Why are we teaching our LLMs and similar AI models backwards? We give them the whole internet or just wikipedia and bunch of high quality books and expect it to generalize. If it is true that the AI transformer architecture is like a human brain, shouldn't we even try to attempt teaching it like a child? When you want your child to become a physicist you don't throw PHD level physics textbooks at them and tell them to read more and more until they start understanding it, so why are we doing this shit when teaching LLMs and multimodal models?

We should be making models that can do basic stuff toddler could do and then scale up from there. Make the LLM understand cause and effect and spelling before giving it texts explaining quantum mechanics. No fucking wonder that all LLMs and multimodal models are extremely predisposed to overfitting and require billion times more learning data then a human when they are learning the fact that trees are green by reading 20 fantasy novels and maybe catching the fact that when characters describe green forest they mean the tree leaves are green rather then like grass or moss. Testing LLMs on PHD level problems and competing with the smartest highschoolers at math, yet having the hardest test be connecting fucking boxes on a image is absolute joke. Moravec's paradox probably isn't the law of the fucking universe and people pretending it is is why there is next to no progress in the field. Why did none of the greatest minds being offered billion dollar salaries by Zuck think of just teaching it simple verbal and special puzzles first and then dumping the internet on it? And if they did then why aren't there news about some dummy toddler model crushing these basic visual benchmarks that hypemen billionaires waste billions trying to edge out extra 2% improvement on?

Who's stopping you?

Skibidi gyat

>>106142025 (OP)
>If it is true that the AI transformer architecture is like a human brain
It isn't.
Or rather it's like a horribly over-simplified model of some of the less sophisticated parts of the brain. It omits the parts that allow fast learning, and it omits the parts that allow handling constraints.

But sama wants trillions of dollars of all your monies so this shit is shilled 24/7 anyway.

>>106142025 (OP)
Because they're not actually learning like the human brain. In fact, they're not learning at all. It's just a silly mathematical model that predicts what the next set of numbers is going to be, based on what the previous numbers were. In other words, you are just telling it to spit out nonsense. LLMs don't learn nor think, they only predict the next numbers. Watch this talk by Andrej Karpathy, he explains it very well:
https://www.youtube.com/watch?v=kCc8FmEb1nY

>>106142025 (OP)
LLMs are basically pic related

>>106142443
And you know that's not how the human brain works how? Thank you for your ted talk.

>>106142025 (OP)
LLMs are basically the brute force way of doing AI. They train them by giving them a billion examples of something until they magically "get it", instead of teaching them basic rules that define that something with much less effort. The fact that they are still able to learn and develop a rudimentary understanding this way is actually a miracle in itself.
The only true innovation of the transformer architecture is the ability to interpret context within a vast dataset. This is far from the being a complete model of the human brain, but it's been significant enough to enable the current wave of generative AI capabilities.

>>106142530
>>106142443
Current models can to some extent do lot of these simple tasks they were not trained on, for example look at the pelican riding a bicycle test with code text. The AIs do learn basic concepts and logic, they just learn it slowly, very inefficiently and it is very hard to tell sometimes when it shows reasoning and logic and when it is just remembering learning data and spitting it back at you. You have to be delusional if you think it can ONLY spit data at you back, so it could be really good if we just made models that can do logic without understanding the world and then go on to teach them about the world.

>>106142025 (OP)
You're mistaken assumption is that a child cannot *eventually* learn *anything* by throwing PhD level physics textbooks at them.

>>106142530
The Chinese room experiment is a retarded sleight of hand trick where the description of the algorithm as "a book" is meant to bias your intuition incorrectly. If you describe the actual scale such a setup would require (multiple warehouses full of instructions and decades to compute a response) it no longer seems nearly as clever an argument.

Academic philosophers are midwits and you should treat them with nothing but contempt.

