AGI is coming

[thenondualtankie]

A lot of people are in denial about AGI.

The reason AGI is coming is because we've already achieved some level of intelligence via things like ChatGPT, and it's a simple scaling of this intelligence. The reason people believe that scaling will happen is because it has shown no sign of stopping. Scaling isn't going to suddenly stop working once we're just below human intelligence level. This is utter anthropocentrism; human intelligence is special and computers cannot ever reach this!

Whether ChatGPT actually attains some level of intelligence: if we just focus on the 'IQ' aspect of intelligence, then ChatGPT absolutely has this. This means ChatGPT is able to solve new problems it has not seen before. For example GPT-4 is able to pass software engineering interviews using unseen questions. You can cope and seethe about how 'it's just combining things from its training data, not creating anything new', but that is literally how intelligence works: combining old information to create new.

People have this massive bias in favour of 'things will always continue the way they currently are', probably because their entire survival strategy depends on it. You see this especially starkly among people like software engineers, where AI hasn't fully hit them just yet, meaning they can still get away with not facing the reality that AI will surpass their abilities.

[erik8lrl]

2 hours ago, zurew said:

Yeah I agree that it can be useful for multiple different things , but I tried to show some of things tthat I and others have recognized regarding its semantic understanding and its reasoning capability.

The article I linked shows many examples ,but here is another one with many examples:

https://medium.com/@konstantine_45825/gpt-4-cant-reason-2eab795e2523

I recommend to check this one as well:

https://amistrongeryet.substack.com/p/gpt-4-capabilities

 

This is also interesting as well:

 

I sent these articles to GPT4 to ask its opinions lol, and this is what it gave me:

ChatGPT

The argument that GPT-4, and by extension, other large language models (LLMs), cannot truly reason is valid within the context provided. It highlights the distinction between the appearance of understanding or reasoning and the actual cognitive processes involved. LLMs are trained on vast datasets to predict the likelihood of sequences of words, which allows them to generate plausible-sounding responses. However, this process does not equate to genuine reasoning or understanding, as it lacks the ability to consider context, apply logic consistently, or understand the underlying principles in the way humans do. The critique is rooted in the current limitations of AI in mimicking the depth and flexibility of human thought processes.

User
Do you think this will be improved as the size of data and parameters increase? 

ChatGPT
Improvements in AI's reasoning capabilities with larger datasets and more parameters are expected to some extent, as these enhancements can lead to better pattern recognition, more nuanced language understanding, and more sophisticated response generation. However, true reasoning involves more than just processing power or data scale; it requires the ability to understand context, apply logical principles, and make inferences based on a deep understanding of the world. While scaling up might address some limitations, addressing the fundamental gap between simulating reasoning and actual understanding will likely require advances in AI approaches and architectures beyond just increasing size.

User
What do you think such advances could be? Search the internet for the latest research if you need to.

ChatGPT
Advances in AI reasoning capabilities are indeed on the horizon, driven by innovative architectures and training paradigms that could potentially bridge the gap between current limitations and more human-like reasoning. Two notable developments are Google DeepMind's Gemini 1.5 and Google's Pathways Language Model (PaLM), as well as insights from Yann LeCun on building AI systems that learn and reason like humans and animals.

Google DeepMind's Gemini 1.5 introduces enhanced performance through a more efficient architecture, capable of sophisticated reasoning tasks across different modalities, including text and coding. This model represents a significant step forward, surpassing previous state-of-the-art models in understanding and problem-solving capabilities across a range of benchmarks. It shows that as models become more capable and architectures more efficient, AI's ability to reason and understand complex scenarios improves significantly.

Google's PaLM, with its breakthrough in scaling to 540 billion parameters, demonstrates impressive capabilities in language understanding, generation, and reasoning tasks. Its use of chain-of-thought prompting for multi-step arithmetic and common-sense reasoning tasks showcases how large language models can be pushed towards more sophisticated reasoning abilities. The model's ability to generate explicit explanations for complex scenarios indicates a move towards AI systems that can not only solve problems but also articulate their thought processes in a way that mimics human reasoning.

