The question is prompted by a post of @Sensei ’s which was sent to trash, precluding any discussion. LLMs are commonly assumed to be AI in the media and this seems to have become the perception of the general public, so I was intrigued to see @Sensei ‘s opinion is that they are not. Would anyone care to expand on this, viz. what qualifies as true AI and how LLMs should more properly be described?

Before this earlier thread devolved into some trollery, it seemed to address this question pretty well:

In another thread, which I can't find, I described LLMs as "stochastic parrots," a term I borrowed from Emily Bender, a computational linguist.

https://en.wikipedia.org/wiki/Stochastic_parrot

Definitely not AI. Possibly AS, though.

2 hours ago, Sensei said:

Let's ask what ChatGPT has to say about it.. ;)

Moderator Note

No, let’s not, if this is to be discussed in Computer Science, as per the rules.

This thread, and any thread outside of speculations, is for what actual people say

3 hours ago, exchemist said:

LLMs are commonly assumed to be AI in the media and this seems to have become the perception of the general public

I remember a company that called itself “Seattle’s Best Coffee” which was a marketing ploy, not a statement of fact

3 hours ago, TheVat said:

Before this earlier thread devolved into some trollery, it seemed to address this question pretty well:

In another thread, which I can't find, I described LLMs as "stochastic parrots," a term I borrowed from Emily Bender, a computational linguist.

https://en.wikipedia.org/wiki/Stochastic_parrot

Definitely not AI. Possibly AS, though.

Yes I remember that thread. Thanks for the stochastic parrot link. That’s quite informative. However what I still lack, after reading the criticisms of LLMs, is an idea of how true AI looks and what it does that is different from what LLMs do.

From my brief time hovering on the edges of this field, my takeaway was that genuine AI required internal goal setting and recursive self-improvement (RSI, which is an AGI system that enhances its own capabilities and intelligence without human intervention, leading to greater intelligence in a RW setting). To develop, I would think it needs a sandbox that's more like a world and not just acres and acres of text.

LLMs lack internal goal setting or anything like RSI and their sandbox is nothing but "stuff people wrote," which is pretty far from what an AGI needs. Seems like you would have to attach something to an LLM which would be unlike present specialized applications and which could somehow broadly develop internal goals and very creative ways to revise it's own software.

Some folks think you can get there with autonomous agents which are driven by simple rules. I have some doubts. I think artificial neural networks, modeled on biological cognition, offer some hope, though I don't quite understand digital simulation of neuronal activity (which is both digital and analog). Digital simulation is driven by algorithms and data, not feelings or survival needs or social imperatives, so it's a real philosophic thicket as to what you really have there.

For now, seems like you need a system that can take raw experiences and contextualize them in a broader "worldview," which would IMO mean some sense of pleasure and pain, some capacity for reflection on experiences which go well or not so well, some way to improvise when faced with novelty. And it would store not just data packages but also slippery sensations of missing something or lacking a pleasing outcome.

AI has been around for over 15 years. It took off with concepts of machine learning and then neural networks.

LLMs are a type of AI, specifically generative, where new combinations of old data are pulled together in statistical ways.

AGI is the big thing that most people think of when discussing the topic, and while some models are getting scary close, we’re not quite there yet.

Different models are good at different things. Some excel at images, others at musical composition, others at coding, and even within that some models are better at coding in some languages than in others.

Ultimately these are all tools. You can’t just hand someone a hammer and expect them to be able to build a cathedral. You can’t put any fool into a race car and have them set track records. You must know how to query them and know how to navigate their various eccentricities.

But yes. LLMs are a type of AI just like arithmetic is a type of maths.

8 hours ago, TheVat said:

From my brief time hovering on the edges of this field, my takeaway was that genuine AI required internal goal setting and recursive self-improvement (RSI, which is an AGI system that enhances its own capabilities and intelligence without human intervention, leading to greater intelligence in a RW setting). To develop, I would think it needs a sandbox that's more like a world and not just acres and acres of text.

LLMs lack internal goal setting or anything like RSI and their sandbox is nothing but "stuff people wrote," which is pretty far from what an AGI needs. Seems like you would have to attach something to an LLM which would be unlike present specialized applications and which could somehow broadly develop internal goals and very creative ways to revise it's own software.

..IOW, you need to create a virtual world for your own artificial intelligence, in which it would live, and learn from “birth” until “death”...

8 hours ago, TheVat said:

And it would store not just data packages but also slippery sensations of missing something or lacking a pleasing outcome.

And seemingly how does your brain store this data?

When the human brain is “happy,” not only data is stored, but also a chemical reaction, injection of hormones, neurotransmitters, etc. etc.

Such ChatGPT is "happy" that it could be useful, but there are no neutrotransmitter injections like in narcoman head after a dose here..

Edited June 10Jun 10 by Sensei

12 hours ago, TheVat said:

From my brief time hovering on the edges of this field, my takeaway was that genuine AI required internal goal setting and recursive self-improvement (RSI, which is an AGI system that enhances its own capabilities and intelligence without human intervention, leading to greater intelligence in a RW setting). To develop, I would think it needs a sandbox that's more like a world and not just acres and acres of text.

