16, 15:54:

What do you mean they have no way of checking? How is not even one way available to AI? I'd assume you generally need three things: access to the content, knowledge of the topic, and the ability to reason fairly. I don't think AI's particularly missing any of these, least of all if the baseline is a person who hasn't much interest in verifying anything.

I posted an example earlier, of someone posting an LLM list of "OpenGL games with MIDI". The LLM had listed games that were released BEFORE ogl was even available. From the 3 you mentioned, it doesn't have the second and third. LLM isn't AI (certainly not the kind that people expect it to be). It just knows which words, images, content, etc. are related and how they are ordered. It has no understanding of the content itself.

vvbee wrote on 2025-12-16, 10:09:

Do you mean that in the true sense.

More like a literal sense. They search or ask a question, they get a response, they assume that response is correct. They trust it to be correct. They rarely bother to verify or question the response. It reminds me of the phrase "I seen it on the internet, so it must be true". The subject matter in question is often of a non-technical nature in relation to my co-workers and friends. They are curious about something, they search it, and trust what the bot says is correct.

Honestly, it's kinda of a weird mix. Many of the younger people seem to trust AI responses much more than the older ones, but many of the older ones distrust the technology, but some still assume the info it spits out to be correct. Though more of the older generation "trust, but verify" the information than the younger ones do. Often by running the answer by other people or digging a bit deeper online when time permits.

For more technical information, like what I use, I've used both ChatGPT and Grok, and both have been amazingly accurate and huge time savers for me, I've learned a lot. I've recently been using Grok more as it seems to give me slightly better results to match my requests when I'm working on a piece of code for a project.

eddman wrote on 2025-12-16, 16:11:

vvbee wrote on 2025-12-16, 15:54:

What do you mean they have no way of checking? How is not even one way available to AI? I'd assume you generally need three things: access to the content, knowledge of the topic, and the ability to reason fairly. I don't think AI's particularly missing any of these, least of all if the baseline is a person who hasn't much interest in verifying anything.

I posted an example earlier, of someone posting an LLM list of "OpenGL games with MIDI". The LLM had listed games that were released BEFORE ogl was even available. From the 3 you mentioned, it doesn't have the second and third. LLM isn't AI (certainly not the kind that people expect it to be). It just knows which words, images, content, etc. are related and how they are ordered. It has no understanding of the content itself.

I tested a couple models and they didn't list pre-GL games. What's a more general test that would convince you that AI has knowledge and can reason? Or is there a test that would convince you?

StriderTR wrote on 2025-12-16, 18:00:

vvbee wrote on 2025-12-16, 10:09:

Do you mean that in the true sense.

More like a literal sense. They search or ask a question, they get a response, they assume that response is correct. They trust it to be correct. They rarely bother to verify or question the response. It reminds me of the phrase "I seen it on the internet, so it must be true". The subject matter in question is often of a non-technical nature in relation to my co-workers and friends. They are curious about something, they search it, and trust what the bot says is correct.

Doesn't sound too bad, it's throwaway knowledge.

vvbee wrote on 2025-12-16, 19:30:

What's a more general test that would convince you that AI has knowledge and can reason? Or is there a test that would convince you?

There is no test that can prove it; a few correct results doesn't invalidate the incorrect ones. Just read up on LLMs. There is a reason the term AGI was invented, because of how the term "AI" got corrupted by equating LLMs to our former understanding of the word.

LLM is like a person that memorized an entire language's shape and sound, every single word, every single possible combination of sentences, which word combinations are a response to other combinations, but yet doesn't understand the meaning of any of them. It's basically advanced mimicry.

It's not the perfect example but is the best one I can come up with right now.

EDIT: or perhaps this. I don't know german but memorize the german pronunciation of all the works of a german writer; I can recite the books from start to finish when asked, or a given line in a page, and yet I have no idea what any of that means.

In this case I didn't limit it to LLMs specifically nor to the AI of today. If you don't have a test then it's just a belief. They're fine to have but you can't do anything with them.

Of course in this case we were talking about practical uses, so the test of letting AI act and then evaluating the results is valid. Earlier in the thread it picked out cognitive biases from posts, now it was able to detect fake OpenGL + MIDI games in a list. So in that sense we can say it can verify the veracity of content. Which then means either it has the three things or whatever's required that I mentioned before, or that to verify content AI only needs the one thing, access to the content. Which is what I was saying earlier, either AI reasoned about the posts to find their biases or the posts weren't actually expressing anything but were just statements of biases.

vvbee wrote on 2025-12-16, 23:30:

In this case I didn't limit it to LLMs specifically nor to the AI of today. If you don't have a test then it's just a belief. They're fine to have but you can't do anything with them.

AFAIK the public ones you interact with are all LLMs. The proper definition of AI is having intelligence (or the one that was used up until a few years ago), and LLMs don't have it. It's not a matter of "belief" as we already know how LLMs work.

Again, if it had intelligence it wouldn't have made the mistakes in the first place. This reasoning you're talking about does not equate intelligence, unless you're using a much simpler definition. These are algorithms that operate upon the relations in the data. If it presents the wrong result, then is challenged by the user, the algorithm could input the challenge as new data and modify the result. If enough people challenge it in a different direction, it could modify the result again. As I've stated before, it has no understanding of the content itself. Being right or wrong is provided by outside factors, and not from an inherent ability of an LLM to "think" (perhaps there's a better word for that).

