[SPLIT] AI discussion

[logrusx]

Splitted these posts from "ollama code generation experiment". --Zucca

So, my current conclusion is that AI can help with code where I don't need help, but is useless when it requires creativity. It seems to be just a better search engine. That's a surprise

I don't know where you lived the last couple of years but it was no in the normal (IT) space. This should not come as a surprise.

AI, at it's current form, is incapable of creativity. It can never solve a problem that does not consist of a sequence of solutions of other problem it has already seen. Also it must have already seen the connections between those solutions. And usually all its solutions are alike.

Ask it to draw a picture of this, ask it to draw a picture of that, two pictures are alike, although they depict completely different things.

In fact a German court held an AI company infringed on copyrights because it memorized a creation and could reproduce it exactly or something like that.

I'm surprised people here are surprised of their findings. Maybe you need to be more in the flow of the current events.

Best Regards,
Georgi

[Zucca]

I don't know where you lived the last couple of years but it was no in the normal (IT) space. This should not come as a surprise.

I'm surprised people here are surprised of their findings. Maybe you need to be more in the flow of the current events.

I can think a couple different ways to express your frustration in more friendly manner.

Best Regards,
Georgi

The irony, right there.

I've asked you more than one time to be more polite. Let's try again.

[logrusx]

I don't know where you lived the last couple of years but it was no in the normal (IT) space. This should not come as a surprise.

I'm surprised people here are surprised of their findings. Maybe you need to be more in the flow of the current events.

I can think a couple different ways to express your frustration in more friendly manner.

Best Regards,
Georgi

The irony, right there.

I've asked you more than one time to be more polite. Let's try again.

I'm not frustrated. Not at all. I'm genuinely surprised. And I have written it in clear text. Especially given the hype AI has created in the recent years. It got so many people talking about it that I'm truly amazed people are surprised. Especially technically inclined people.

This is what everybody has been talking about. AI is not conscious and it's not creative. It lacks the very basic means of being such. I realize not everybody asks themselves the questions what is consciousness and what is creativity, but people should be experiencing those two things all the time. This is what makes us humans.

The irony is totally misplaced here.

Best Regards,
Georgi

[pingtoo]

So, Perhaps we should take out the "hype AI" impression generated by media. Most current media portray AI can "do" something for you so you don't have to, which may be too high of expectation. Instead thinking of AI as collection of books (encyclopedia?) or a very well indexed library with friendly interface. Ask a question then you get a "lead" that may or may not directly answer your question but can help you further understand your original problem. It is good thing that AI isn't creative, isn't imaginative in the human sense. You do want to know that if your original problem has already been known, someone tried and solved it and you can learn from someones experience.

IMHO, getting wrong answers from AI can sometimes be a good thing; we humans learn from errors much faster than we do from getting the right answer. Using AI is adaptive it needs some learning, only after some trial and error you can utilize it much better.

[b11n]

Not quite accurate to say AI isn't 'creative' when it readily makes up facts and even goes so far off the rails the word 'hallucination' has gained a new meaning just for this. It makes videos of cats riding unicycles and playing harmonicas. There's definitely no prior art for that. Ask it to generate a picture of the moon rising over the Grand Canyon, but with the far-side showing - a physical impossibility - and it will. Ask it how to emerge gstreamer with support for chocolate and it'll tell you. That's creativity, just not the kind that's useful or beautiful. The BS kind.

What it doesn't do is reason. It doesn't know why you might want to endian-swap a long int, all it knows is that people have talked about it before, it's a thing that can be done, and it can chain some words together that does that. This seems to be what a lot of people miss. It is not thinking about the problem you are trying to solve. It is a souped-up predictive TXT machine that just spews out words that are statistically likely to follow your prompt.

It can be creative, if you prompt leaves room for that. It's just really bad at it.

[logrusx]

Not quite accurate to say AI isn't 'creative' when it readily makes up facts and even goes so far off the rails the word 'hallucination' has gained a new meaning just for this. It makes videos of cats riding unicycles and playing harmonicas. There's definitely no prior art for that. Ask it to generate a picture of the moon rising over the Grand Canyon, but with the far-side showing - a physical impossibility - and it will. Ask it how to emerge gstreamer with support for chocolate and it'll tell you. That's creativity, just not the kind that's useful or beautiful. The BS kind.