LLMs lack internal goal setting or anything like RSI and their sandbox is nothing but "stuff people wrote," which is pretty far from what an AGI needs. Seems like you would have to attach something to an LLM which would be unlike present specialized applications and which could somehow broadly develop internal goals and very creative ways to revise it's own software.

Some folks think you can get there with autonomous agents which are driven by simple rules. I have some doubts. I think artificial neural networks, modeled on biological cognition, offer some hope, though I don't quite understand digital simulation of neuronal activity (which is both digital and analog). Digital simulation is driven by algorithms and data, not feelings or survival needs or social imperatives, so it's a real philosophic thicket as to what you really have there.

For now, seems like you need a system that can take raw experiences and contextualize them in a broader "worldview," which would IMO mean some sense of pleasure and pain, some capacity for reflection on experiences which go well or not so well, some way to improvise when faced with novelty. And it would store not just data packages but also slippery sensations of missing something or lacking a pleasing outcome.

Yes I wasn’t really thinking of AGI but how LLMs compare with the so-called AI applications that have been developed for specific purposes, such as interpreting medical X-ray mages and things like that. Something to do with using neural networks to learn, perhaps?

3 hours ago, exchemist said:

Yes I wasn’t really thinking of AGI but how LLMs compare with the so-called AI applications that have been developed for specific purposes, such as interpreting medical X-ray mages and things like that. Something to do with using neural networks to learn, perhaps?

LLMs use transformers in their architecture which is a type of NN called a recurrent NN. It's good for temporal relations and sequences of textual data.

Pattern recognition like image or video analysis use CNN, convolutional NNs. CNNs are more "biological." Here's a basic look. (I want to read these, too, and brush up)

https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)

https://en.wikipedia.org/wiki/Convolutional_neural_network

LLM's are hopelessy bottom of the class in Mathematics.

One problem with this is if you ask a mathematical question in words it has no way of discerning what the actual maths is.

I think this is because it trawls more respectable sites (such as Wikipedia) for the words, without the mathematical understanding.

Here is an example of an utterly incorrect declaration from Google AI

The moral of this is do not use a language based construct to do or check your maths homework.

( in this case Wiki also has it wrong)

2 minutes ago, studiot said:

LLM's are hopelessy bottom of the class in Mathematics.

One problem with this is if you ask a mathematical question in words it has no way of discerning what the actual maths is.

I think this is because it trawls more respectable sites (such as Wikipedia) for the words, without the mathematical understanding.

Here is an example of an utterly incorrect declaration from Google AI

The moral of this is do not use a language based construct to do or check your maths homework.

( in this case Wiki also has it wrong)

This is not surprising to me. As I have understood the issue so far, LLMs are AI in the sense they use neural networks to learn how to construct sentences and create bodies of original text. So they are very ingenious at that. But that's all. They are brain-dead when it comes to the meaning of the words, i.e. the content of the sentences they read and write. Emily Bender describes them as "stochastic parrots" because they simply look up a load of references on the subject they have been given and construct a reply out of the statistically most common features they encounter. A further characteristic of LLMs is that, because they are designed to engage the user in chat, they will obsequiously try to construct an answer that tells the user he or she is right, or at least one that lets them down very gently if they are wrong.

The obvious danger is that lazy users, or those lacking adequate critical faculties, will feel their wrong ideas are vindicated because "AI has told me I'm right". As these false responses build up in the on-line bank of knowledge, there is the further danger that this "botshit" will get start to be ingested by other LLMs and repeated, causing progressive contamination of the knowledge base.

58 minutes ago, exchemist said:

The obvious danger is that lazy users, or those lacking adequate critical faculties, will feel their wrong ideas are vindicated because "AI has told me I'm right". As these false responses build up in the on-line bank of knowledge, there is the further danger that this "botshit" will get start to be ingested by other LLMs and repeated, causing progressive contamination of the knowledge base.

There are even worse consequences. One we have seen is that folks are not only taking in LLM or AI outputs uncritically. More and more, they are not even reading the outputs and just parrot them. Basically folks parrot those stochastic parrots. The danger here is that folks are not only losing the ability of critical thinking- basic skills like reading and reading comprehension is also diminishing.

I am aware that in the past there there was always a kind of moral outrage from the older generation lamenting how tools are going to degrade human skills. However, we are starting to see real shifts in basic reading skills especially (but not exclusively) in the younger generation, which are not filled by other means. And in contrast to archaic skills that have become obsolete, reading and being able to process information is, in my mind, a foundational skill that is needed in every society, regardless how technologically advanced it is.

For the first time we are seeing evidence of things happening that old folks have been yelling about for millennia, mostly because it is happening way faster than anything else I am aware of.

Just now, CharonY said:

There are even worse consequences.

Agrreed in that there is a self reinforcing aspect.

The LLM returns something as an answer

Others then quote the LLM

So the next time a similar question is asked the LLM weighs into account a quote of its previous statement.

Und So Weiter.