BREAKING: Thousands of AI researchers with decades of experience in the field utterly btfo by anons making crude analogies of AI

>>106142814
Ok anon but if you abstract the academic masturbation away and just think about a blackbox, it still makes sense

>>106142841
Reminder that the thing that made Language models go from making barely coherent sentences to replacing virtually all copywriters overnight was one Russian IT guy saying there should be more virtual neurons stacked on top of eachother rather then next to eachother.

several dangerous assumptions
1 - current education system is good at producing intelligence (or at least that it's better at producing intelligence than throwing a book at someone and telling them to just read)
2 - once the book is dumped on the AI, its reinforcement process isn't naturally optimizing itself based on rulesets
3 - the basic rules ("less advanced) -> advanced rules (phd level, etc) pipeline holds true (it doesn't, highschool physics models regularly treat electrons as a particle and newtonian physics as law, even basic concepts like friction resistive is sold as a basic coefficient * downward force which every engineering course needs to correct as soon as the freshmen get in for 101)

>>106142901
>there should be more virtual neurons stacked on top of eachother rather then next to eachother
That was just a matter of time after they had enough computing power, but already way better than saying AI is like a child or like a chinese room. Can you say how you think neurons should actually be stacked in a Transformer model?

>>106142734
Again, it has no underlying way of understanding the data. It only regurgitates it back. Any kind of semblance of logic is just pattern recognition. You can basically show that it's flawed "understanding" falls apart when you slightly modify popular word or math problems that it gets right before.

>>106142880
If you treat it as a black box it's entirely unsurprising that a room large enough to hold a chink could communicate in Chinese. The entire thought experiment hinges on white box understanding of what's going on inside plus an argument from incredulity of "hey wow if you hilariously underestimate the complexity of the algorithm isn't it surprising that it could speak a human language?"

>>106142025 (OP)
You are wrongly assuming that phd-level physics is harder than basic spatial reasoning and deduction. It's not, it's just sparser in the training data.

>>106142671
>They train them by giving them a billion examples of something until they magically "get it"
Not magically practically is neural-compression of ALL examples
that's why took them so much to make them be able to respect pemdas.
>>106142025 (OP)
Current LLMs should be called Large Language retrival Generative models
I am starting to think that at this point companies are aware and desperate, and they know they are only scaling knowledge but not capabilities.
We already know this because a very tiny model fuck it, I wonder why it haven destroy the LLM bubble yet.
https://github.com/sapientinc/HRM

>>106142970
>You can basically show that it's flawed "understanding" falls apart when you slightly modify popular word or math problems that it gets right before.
Have you actually tried? I doubt you can come up with a single puzzle that would trick the biggest AIs. You just remember what worked on chatgpt 2 years ago and think it's still the same

>>106143033
no current LLM can solve a large sudoku or complex puzzle without cheating retrieving the code.
If you think they do you are just AI illiterate. that's why HRM was created.

>>106143081
Sudoku is a pretty long term task though, if they could do so many steps we'd already have agents. But there's no longer any puzzle that they can't understand. I remember the threads that AI couldn't figure out this 3 digit lock code or whatever. We don't have these threads anymore.

>>106142354
It doesn't omit the part that handles constraints, you can give the new multimodal models a ruleset and ask it to generate something within that ruleset, and it will follow it.

>>106142993
gosh, why is that?

>>106142959
Processing power was scaling for long time and in very predictable ways, plus lot of today's models can be trained on home computers. LLMs basically jumped from nothing to something with the GPT series and image generators with Dall-E. There were genuine innovations happening, its not all just computer architecture models from 80s being finally given NVIDIA RTX GPUs and suddenly jumping from nothing into something.

>>106143081
Yeah, it fucking cant because they tell it to read the whole internet and spontaneously develop understanding of how board games work, and also be able to play them when given recording and images of individual turns. This is why I think we should teach it special awareness and basic logic before throwing random reddit forums and Wikipedia at the models during the training.

>>106143110
Closed source AI cheats with programmatic methods under the hood for basic tasks like counting. That's why you're not seeing threads for basic stuff like this anymore. But you can throw the same old challenges at open source stuff and watch it fail miserably. Here's Kimi K2 with counting. There were also a shit ton of logical puzzles (which no doubt probably already made it into the AI training set) that you could fool GPT 4 just by changing them slightly even after ClosedAI had updated it to give correct answers. Popular one is the one with the man, son and surgery, and there's probably many more variations. When you get to complex coding problems, that's where you really notice it start to break apart with just slight changes.

>>106143110
>Sudoku is a pretty long term task though, if they could do so many steps we'd already have agents. But there's no longer any puzzle that they can't understand.
as I said you have no idea what are you talking about, those "agents" are just pure control loops "chained" using "LLM" orchestrators.
pic-related are the results of SOTA models being released last month and aren't LLMs.
The issue is that not everybody is working on real AI applications to be aware that LLMs are incapable of understanding or solving complex puzzles by itself without retrieving code based solvers

>>106142993
Physics is really not that hard when you have all the knowledge in the world. There is rarely a huge logic leap, if you understand the intuitions and the equations it's usually enough.