Yann LeCun proposes an architecture for autonomous intelligence that could enable machines to learn world models in a self-supervised fashion. This approach involves constructing predictive models of the world that can simulate possible future states and use these simulations to predict, reason, and plan. Such an architecture, which includes modules for perception, world modeling, and action planning, represents a comprehensive framework for developing AI systems capable of complex reasoning and decision-making.

These advancements suggest that while scaling up data and parameters can lead to improvements in AI's reasoning capabilities, significant leaps also require innovative approaches to model architecture and training. By developing models that can better understand and simulate the world, and by employing strategies like self-supervised learning and chain-of-thought reasoning, we move closer to AI systems that can reason in ways that are more aligned with human cognitive processes.

For more details, you can explore the developments in Gemini 1.5 on Google DeepMind's website and PaLM on Google's Research Blog, as well as Yann LeCun's vision for AI on Meta AI's page.


 

---

The world simulation model is exactly what OpenAI, Google, and even Tesla are working on right now. We can already see it with Sora, GPT5 will likely include Sora in its model understanding. Similar with Google Gemini going multi-modal. 

Edited February 18 by erik8lrl

[Butters]

If you want to make some money this year, remember: the people who made most during the gold rush were selling the shovels. 

[Scholar]

2 hours ago, erik8lrl said:

Very cool! Yeah, it's not perfect, we'll see how it develops. 

It's not about perfection. Humans aren't perfect either.

 

You are missing that this is fundamentally doing something different than humans do when using their intelligence. We don't inform intelligence through unreasonable amounts of data, we inform intuition through unreasonable amounts of data.

[erik8lrl]

 

[erik8lrl]

1 hour ago, Scholar said:

It's not about perfection. Humans aren't perfect either.

 

You are missing that this is fundamentally doing something different than humans do when using their intelligence. We don't inform intelligence through unreasonable amounts of data, we inform intuition through unreasonable amounts of data.

Yes, if you want to word it this way. It's true, that we don't inform intelligence through data alone, we inform them through connection and pattern recognition, which is what neural networks do also aside from data collection.  
My point is we don't know that it is fundamentally different for sure, it's still too early to tell. We don't know enough about neural networks to definitively say that it might not be the starting point that could lead to Qualia. 

Edited February 18 by erik8lrl

[erik8lrl]

 

[Shodburrito]

@Leo Gura They can, the problem is simply they can't put powerful enough chips that will fit into a car.

 

I believe I read an article that Waymo already has less crashes per mile than the average driver. Of course this isn't all terrain and road conditions but the models now are the worst they will be from here on out. They will only improve. If clunky biological humans can drive a car then precise machines will be able to one day too. 

Edited February 19 by Shodburrito

[Shodburrito]

 l love how everyone in here is acting like their predictions about AI just being a hype cycle and how it is overblown is true because y'all are clearly the experts on AI. Dunning Kruger Effect on full display

[Shodburrito]

I'm sure the experts in the field who nearly all are saying we will have AGI by the end of the decade are just wrong because you all are obviously top executives at the largest AI companies and have access to their best models. There was a leak over a year ago that SORA was developed in MARCH 2023 but they've delayed release until now due to public fear over AI. They have way better models behind closed doors but you all aren't ready for that conversation. 

[Shodburrito]

Watch this video: 

 

 

[erik8lrl]

3 hours ago, Shodburrito said:

I'm sure the experts in the field who nearly all are saying we will have AGI by the end of the decade are just wrong because you all are obviously top executives at the largest AI companies and have access to their best models. There was a leak over a year ago that SORA was developed in MARCH 2023 but they've delayed release until now due to public fear over AI. They have way better models behind closed doors but you all aren't ready for that conversation. 

