The pioneering computer scientist on why Trump is using ‘superintelligence’ wrong, whether AI could wipe out humanity – and how our biggest challenge will be accepting we’re not the only intelligent beings
you might be thinking you’re pulling a Socrates here, but you began sounding like a annoying 5-year-old that keeps repeating “what’s this, what’s that, why”
The capacity of mind, especially to understand principles, truths, facts or meanings, acquire knowledge, and apply it to practice; the ability to comprehend and learn; the ability to process sentient experience to generate true beliefs with a justified degree of confidence.
Breaking this down:
Understand principles, truths - Sure, sounds like an LLM to me. If you ask it to define a scientific principle or fact or Newton’s laws, it can tell you what it is.
facts or meaning - And, depending on the kind of thing we’re talking about, it can you tell you why it exists.
acquire knowledge - While it’s not permanent (because context are just thrown away like short-term memories), LLMs have MCPs and other tools to acquire new knowledge.
apply it in practice - It takes the knowledge it has naturally (training data), acquires new knowledge from its tools, and then applies it to the task it’s looking at.
the ability to comprehend and learn - Here we go with “comprehension” again. But, let’s assume we all know what “learn” is. A LLM can learn about things, but the main problem is the context size and how we throw that away each session. But, that can be solved by storing facts elsewhere, Markdown rules, other techniques to condense the immediate context into long-term storage. In humans, we call that “long-term memory”. We have our own process to stash short-term memory into long-term memory. We just haven’t experimented enough with the long-term storage part in LLMs.
the ability to process sentient experience - sentient… well, that’s a pretty weighted word. I also feel like there’s a clear difference between intelligence and sentience. We’re now entering the realm of sensation, thought, and feelings, which are entire different parts of the brain.
I also don’t think “sensations” are impossible to emulate, either.
I’m pretty sure comprehension was like the original success of LLMs, and the responses were a bit wonky. When was the last time an LLM failed to comprehend?
LLM’s comprehend statistical correlations. If you feed it every chess game, it can tell you which spaces the knight is likely to move to, and which it very unlikely to move to, but it does not understand how the knight moves.
You might say that’s a distinction without a difference, but programs can actually know thing, like a chess program actually does know how a knight moves, and it does so with minimal electricity.
Large neural networks have been shown to build parallel models of the world that they learn through their training data. A chess model trained only on chess move notation does have an interior board and state that it tracks, despite never being told about a board, the number of pieces, how they move, etc.
LLMs do something similar with language, mapping tokens to ideas in ways that can produce meaningful and grammatically correct output without needing to be formally taught any of this.
Our brains also do things like this - we observe X and then try to intuit Y, the truth of the world and reason for X.
Where this breaks down is humans can then take our guess for Y and propose Z, and try to find it out in the world. If we can’t, then Y is probably wrong, and our new understanding of the world is Y* that is consistent both with X and the absence of Z.
I cannot stress enough how far we are from machines that can do this. They can learn to approximate Y through training, but they have no means to extend that to propose an unobserved Z, no means to update their internal state given the results of that observation (without fully retraining the model on new data - which makes it a new model), and no way to question X given Z - they can talk at great length about data biases but can literally never excise bad data from their training weights. There have been experiments with training models on time-boxed datasets (e.g. stuff up to AD 1880) and asking it about modern concepts - predictably hilarious results.
All the current ‘AI’ stuff is smoke and mirrors around this very hard problem. The machines can talk, but they can’t think (Z) and can’t learn (Y*).
they can talk at great length about data biases but can literally never excise bad data from their training weights
And you can’t excise 20 years of Fox News from some boomer conservative. They can literally never back down from their biases.
Training data is a lot more permanent than context, but at the same time, it’s not absolute, either. I’ve seen plenty of thinking blocks where it knows about X from its training data, but perhaps it’s out-of-date and should check other sources. It has the web tools, finds newer information, processes it as part of its problem set, and solves the problem with newer information.
The core issue is we throw all of this data away when we’re done with the problem! A LLM only has so much context, so it starts to compact after a while. And even if it doesn’t, we voluntarily chuck that shit out the window, as soon as the session ends.
It’s like Lenny from Memento. You got a baseline person who constantly believes he’s got to find his wife’s killer. And even when he tries to supplant his lack of long-term memory with photos and tattoos, it’s a shit system with shittier outcomes. Also, he doesn’t get to choose when his memory disappears. It just happens with some kind of stress trigger. Because of how flawed Lemmy is, do we call him somebody who can’t think and can’t learn? Is he intelligent? Is he even sentient?
Though, in the case of an LLM, it doesn’t have to be set up this way. We can still stash a long-term version of memories as rules and other permanent files. In fact, this is encouraged. Everything starts with a system prompt, a set of rules, and the user prompt. All of that is adjustable.
And yes, you can train out biases. Rather easily, actually. You don’t even have to adjust weights. You post-train through rules files.
But, in terms of philosophical questions, can we call LLMs intelligent if we can edit their memories at a whim? Is a brainwashed boomer intelligent because Fox News hacked their memories and biases? Is Lenny intelligent because he can’t form new memories?
You have to remember that just about every argument for or against LLMs can also apply to humans, and you can’t really arrive at definitions without clearly separating what doesn’t apply to LLMs and does apply to humans, in all cases.
And yes, you can train out biases. Rather easily, actually. You don’t even have to adjust weights. You post-train through rules files.
