LLMs can not become AGI. It's not possible. You don't fundamentally change a thing just by adding more compute

[inyerta]

A calculator that has infinite compute is still just a calculator, just like an LLM with infinite compute is still just an LLM. It will never be anything more than "predict the next token based on the preceding context." Its "knowledge" is entirely parasitic.

"AI" is used as an umbrella term so people can say, "what about robots? what about self driving cars?" which are entirely different classes of things from LLMs. They are not the same thing and shouldn't be lumped together.

Word prediction algorithms are factually incapable of evolving into whatever they're calling a digital superintelligence this week. This kind of shit is just more marketing wank.

 

[dimocel]

A calculator that has infinite compute is still just a calculator, just like an LLM with infinite compute is still just an LLM. It will never be anything more than "predict the next token based on the preceding context." Its "knowledge" is entirely parasitic.

"AI" is used as an umbrella term so people can say, "what about robots? what about self driving cars?" which are entirely different classes of things from LLMs. They are not the same thing and shouldn't be lumped together.

Word prediction algorithms are factually incapable of evolving into whatever they're calling a digital superintelligence this week. This kind of shit is just more marketing wank.

Well said, completely agree. PS: Have you read AI 2027 before? Great article, and it's scary how similar it mirrors our current reality.

 

[DeborahAnnWollFARTS]

Human cope
Get replaced by Big Black AI

 

[o_Owtf]

My guess is that the core idea of the LLM architecture, where you have these huge Hilbert spaces that are trained with vast amounts of data, is indeed the right idea. However, the thing that constrains LLMs is how you interface with them, everything is just tokens. Researchers are hopelessly trying to represent images, audio, even videos as strings of tokens to make it readable for the LLM.

 

[inyerta]

My guess is that the core idea of the LLM architecture, where you have these huge Hilbert spaces that are trained with vast amounts of data, is indeed the right idea. However, the thing that constrains LLMs is how you interface with them, everything is just tokens. Researchers are hopelessly trying to represent images, audio, even videos as strings of tokens to make it readable for the LLM.

Which means they will need a nuclear reactor to power the damn thing, progress is not infinite

@Jason Voorhees

 

[amisci]

A calculator that has infinite compute is still just a calculator, just like an LLM with infinite compute is still just an LLM. It will never be anything more than "predict the next token based on the preceding context." Its "knowledge" is entirely parasitic.

"AI" is used as an umbrella term so people can say, "what about robots? what about self driving cars?" which are entirely different classes of things from LLMs. They are not the same thing and shouldn't be lumped together.

Word prediction algorithms are factually incapable of evolving into whatever they're calling a digital superintelligence this week. This kind of shit is just more marketing wank.

High iq thread. The problem with an AGI is we have no actual mathematical foundation for it we literally don't know how to make a system that doesn't need input to generate output. Pretty sure we will get AGI when a team of 20 Terrance taos spend their whole life studying what ever obscure math branch Machine learning comes from

 

[amisci]

My guess is that the core idea of the LLM architecture, where you have these huge Hilbert spaces that are trained with vast amounts of data, is indeed the right idea. However, the thing that constrains LLMs is how you interface with them, everything is just tokens. Researchers are hopelessly trying to represent images, audio, even videos as strings of tokens to make it readable for the LLM.

You're right with the direction part. LLMs for me seem like a single part of an actual AGI. It is definitely the key but we can't spend ages just scaling up and scaling up the same random bullshit model The human brain barely needs any electricity jfl

 

[o_Owtf]

Which means they will need a nuclear reactor to power the damn thing, progress is not infinite

@Jason Voorhees

Machine learning needs a fuckton of data which needs a fuckton of flops, comes with the territory.

 

[o_Owtf]

You're right with the direction part. LLMs for me seem like a single part of an actual AGI. It is definitely the key but we can't spend ages just scaling up and scaling up the same random bullshit model The human brain barely needs any electricity jfl

High iq thread. The problem with an AGI is we have no actual mathematical foundation for it we literally don't know how to make a system that doesn't need input to generate output. Pretty sure we will get AGI when a team of 20 Terrance taos spend their whole life studying what ever obscure math branch Machine learning comes from

Yeah, it's quite remarkable how it's so hard to capture the efficiency and generality that neurons seem to bring. There's been experiments where researchers put rat neurons in petri dishes hooked up to some kind of pong game, and the neurons organize themselves to play the game within like 5 minutes. It feels like it should be trivial to model this mathematically, alas we're missing something that even really competent mathematicians can't suss out.

 

[shngstaaaa107]

Yeah, it's quite remarkable how it's so hard to capture the efficiency and generality that neurons seem to bring. There's been experiments where researchers put rat neurons in petri dishes hooked up to some kind of pong game, and the neurons organize themselves to play the game within like 5 minutes. It feels like it should be trivial to model this mathematically, alas we're missing something that even really competent mathematicians can't suss out.