I'm sure one day there will be proper AI, but what we have now is not it.

We are seemingly talking about different things, or it's just my english getting in the way as it isn't my first. In any case, there isn't really anything else to add to our conversation.

EDIT: To clarify, I'm not saying they are useless, just that we should know their nature and how to properly use them.

eddman wrote on 2025-12-16, 20:28:

LLM is like a person that memorized an entire language's shape and sound, every single word, every single possible combination of sentences, which word combinations are a response to other combinations, but yet doesn't understand the meaning of any of them. It's basically advanced mimicry.

It's Markov Chains. The only innovation this time is the unprecedented computing power being thrown at it and the vast resources being consumed to power that number crunching. Fundamentally it's a calculator if it ate the Sun instead of running off a dinky solar cell. Someone on Slashdot put it very well:

Computing used to be measured in the number of computations per second, the storage available or the number of bits in and out of a data centre. In other words, its output. Now it seems to be measured in the power consumed, its inputs. Is this because the output is so pointless it's not worth measuring?

eddman wrote on Yesterday, 00:12:

vvbee wrote on 2025-12-16, 23:30:

In this case I didn't limit it to LLMs specifically nor to the AI of today. If you don't have a test then it's just a belief. They're fine to have but you can't do anything with them.

The proper definition of AI is having intelligence (or the one that was used up until a few years ago), and LLMs don't have it. It's not a matter of "belief" as we already know how LLMs work.

Again, if it had intelligence it wouldn't have made the mistakes in the first place. This reasoning you're talking about does not equate intelligence, unless you're using a much simpler definition. These are algorithms that operate upon the relations in the data. If it presents the wrong result, then is challenged by the user, the algorithm could input the challenge as new data and modify the result. If enough people challenge it in a different direction, it could modify the result again. As I've stated before, it has no understanding of the content itself. Being right or wrong is provided by outside factors, and not from an inherent ability of an LLM to "think" (perhaps there's a better word for that).

The mistakes you're talking about were according to you made by one model but then not by the others tested. But since you defined intelligence in a way that excludes those who's kind has been known to do this or that then this version of intelligence is defined by group identifiers anyway. Even though it's derived by comparison to humans it's not even recommended to apply this thinking to humans, so I don't see how you'd make it a useful metric here.

You say the output (mistakes) disqualify, but that even if it didn't the internal state (algorithms) would. I'm not aware of a realistic measure of intelligence that's derived from the internal state, maybe each person can have their private measure this way but you can't generalize it at all so this approach can't be useful here. That just leaves the output, which we can evaluate and compare between AI and humans.

It's valid to say mistakes in the output suggest less intelligence, but if you have a hard threshold then your approach has no explanatory power past the threshold. So then it's not useful to select a threshold that already excluded the category that was going to be evaluated.

However you choose to see it, it's fine. Read up on the inner workings of LLMs and you'll understand it better.

eddman wrote on Yesterday, 13:33:

However you choose to see it, it's fine. Read up on the inner workings of LLMs and you'll understand it better.

AI's take on what would've been a more effective way to argue the position:

Person 2’s main argument was: “LLMs have no way of checking the veracity of the content.” With a reasoning model connected to th […]
Show full quote

Person 2’s main argument was: “LLMs have no way of checking the veracity of the content.” With a reasoning model connected to the internet, this is functionally false. ... If Person 2 insists "it still doesn't understand," Person 1 will rightly view this as a distinction without a difference. If the AI catches the OpenGL error by looking up the dates, it has vetted the content.

Since Person 2 can no longer argue that the AI can't verify facts (because it can via search), they must shift their argument to Source Bias and Algorithmic Amplification. To convince Person 1, Person 2 would have to say:

"Okay, the model can look up dates and verify facts better than a lazy human. But 'vetting' isn't just about fact-checking dates; it's about evaluating truth.

If a reasoning model searches the internet, it is still limited to what is popular or SEO-optimized on the internet. If the internet is flooded with a common misconception, the AI will 'verify' that misconception as fact because 10 out of 10 search results say so.

A human might have existing knowledge or intuition to say 'that sounds wrong.' The AI will see 10 matching search results and call it 'Verified Truth.' Therefore, this tool doesn't fix the problem of misinformation; it automates the consensus of the internet, which is often wrong."

It's a decent strategy, better than the alternative. I'd still say I'm not limiting what I'm saying to just the AI of today, but in principle arguing in favor of what it can already do and will do even better as time goes on. Though I also don't see reason to immediately assume that AI can't even now have the "knowledge or intuition" to say the "verified truth" sounds wrong, but what it may well not have is the initiative (or budget) to trigger more iterations of evaluation.

That's exactly what a language model would churn out. I know what's behind the curtain. Whatever definition of "intelligence" you want to go with, that's perfectly fine; it simply does not equate to the proper definition that I've known, which would be what nowadays is called AGI, which hasn't been achieved yet.

On the previous page it was pointed out that incorporating AI increased the amount of critical thinking potential in the discussion, and I'd say that's again true here. This also suggests there's a threshold, likely different for each community, past which it's simply more conductive to explore ideas with AI directly, and for this AI doesn't need to excel to the extent that the threshold isn't at the maximum of human ability. What effects this would have on the various communities I don't know, maybe they become stale holdouts due to population shift if you will.