AI models can generate code to solve complex problems that the core LLM struggles with, but that doesn't count as intelligence because...because it just doesn't, okay!

>>106143202
>Closed source AI cheats with programmatic methods under the hood for basic tasks like counting.
yea, but he won't understand that because probably had never run a local "agent"

>>106143251
>because it just doesn't, okay!
is like saying that A* understands why it chooses the shortest path.
We are not arguing if is AI or not, we are arguing that is not understanding the problem.
If that were true would be able to implement code of novel papers which is not even the case.
Is just feeding the bubble from public repositories I know this because the shit one recommend me one of my own repos.

>>106142025 (OP)
A child is exposed to the world world of noise. AI is exposed to the entire world of noise. Through artificial guidance or through guidance-less mechanism, children/ai learn to recognize patterns. This phase of AI was done early on. Now we know the mechanism on how to get AI to go from 0 -> PhD level. We dont need to go back to 0->Baby -> Teen -> College -> Grad anymore. That phase is done. Its now 0-> Grad+

>>106143496
>A child is exposed to the world world of noise. AI is exposed to the entire world of noise.
this is a lie, not all models starts learning from noise, we wish sadly is even close to the mechanism of how neurons works.
>Now we know
>we
kek.

>>106142841
AI researchers BTFO

>>106143551
Everything is noise until it has enough data. Thats true of babies, its true of ai.

>>106143636
>Everything is noise until it has enough data.
Stop making false assumptions. That's a lie,
read a book about information theory and search for https://en.wikipedia.org/wiki/Entropy_(information_theory)
Not all models works as de-noisers and in the domain of NLP/NLU your analogy based mentality doesn't work. If by noise you mean the first random weights that's not noise is the first initial guess of the semantic relations and is not how human brain works.
A 4yo child consumes more information in the visual cortex that the whole information trained on SOTA LLMS at the moment and is stuck It wont scale reasoning until we change the kind of models.

>>106143688
The first semantically coherent output requires lots of data. Otherwise, its all junk noise with random luck. A baby/ai repeating a word has no meaning, but a baby that can form chain reasoning endlessly is intelligence.

>>106143496
Yeah, except we aren't. O-3 or something was as good as graduates on most tasks, yet in almost all jobs it didn't replace graduates, why? Because it did well only on graduate tests, but not on all abilities of graduates. Humans are all roughly equally good at basic human tasks that even children can do. People with low IQ can struggle with them, but they are still easy, so when we test graduates we only test them on things that average human cant do but graduates can. But current AI models cant do what baby or teen can do, they can only do the graduate stuff, which is mostly just memorization and then using that memorized knowledge with basic human critical thinking to do the work. But it does not do that, AI cant do that basic stuff and it still hallucinates like crazy. The so called AGI benchmarks are stuff that 9 year old kids could do, so there is clearly a need to make AIs as capable as elementary school students and not just give them phd level collage tests to see if they score well. Nobody is paying people to score well on tests, so doing well on collage tests alone has no practical merit.

>>106143744
>The first semantically coherent output requires lots of data.
this is how I know you have 0 clue about the topic.
You can literally train an small model on an small set of language and get semantically coherent results.
And is unrelated to scaled reasoning, as I said there is no utility on making false statements.

>>106142025 (OP)
>We give them the whole internet or just wikipedia and bunch of high quality books and expect it to generalize.
That's actually not what people do. They give them a bunch of data and then try to reinforce out all the bad behavior. The hope is that the good responses will be all that's left, like a trained dog. The problem is they don't really learn a model of the world, they only learn how to predict tokens. Get it to imitate a phd enough and have phds give it a thumb up during training, and it will say things that a phd really would sometimes when prompted appropriately. But they don't "know" anything.
The only bubble is with this family of LLMs and the sunk costs of the affluenza addled Sillicon Valley greedheads. These problems are slowly getting solved in the research but they don't want to stop to understand. They just want to shove money into every orifice.

>>106143800
>small

>>106142530
If LLMs behaved like the Chinese Room it would be more useful in many cases because it would be deterministic. In practice there's a probability the guy puts it in bucket 1 and a probability he puts it in bucket 2 and you can never force or instruct him to put it in the one you want.