Idk biology makes for a wonderful computing system by itself

 

[o_Owtf]

Idk biology makes for a wonderful computing system by itself

Yeah apparently it has been proposed that you could just use big loaves of neurons hooked up to input from computers for training. The problem imo is that neurons are mechanically limited in a way that computers aren't. Transistors don't get fatigued in the same that neurons do, they are a lot faster, they are a lot more compact, etc.

The dream would be if you could aid the biological system with computer processing, but then you're just back at square one with like what we're trying to do with LLMs today. You'll just end up making more and more add-ons to try and circumvent the inevitable flaws in the original method.

 

[shngstaaaa107]

Yeah apparently it has been proposed that you could just use big loaves of neurons hooked up to input from computers for training. The problem imo is that neurons are mechanically limited in a way that computers aren't. Transistors don't get fatigued in the same that neurons do, they are a lot faster, they are a lot more compact, etc.

The dream would be if you could aid the biological system with computer processing, but then you're just back at square one with like what we're trying to do with LLMs today. You'll just end up making more and more add-ons to try and circumvent the inevitable flaws in the original method.

I suppose the idea would be to abandon LLMs completely and go invent consciousness

Really though, I don't doubt we could let AI try to reason and "Think critically" and get close enough but really most scientists also struggle with the magic of how LLMs work or getting AI to produce something completely new under its current framework

 

[o_Owtf]

I suppose the idea would be to abandon LLMs completely and go invent consciousness

I think that would be the final invention humanity would ever need

Really though, I don't doubt we could let AI try to reason and "Think critically" and get close enough but really most scientists also struggle with the magic of how LLMs work or getting AI to produce something completely new under its current framework

For sure AI can do a lot of crazy shit nowadays but the question always remains whether or not the system is actually creating something novel or if it's just overfitting the input material.

This is an especially worrying prospect since there's been worries that new training material for AI today is being contaminated with AI generated data.

You can see it with how AI art tends to have a yellow tint to it. It's being hypothesized that image learning data is being re-fed into the same models that created it, reinforcing the slight bias for yellow further and further. A similar process could be happening with LLMs where their sometimes idiosyncratic language is being put into itself, making it even more and more odd/noticeably computer generated.

 

[shngstaaaa107]

I think that would be the final invention humanity would ever need

For sure AI can do a lot of crazy shit nowadays but the question always remains whether or not the system is actually creating something novel or if it's just overfitting the input material.

This is an especially worrying prospect since there's been worries that new training material for AI today is being contaminated with AI generated data.

You can see it with how AI art tends to have a yellow tint to it. It's being hypothesized that image learning data is being re-fed into the same models that created it, reinforcing the slight bias for yellow further and further. A similar process could be happening with LLMs where their sometimes idiosyncratic language is being put into itself, making it even more and more odd/noticeably computer generated.

Yeah we are probably approaching a point where AI trains off good AI images

 

[WELOVELOOKS]

idk what LLMs even are so ima just tag @Jason Voorhees

 

[shngstaaaa107]

idk what LLMs even are so ima just tag @Jason Voorhees

Unironically good person to tag

 

[MoneyMaka]

right the human biology is a bunch of different super computers doing different tasks connected to a processing node
llms are siply one part of the super computer

they're like training biceps in hopes that the bicep of the body will be able to think
it won't

 

[MoneyMaka]

to add: language is just the front facing mehanism that's the easiest to associate with human beings
so it's like the interface to interact with a human being

which makes it the most profitable. you advance it just far enough and you get people to believe it's a human
but a humn, and general intelligence is muh more than just the interface

 

[Jason Voorhees]

idk what LLMs even are so ima just tag @Jason Voorhees

I made a thread if you are interested in reading

https://looksmax.org/threads/the-2026-ai-landscape-with-sources-under-every-claim.2179541/

 

[Informationcapitali]

I made a thread if you are interested in reading

https://looksmax.org/threads/the-2026-ai-landscape-with-sources-under-every-claim.2179541/

Will jason voorhees be replaced by ai soon or are people still trying to figure out how to replace you?

And what will u do if they finally get you replaced?

Is it possible that you could already be ?

 

[inyerta]

Will jason voorhees be replaced by ai soon or are people still trying to figure out how to replace you?

And what will u do if they finally get you replaced?

Is it possible that you could already be ?

Ai LLM's lack context sensitivity and nuances of human interaction.

So it will never truly replace people no matter how much money is thrown at it

 

[Informationcapitali]

We will not be replaced @Mai Sakurajima.

Can we say were truly unique and special people at this very moment with our unique quirks if Not even ai can replace us. Are we truly special as society says or are we just basic rehashes of other people who have tragically died a scary fate?