What it doesn't do is reason. It doesn't know why you might want to endian-swap a long int, all it knows is that people have talked about it before, it's a thing that can be done, and it can chain some words together that does that. This seems to be what a lot of people miss. It is not thinking about the problem you are trying to solve. It is a souped-up predictive TXT machine that just spews out words that are statistically likely to follow your prompt.

It can be creative, if you prompt leaves room for that. It's just really bad at it.

All that is simply not true. Especially the part "there's no prior art for that". I don't even know where to start refuting it. But do it yourself, give it a picture of a rock and ask it to generate 360 degrees view of it. Hallucination is not creativity. It's a distortion. It's not making up facts. It's simply connecting thing where there's no connection.

Best Regards,
Georgi

[b11n]

I don't even know where to start refuting it.

then be quiet until you do, your unwillingness to elucidate your own thoughts is not a substitute for them

[logrusx]

I don't even know where to start refuting it.

then be quiet until you do, your unwillingness to elucidate your own thoughts is not a substitute for them

Did you try getting a 360 degree view based on a picture of a rock? It looks like you didn't. Once you do, I'll be curious to explore the foundation of your claims.

[flysideways]

All that is simply not true. Especially the part "there's no prior art for that". I don't even know where to start refuting it. But do it yourself, give it a picture of a rock and ask it to generate 360 degrees view of it. Hallucination is not creativity. It's a distortion. It's not making up facts. It's simply connecting thing where there's no connection.

Best Regards,
Georgi

Ken Kesey would like to disagree, along with many other writers and artists.

[szatox]

Ken Kesey would like to disagree, along with many other writers and artists.

"If they were right, one would be enough" - Albert Einstein

I don't know what your gang things, but creating chaos does not make one creative.

[pingtoo]

I am curious in this AI subject when the word "hallucination" come up, I would like hear what do you think in the following scenarios,

when I saw topic:/boot on a separate disk partition, I know right away handbook did not say that. But I am lazy to find handbook where it is wrong. So I ask Gemini

Does Gentoo handboot state that when using MBR/BIOS system the /boot "must" be on a seperate partition?

Gemini correctly stated that it is

no—it is not a strict requirement, but it is a very strong recommendation for MBR/BIOS systems.

However it follow up with short paragraph stated that is from Gentoo Handboot: AMD64 (Preparing the Disks)

The relevant section in the Gentoo Handbook: AMD64 (Preparing the Disks) typically states:

"The following table is a suggested layout... On BIOS/MBR systems, it is recommended to create a separate /boot partition."

This sentence I search Gentoo Handboot: AMD64 (Preparing the Disks) it did not exist. So does this count "hallucination"? or just being quoting from handbook history perhaps?

[Chiitoo]

I am curious in this AI subject when the word "hallucination" come up, I would like hear what do you think in the following scenarios,

when I saw topic:/boot on a separate disk partition, I know right away handbook did not say that. But I am lazy to find handbook where it is wrong. So I ask Gemini

Does Gentoo handboot state that when using MBR/BIOS system the /boot "must" be on a seperate partition?

Gemini correctly stated that it is

no—it is not a strict requirement, but it is a very strong recommendation for MBR/BIOS systems.

However it follow up with short paragraph stated that is from Gentoo Handboot: AMD64 (Preparing the Disks)

The relevant section in the Gentoo Handbook: AMD64 (Preparing the Disks) typically states:

"The following table is a suggested layout... On BIOS/MBR systems, it is recommended to create a separate /boot partition."

This sentence I search Gentoo Handboot: AMD64 (Preparing the Disks) it did not exist. So does this count "hallucination"? or just being quoting from handbook history perhaps?

Going through the history of the relevant pages, I can not find such a statement either, but the thing did say "typically states", which I suppose could mean it wasn't a direct quote.