An MIT study has been doing the rounds.

https://www.cnet.com/tech/services-and-software/do-you-really-learn-when-you-use-ai-what-mit-researchers-found/

https://www.edweek.org/technology/brain-activity-is-lower-for-writers-who-use-ai-what-that-means-for-students/2025/06

Some main points:

People who used LLMs to do the heavy lifting for their essays often didn't even remember it was their essay when shown it later. (So much for learning.)

People who used LLMs to do the heavy lifting often still did worse when later not using an LLM.

I will also note that there are some studies showing a beneficial effect, but they were all in the context of a study where use was directed (e.g. in classroom settings). There sometimes benefits are seen, but I think this is a common issue in testing benefit of approaches and technologies for education.

These studies are often with selected or self-selected students and because the it is a novelty, almost everything you do (multimedia, AI, exercise, dancing) students get more engaged and perform better (for a while). But after implementation in large scale, the benefits vanish or things get worse.

One area where I have seen actual improvement is in underserved regions, provided the infrastructure has at least some level of computerization (e.g. in form of cell phones).

6 hours ago, studiot said:

LLM's are hopelessy bottom of the class in Mathematics.

Have you tried the Wolfram Alpha AI or the Julius AI?

Just like not all cars can be used to tow a camper or a boat, not all models should be used to help with math. Doesn’t mean it’s impossible to pull a boat with a vehicle or do math with a model.

On 7/6/2025 at 7:14 PM, studiot said:

Here is an example of an utterly incorrect declaration from Google AI

Honestly. Google AI, at the moment, is a crap.

I asked ChatGPT “how many neutrons are in a water molecule, with separation into isotopes,” and it provided this:

The problem is trust. Overconfidence.

People who are considered specialists in a certain field are also trusted. Which can have disastrous consequences if they make a mistake.

A person who is a layman, and just wants to get a quick answer to a question that concerns him/her will not want (or will not know how to do it at all) to verify the answers from a “specialist”..

Much greater reliability is expected from a machine than from a human being. It doesn't make sense.

LLMs are AI models that simulate human language understanding through complex patterns and data. While they aren't conscious, they demonstrate genuine AI capabilities in generating coherent, context-aware text, making the claim that they are just hype an underestimation of their technological advancements.

Wow. Powerful insights there. It's phenomenal what humans can post nowadays

Wait for neurosymbolic AI guys.

19 minutes ago, Otto Kretschmer said:

Wait for neurosymbolic AI guys.

Wot dat?

3 hours ago, exchemist said:

Wot dat?

It's an AI architecture that will likely replace the currently dominant transformer architecture within a few years.

The "symbolic" part of it (usage of knowledge graphs, logic rules etc.) allows neurosymbolic AI to derive the same knowledge from drastically smaller amount of information.

As an example - in order to learn what a cat is, regular neural networks need to be given tens of thousands of cat photos - if the training data only includes cats that are indoors, the model will struggle to recognize a cat in a forest, a city etc. since it learns correlation, not the underlying concept.Neurosymbolic AI does not have this problem because it can learn what "catness" is about from just a handful of pictures or even a single one. When teaching neurosymbolic AI to stop at red, you no longer need to show it dozens of hours of cars stopping at red lights in every possible weather and lightning conditions - it's enough to show it a single video and it'll derive the general rule. The neural part handles perception ("That's a traffic light," "The color is red"). The symbolic part has a hard-coded rule: IF object_is_traffic_light AND color_is_red THEN initiate_stop(). The model doesn't need to learn this core traffic law from data; it's given as a fact. It only needs enough data to learn to recognize the objects reliably. It's creativity is also going to be higher since the AI will no longer have the need for every possible outcome to be included in the training data in order to know it.

This is a rather basic answer and I guess an AI enthusiast more knowledgeable than I could provide a more detailed answer.

Edited July 16Jul 16 by Otto Kretschmer

Just now, Otto Kretschmer said:

When teaching neurosymbolic AI to stop at red, you no longer need to show it dozens of hours of cars stopping at red lights in every possible weather and lightning conditions - it's enough to show it a single video and it'll derive the general rule. The neural part handles perception ("That's a traffic light," "The color is red"). The symbolic part has a hard-coded rule: IF object_is_traffic_light AND color_is_red THEN initiate_stop(). The model doesn't need to learn this core traffic law from data; it's given as a fact. It only needs enough data to learn to recognize the objects reliably. It's creativity is also going to be higher since the AI will no longer have the need for every possible outcome to be included in the training data in order to know it.

It would be much more interesting to know what rule it would derive for a green light.

Can you shed any light on that ?

23 minutes ago, studiot said:

It would be much more interesting to know what rule it would derive for a green light.

Can you shed any light on that ?

Like I said, I am not an AI expert or even a major AI buff, just a regular user unfortunately. :( I would recommend asking on Reddit in some AI or tech-related subreddits.

Just now, Otto Kretschmer said:

Like I said, I am not an AI expert or even a major AI buff, just a regular user unfortunately. :( I would recommend asking on Reddit in some AI or tech-related subreddits.

I asked you because you are the one who posted the material I quoted and are responsible for its veracity or otherwise.,