>>106142880
The problem isn't that its a black box or that its got a homunculi trapped inside the model doing all the "real" work. Its just that there's no homunculi or central control node, its just a large network. Similar to a slime mold intelligence. There's no central intelligence hidden inside the network.

>>106142901
What is this referring to?

>>106143864
Thing is the same is true of the human brain. It's true that there seems to be an executive cortex that functions as a decision agent, but it's still a stochastic system. The key difference is that the brain is trying to balance a bunch of needs and motives, and they way it does so is by behaving intelligently or at least effectively. What keeps it on task is this balancing act. It's not just completing sequences.

>>106144091
Yeah, the "mixture of needs" is prob what drives human's effective intelligence with the different needs converging together to form a general intelligence

>>106143836
isn't the raw output of the model deterministic tho? it just outputs the probabilities alongside the tokens. I'm sure all the online services just pick tokens weighted-randomly for variation, but the model itself should always output the same tokens and same probabilities for the same input.

>>106142025 (OP)
Generalization is bloat. The best AI tools are the ones that do one thing (e.g. chess engines that just play chess and nothing else).
>AI transformer architecture is like a human brain
Human brains have specialized areas. Each task is done by a specific area except in rare cases where brain damage forces it into another area (e.g. you suffer damage to Broca's area so you re-learn speech using the bit of the brain normally reserved for singing). Each neural structure was fine-tuned by hundreds of millions of years of natural selection.
If we assume AGI will develop like a brain (it may not, but it's a good guess), then attempting to generalize is a waste of resources. Instead, you should develop lots of highly specialized tool models that can be operated by agentic models. Each agent can then have a toolkit relevant to its job and never has to learn anything else. Coordination should be done by a higher order agent that directs the tool-using agents without ever needing to understand how they work. At the top of the hierarchy you'll have an executive agent to manage the big picture.

>>106144248
>outputs the probabilities alongside the tokens.
When I think probability I think indeterminism, even if the output of the probability function is a reliable consequence of its inputs.
This gets at the heart of why LLMs aren't perfectly reliable. The kind of determinism I'm thinking of is logical inevitability, if A then B type shit. They don't emulate deductive rules. They are just a correlation engine.

I don't think the human brain emulates deductive logic either, but we can "discipline ourselves" to do so, and in doing so we in a way "surrender ourselves" to the deductive inevitability by following out where the consequences lead. This ability to emulate logic and not just correlate statistics is what sets us apart. We don't just imitate training examples of reasoning, we somehow "obey the symbols"

>>106144445
>I don't think the human brain emulates deductive logic either,
oops I contradict myself here later, I don't think "emulate" is the right word here but I can't think of one right now

The llms we get aren't the one the glowies and military have got. You can't just release an unrestricted llm on the people partly because you nerds all try to turn them into hitler and stuff. Everything we get told about them is possibly misdirection.

It was on a good track like 10-15 years ago with all the vidya and board game stuff. Then geepeeters at OAI did their high-verbal stuff on the field, investors and normies.
It seems to be retvrning though with people like Davis Silver (OG alphago guy) Sutton and Carmack trying to steer it back to agents learning the super basic stuff on up, for themselves, from scratch, from their pure experience of interacting with the world they are in.

>>106142232
Doing the experiment would require (at least) hundreds of dollars worth of compute

>>106142025 (OP)
i feel like my chatgpt, because it has a memory of discussions, has gotten better on solving problems, by virtue of being walked thru solving many different problems, and now is more robust than it was before. i think you're right, ai has this amazing access to the encyclopedia of human output, but it has the experience of a 0 year old baby.

>>106144566
We teach thousands of years worth of cumulative human knowledge to people. The bits we don't teach are hardwired into the DNA.
There is very little you can learn just from interacting with the environment in a reasonable amount of time without some very strong priors.

>>106142025 (OP)
It took billions of years of evolution for our brain to learn how to learn efficiently. When we grow up we are getting bombarded with billions of training tokens every day, yet our brain is inherently capable of making sense of that torrent of data and discriminate between useful and useless information. We have nothing that is even remotely as good as the underlying software of the brain so we have to use the brute force approach and hope that true intelligence is an emergent property that just appears after a certain amount of compute.

>>106142443
midwit post