Totally. It's possible they already have AGI or getting really close.

[zurew]

6 hours ago, Shodburrito said:

l love how everyone in here is acting like their predictions about AI just being a hype cycle and how it is overblown is true because y'all are clearly the experts on AI. Dunning Kruger Effect on full display

I love how you were incapable to engage directly with anything that was said and had to strawman everyone and you had to use one of the worst arguments to make a case for your position.

Most of us didn't say that AGI definitely won't come in the next 5-10 years, most of us just tried to point to the fact, that the arguments that are used to make a case for such claims are weak and have a lot of holes in them and some of you have unreasonably  high confidence in such claims.

 

If you really want to make the "but experts said this  and you reject that" argument, then

First show us a survey where there is a high consensus between AI experts on when will AGI emerge (But it has to be such a survey, where AGI has a concise definition, so that those experts have the same definition in mind when they answer the question)

Secondly, you will have to make a case for why this time the experts prediciton will come true (because there is a clear inductive counter argument to this survey argument - namely, that most of the AI experts predictions in the past have failed miserably).

Edited February 19 by zurew

[zurew]

21 hours ago, erik8lrl said:

The world simulation model is exactly what OpenAI, Google, and even Tesla are working on right now. We can already see it with Sora, GPT5 will likely include Sora in its model understanding. Similar with Google Gemini going multi-modal.

WIll look forward to GPT5.

I will give one more thing that they will eventually need to solve: Right now it seems to be the case, that AI doesn't have an abstract understanding of most things (I already said this), but more specifically,  doesn't have an abstract understanding of negation or the negative. The proof for this is the fact, that in a lot of cases, when you tell it to not do something - it will still do it.  Yes in some cases it might get it right, but this is a principle problem, where you either have a real understanding of negation or you don't. 

This includes instances, where you want to create an image and say 'please don't inlcude x on the image' or if you want it to not include a specific thing in its answer. This problem seem to be a tough one , given  that you can't just show it a clear pattern what negation is just by using a dataset that has a finite set of negating things. - in fact, I would say the whole category of 'abstract understanding of things' cant be learned by purely just trying to find the right pattern between a finite set of things. - it seem to require a different approach in kind. Because, for instance, if I have a prompt of 'don't generate a mushroom on the image" - I would need to show an infinite number of instances of images that don't include a mushroom on it,  and even then the prompts full meaning wouldn't be fully captured.

I will grant though, that AI won't necessarily need to have a real abstract understanding of things to do a lot of tasks, but still; eventually - to make it as reliable as possible, some solution will need to be proposed to this problem.

Edited February 19 by zurew

[erik8lrl]

4 hours ago, zurew said:

I will give one more thing that they will eventually need to solve: Right now it seems to be the case, that AI doesn't have an abstract understanding of most things (I already said this), but more specifically,  doesn't have an abstract understanding of negation or the negative. The proof for this is the fact, that in a lot of cases, when you tell it to not do something - it will still do it.  Yes in some cases it might get it right, but this is a principle problem, where you either have a real understanding of negation or you don't. 

Yeah, I think this is mainly caused by the size of the context window and the architecture. Right now, even with GPT4, the context window is small, meaning it can't remember long-term conditions naturally. You could get around this sometimes with prompting, something like "Don't do (this) from now on when answering my questions". You can also use the new memory feature that's released recently, it basically allows you to add a system prompt to the GPT so that it remembers it and applies it globally.  
 

5 hours ago, zurew said:

This includes instances, where you want to create an image and say 'please don't inlcude x on the image' or if you want it to not include a specific thing in its answer. This problem seem to be a tough one , given  that you can't just show it a clear pattern what negation is just by using a dataset that has a finite set of negating things. - in fact, I would say the whole category of 'abstract understanding of things' cant be learned by purely just trying to find the right pattern between a finite set of things. - it seem to require a different approach in kind. Because, for instance, if I have a prompt of 'don't generate a mushroom on the image" - I would need to show an infinite number of instances of images that don't include a mushroom on it,  and even then the prompts full meaning wouldn't be fully captured.