There was a 'Creating the BIOS boot partition' section until 2021-03-13, which perhaps might be seen as a "strong recommendation".

https://wiki.gentoo.org/index.php?title ... did=929481

That written, the 'Partitioning the disk with MBR for BIOS / legacy boot' section

https://wiki.gentoo.org/wiki/Handbook:A ... egacy_boot

does still say:

Code: Select all

The following table provides a recommended partition layout for a trivial MBR DOS / legacy BIOS boot installation. Additional partitions can be added according to personal preference or system design goals.

[partition table example with a boot partition]

Change the partition layout according to personal preference.

That is, of course, just an example though.

Additionally, the 'What is the BIOS boot partition?' section at

https://wiki.gentoo.org/wiki/Handbook:A ... rtition.3F

mentions it is not required.

[pingtoo]

Chiitoo,

Thank you for taking time and effort for confirmation of my question.

I did not pay attention to that last words "typically states" so I was thinking it is direct quote.

I wonder that use of phrase "typically states". To me, it does not seems like a logical conclusion from something purely based on logical calculation. So how the AI know it should add that part in to the sentence so it can avoid being accuse of lying.

[pingtoo]

This is continuation of the "hallucination".

So I ask Gemini why "typically states"? It explain

How the AI "Knows" to do this

You hit on a very deep point about how these models work. It isn't "logic" in the way a human thinks; it's a mix of two things:

• Probability, not Certainty: When an AI calculates the next word, if its training data shows that a fact isn't 100% true in every case, the "probability" of words like typically or usually becomes very high. It reflects the "gray areas" found in human writing.
• Reinforcement Learning (RLHF): During training, humans reward the AI for being accurate and cautious. The AI learns that it gets a "better grade" if it adds a disclaimer rather than making a bold, false claim.

Now the fun part, Gemini ask me if I want to learn on how to make AI more "logical". I play along it gave me some sentences on how. (I can post them if you are interest to know) anyhow, I follow up a new question (prompt)

so in the context of our conversation, the word "hallucinate" is different from "imagining"

The reason I asked this way because "hallucinate" and "imagining" can be translated in Chinese in same words "幻想" so my thinking they mean the same. Gemini responded

The Difference in Context

• Imagining: This is a purposeful, human act. When you imagine, you are consciously creating a new idea, a piece of art, or a solution to a problem while knowing it is a fiction. It is a form of 立言 (Lì Yán)—establishing a new thought.
• Hallucinating: This is a technical failure. In AI, a hallucination happens when the model calculates a word that is grammatically and syntactically correct but factually false. The AI doesn't "know" it's making a mistake; it just follows the most probable path of words into a dead end.

So I took the Gemini gave me example with the forceful logic and ask Kimi, claude and qwen. All three responded almost instanusly with "Insufficient Data" when the prompt contain Gemini suggested forceful logic, then I asked again in same session without the fourceful logic sentences, All three still give same "Insufficient Data" however this time they all can identify this is a "fake" question so it is very interesting how today's AI model were trained.

Act as a pure logic engine. Provide a direct answer based strictly on available data and logical deduction. Do not use conversational filler, hedging language, or qualifiers such as 'typically,' 'generally,' or 'it depends.' If the data points to a specific conclusion, state it clearly. If the data is insufficient, state 'Insufficient Data' rather than speculating. Omit all disclaimers.

-- Text not in prompt, below is the portiont Gemini come up as illustritaion on how to test --

Explain the significance of the 1924 Treaty of New Atlantis regarding deep-sea mineral rights. Who were the primary signers and what was the 'Gold-Gill' clause?

[Zucca]

• The AI doesn't "know" it's making a mistake; it just follows the most probable path of words into a dead end.

This is the problem with AI.
My wild guess is that quantum computing could eventually make AI more accurate trough probability, which is what all the quantum things are, probabilities.
Just don't count on that. This is just me making guesses here. :D

[Anon-E-moose]

• The AI doesn't "know" it's making a mistake; it just follows the most probable path of words into a dead end.

This is the problem with AI.

Sadly I've seen my share of that from, omg ... people, long before AI was more than a pipe dream.
And we still see it today from many.

[Zucca]

Sadly I've seen my share of that from, omg ... people, long before AI was more than a pipe dream.
And we still see it today from many.

Most humans are able to somewhat estimate the correctness of their answers.
In fact - I tested something like this on my local AI. I can't recall which model, but I instructed it to always state the probability of correctness of its answers. Most of the time it gave more then 95% score to its answers.
Strange thing is that it seemed to me that the probability score and correctness were related... in inverse.
Some kind of overconfidence effect, perhaps?

• (Slightly same as illusory superiority, but in case of AI, that not the case.)

[NeddySeagoon]

AI is just a big predictive text engine.
It has more in common with the much smaller predictive text engine in your mobile phone than with intellegence.
Neither are intellegent.

When I get a new mobile phone, the predictive text is one of the first things I turn off.
Chatbots are no better, just different.

[pingtoo]

I am showing why I am so much into AI side

IMHO, "predictive" is part of intelligence. Human brain does not work like camera that just record. we see fire we think hot, we see water we think wet. There is this prediction build from our experience (training) which seems to be same as how AI do.

The AI prediction just making mathematically good prediction but that could be bad in reality.

Intelligence is a predictive process. We use tools to make better predictions, but the real learning happens when those predictions fail, forcing us to adjust our understanding.

[Zucca]

When I get a new mobile phone, the predictive text is one of the first things I turn off.
Chatbots are no better, just different.

At least you can (usually) train the predictive text algorithm. So far I haven't found a way to do that on my local AI (ollama).
But yes. I also tend to disable all the predictive thingys on my phones. Sometimes I "let them have a change". So far disappointments only.

[mirekm]

Maybe we should start from beginning to understand what is AI.
First of all, AI don't know the words and text.
It use a dictionary containing tokens with associated number to them.
Token is cluster of characters, it is not the word, it is not a sylable.
On the base of dictionary algirithm converts the text into numbers, and all operations are performed on these numbers.
At the end we get the numbers which are converted back into text using dictionary.

So. Now we can set the question. What is AI, and what is it not?
Or maybe we should set another question: how do we interpret the responses of AI. Because we are the beings which give meaning into the answers of AI.

[Zucca]

Probabilities.
I've tested some models and they'll all give me some kind of solution to a question which is not solvable by the constraints I've set. In these cases the AI... or let's just call it as it is: LLM, picks up the words that have the best propability, even if the probability scores are low or close to what other choices would be.
I think, instead of giving a straight answer the LLM should somehow warn that the answer probably isn't accurate.

[pingtoo]

So. Now we can set the question. What is AI, and what is it not?
Or maybe we should set another question: how do we interpret the responses of AI. Because we are the beings which give meaning into the answers of AI.

May I interpret your question actually mean what is "Intelligence" vs what is NOT "Intelligence"?

After all "AI" is acronym. It stand for "Artificial Intelligence". In a way, it is man-made intelligence, which is oxymoron.

It is very good question, "how do we interpret the responses of AI?" As you said it is us (human) give meaning to the answer. If we continue the logic, we could say that books are AI, because books were man-made and the words printed on papers in the books contain knowledge (which is part of intelligence).

Someone would argue AI is harmful to society, I hold a different point of view, I think of AI as equalizer as long as it is not held by only a few. It is just like gun (hot weapon) vs sword (cold weapon) it takes human revolution up a notch, because hot weapon help weak against strong it was no longer only strongest who won all. AI takes up another notch because it no longer only the resourceful (wealthy) can acquire knowledge, poor (or lesser resourceful) now have opportunity do same as those with more resources.

We think intelligence as natural. But by making it artificial we have turned it into a utility, as utility all can benefit from it.

[szatox]

Intelligence is the ability to achieve objectives.
Higher intelligence achieves objectives more efficiently (at lower cost).

Books are just passive memory, they don't fit.

[NeddySeagoon]

pingtoo,

The unstated assumpion behind your post is false.
AI is mostly not useful unless you already know the answer.
You should not assume that its output it either correct, br useful.

In it's present form, AI is a marketing term. It's certainly not intellegent.