I think if you are using GPT4/Dalle3 to generate images, you won't have a good time with customization. Dalle-3 is really good at text interpretation, but it can't really do inpainting or negative prompts. How it changes images is by reverse prompting text description from your images, and then using that prompt to generate similar results, but it's not editing the images, so you won't have consistency. GPT4/Dalle-3 is really only good for generating initial images that require complex text interpretation. For image generation, I think it's best to use Midjourney, SD, or if you can, learn comfyUI. These give you far more customization ability and negative prompting works great for them. With comfyUI you can even customize the lower-level processes of the models to fit your needs, there are also endless workflows shared by other users for specific image generation tasks. GPT4/Dalle-3 is really not designed for image generation, GPT5, however, will be a different story.  

[Scholar]

5 hours ago, zurew said:

The proof for this is the fact, that in a lot of cases, when you tell it to not do something - it will still do it. 

Because it is intuitive. It is for this same reason that, when you think "Don't think about an elephant", that your subconscious mind will provide you with an elephant.

The subconscious and unconscious mind are basically neural networks, and they aren't intelligent because they aren't conscious.

 

We are on actualized.org, how can people be so confused about the nature of intelligence? Just go sit down and meditate for a while. You can directly see what it is, and how absurd the notion is that somehow complexity will lead to individuated consciousness.

 

Edited February 19 by Scholar

[zurew]

36 minutes ago, erik8lrl said:

Yeah, I think this is mainly caused by the size of the context window and the architecture. Right now, even with GPT4, the context window is small, meaning it can't remember long-term conditions naturally. You could get around this sometimes with prompting, something like "Don't do (this) from now on when answering my questions". You can also use the new memory feature that's released recently, it basically allows you to add a system prompt to the GPT so that it remembers it and applies it globally.  

I disagree that it can be solved by a better context window. The nature of the problem is much deeper than that.  

Its not a matter of forgetting something , its a matter of not doing it in the firstplace.The examples that I mentioned were ones, where you give it a prompt and it immediately fails to do what the prompt says  (not like it fails to maintain a long term condition that you gave it few responses ago) . Its like User: Hey GPT dont do  x GPT4: does x. 

Edited February 19 by zurew

[thenondualtankie]

Zurew, I'd like you to provide some examples of that.

You have no idea what you're talking about.

GPT-4 does not have perfect reliability in following instructions, including negations. This also applies to humans. It says nothing about its general ability to understand.

If it doesn't understand, then how do you explain its ability to solve unseen problems? For example, as I mentioned above, it can solve a wide array of programming problems it has never come across. You could probably google more examples of its problem solving.

[zurew]

1 hour ago, thenondualtankie said:

GPT-4 does not have perfect reliability in following instructions, including negations. This also applies to humans. It says nothing about its general ability to understand.

You still don't get the depth of the problem. Its not a matter of sometimes I can follow the instructions and sometimes I don't. Its not a cherry pciking problem, its different in kind.

You either know the abstract meaning of negation or you don't. If you tell a human to not do x - that human won't do that action , given that he/she understands what that 'x' is. If you do the same thing with an AI, it will still do the action even if it knows what x is. And not just that, but in its response it will give the bullshit memorized take of "Yeah I apologize for my mistake, I won't put x on the image anymore" and then right after that it generates an image that has x on it. 

I will apply this to a programming problem, because it will illustrate perfectly whats wrong with the saying of "sometimes it can work and sometimes it cant" - If you know programming you will understand this: I can write a function that will ask for an integer input, and after that it will print out that input. Now given that the condition is met correctly ( that I give an integer as an input, that the computer can read) it wouldn't make any sense to say that yeah this function will only work 40% of the time or 30% of the time - no, once all the required conditions are met - It either can perform the function or it cant.

Another way to put it is to conflate inductive and deductive arguments. When it comes the deductive arguments the conclusion will always follow from the premises, but when it comes to inductive ones, the conclusion won't necessarily follow 100% of the time.

Another way to put is to talk about math proofs. It would make 0 sense to say that this x math proof only works 90% of the time.

 

1 hour ago, thenondualtankie said:

If it doesn't understand, then how do you explain its ability to solve unseen problems? For example, as I mentioned above, it can solve a wide array of programming problems it has never come across. You could probably google more examples of its problem solving.

I already gave a response to this:

Quote

Right now you could do this with me:

Give me a foreign language that I understand literally nothing about - in terms of meaning of sentences and words - and then give me a question in that foreign language and the right answer below. If I memorize the syntax (meaning, if I can recognize  which symbol comes after which other symbol) then I will be able to give the right answer to said question even though I semantically understand nothing about the question nor about the answer - I can just use the memorized patterns.

The AI seem to be doing the exact same ,except with a little twist that it can somewhat adapt said memorizedpatterns and if it sees a pattern that is very similar to another pattern that it already ecountered with in its training data, then - in the context of answering questions - it will assume the answer must be the exact same or very similar to it, even though changing one word or adding a , to a question might change its meaning entirely.

https://medium.com/@konstantine_45825/gpt-4-cant-reason-2eab795e2523 - this articles breakes down pretty thoroughly the problems with  GPT4's reasoning capability and understanding of things.

Edited February 19 by zurew

[erik8lrl]

8 hours ago, zurew said:

You still don't get the depth of the problem. Its not a matter of sometimes I can follow the instructions and sometimes I don't. Its not a cherry pciking problem, its different in kind.

You either know the abstract meaning of negation or you don't. If you tell a human to not do x - that human won't do that action , given that he/she understands what that 'x' is. If you do the same thing with an AI, it will still do the action even if it knows what x is. And not just that, but in its response it will give the bullshit memorized take of "Yeah I apologize for my mistake, I won't put x on the image anymore" and then right after that it generates an image that has x on it. 

I will apply this to a programming problem, because it will illustrate perfectly whats wrong with the saying of "sometimes it can work and sometimes it cant" - If you know programming you will understand this: I can write a function that will ask for an integer input, and after that it will print out that input. Now given that the condition is met correctly ( that I give an integer as an input, that the computer can read) it wouldn't make any sense to say that yeah this function will only work 40% of the time or 30% of the time - no, once all the required conditions are met - It either can perform the function or it cant.

Another way to put it is to conflate inductive and deductive arguments. When it comes the deductive arguments the conclusion will always follow from the premises, but when it comes to inductive ones, the conclusion won't necessarily follow 100% of the time.

Another way to put is to talk about math proofs. It would make 0 sense to say that this x math proof only works 90% of the time.

 

I already gave a response to this:

https://medium.com/@konstantine_45825/gpt-4-cant-reason-2eab795e2523 - this articles breakes down pretty thoroughly the problems with  GPT4's reasoning capability and understanding of things.

Yeah, I think current models definitely have this problem, especially for complex logical reasoning. It might be solved with more parameters or better architecture. We do see that the scaling of "intelligence" and reasoning ability increase as we scale. I think the new Gemini 1.5's reasoning ability scored 80-90% on benchmarks. We'll see if newer models like GPT5 improve this area, it might or might not be a problem related to scale since we still don't fully know what emergent property can appear as we scale larger and larger. I think for me, a lot of the times when GPT4 doesn't do what I ask, it's mostly because of my prompt not being specific or detailed enough, most of the time I find improving my prompting helps, but not always tho. 
I think for coding, context window size does help, for me, most of the code GPT4 gives me is out of context, so they are often the most common implementation of something. If Gemini can generate code in the context of the whole system, then it should produce better results.