The assertion is not about whether ‘you’ can do it. It is that the model cannot do this itself.
A human recognizing a bias in the model and correcting it is not evidence robot intelligence, it is further evidence for human intelligence.
Training is absolute, it determines the set of relationships the model can represent between input and output. Context, rules files, system prompts, etc. operate only on the conditional probabilities within this set and are impermanent. This is ‘priming’ in human psychology, and why a person in a herpetology class will talk about Pythonic features differently than in a coding class.
As an example of how this is not real learning and is inadequate to solve real problems, here is a paper where the authors challenged multiple models to solve unfamiliar games where they need to discover the rules through play and feedback. Even the ‘best’ model performed pitifully compared to humans.
If you feed it every chess game, it can tell you which spaces the knight is likely to move to, and which it very unlikely to move to, but it does not understand how the knight moves.
Well, it knows how the knight moves, but it might forget that’s it’s an important rule after it starts to lose context. That’s also how LLMs start to lose track of what’s actually on the board.
There are definitely differences between LLMs and Neural Networks, but at the end of the day, they are both just gigabytes of weights. And really, both types are similar to what the human brain possesses. People tend to forget that the human brain isn’t just some monolithic thing. There’s a lot of difference specialized pieces. If you ask Magnus Carlsen how he calculated the best move on the chess board in a blitz game, he literally cannot tell you! It just appears, because that part of the brain is trained and hardwired to calculate good moves, based on the current board state.
A LLM uses logic and reasoning to arrive at a solution. That’s not really compatible with intense pre-trained calculation. You can’t talk yourself into a chess move at the speed the game requires. You can quickly think about possibilities, but you have to already have pre-trained calculations to get at least 90-95% of the way there, to pick from a few different lines.
On the other hand, Neural Networks beat the shit out of the greatest grandmasters in chess. Hands down. On average, they are at least +500 ELO to the best GMs. And everybody in the chess world knows this, which is why everybody uses these tools outside of 1v1 human to human games. (Obviously, cheating is taken very seriously, and has made plenty of headlines over the years.) Every GM analyzes their own games using those engines, to pick apart mistakes.
They never comprehend. They ingest information and find close approximations for responses, but they don’t comprehend any part of the process. If they did, they wouldn’t hallucinate complete bullshit answers and spit them out like they were fact
To be honest human spew bullshit all the time when they do not understand something. Or people are confidently incorrect all the time as they are remembering incorrectly. The amount of bullshitting is more a meter for intelligence, not something that discriminate between intelligence or the lack thereof.
They have not comprehended anything to date, as they don’t actually understand anything, or have an intellect with which to do so. They are probabilistic next word guessing machines, that is what is happening behind the smoke and mirrors, not “thought”.
pretty sure it comes from latin “inter-lego” (collect or pick out information, like you would pick out the good fruit from a bush with good and bad fruit).
so it’s about filtering what you want to have from what you don’t want to have. i.e. a mathematical function f : X -> R which maps an object x to a real number to figure out how desirable that object is, i.e. assign value to it.
technically you don’t even need neural networks for it, sorting algorithms employ this kind of function all the time to sort objects according to some criteria.
Define intelligence.
Comprehension, something LLMs are intrinsically incapable of demonstrating, let alone possessing.
you are just replacing one underspecified concept with another. what’s “comprehension”?
Literal dictionary definitions:
Comprehension: the ability to understand something.
Understand: perceive the intended meaning of something.
Things LLMs are categorically incapable of embodying, and anyone who believes otherwise is demonstrating an innate challenge of grasping themselves.
what’s “perceive”?
Alright Mr Peterson, calm down.
What’s “what”?
what?
Hmmm, what?
you might be thinking you’re pulling a Socrates here, but you began sounding like a annoying 5-year-old that keeps repeating “what’s this, what’s that, why”
people of ancient Greek found Socrates so annoying they sentenced him to death by poison. so… thanks, i guess?
It’s what you don’t posses about the workings of LLMs (or basic English).
https://www.merriam-webster.com/dictionary/comprehension
Jesus fuck… did you actually read your definition?
Okay, we’ve replaced the concept of “intelligence” and just swapped it out with “grasping with the intellect”.
Nice, using the verb word of “comprehending” to define “comprehension”.
Okay, maybe that’s a start, but it’s incredibly vague.
Also, clicking on the links just resolves back to the most generic circular definition possible. Remind me to never use Merriam-Webster.
At least this gets further: https://en.wiktionary.org/wiki/comprehension
If we settle on “A thorough understanding” and define understanding as… fuck, we’re back to this circular bullshit again.
Okay, fine, let’s just look at intelligence:
Breaking this down:
I also don’t think “sensations” are impossible to emulate, either.
Words have meanings, and you don’t get to change them in order to strengthen a weak argument.
I’m sorry that those meanings aren’t convenient for your untenable position.
anyone can quote a dictionary without comprehension.
I didn’t quote it, I sent you the entry in the hopes that you might read it yourself and learn something
I’m pretty sure comprehension was like the original success of LLMs, and the responses were a bit wonky. When was the last time an LLM failed to comprehend?
LLM’s comprehend statistical correlations. If you feed it every chess game, it can tell you which spaces the knight is likely to move to, and which it very unlikely to move to, but it does not understand how the knight moves.
You might say that’s a distinction without a difference, but programs can actually know thing, like a chess program actually does know how a knight moves, and it does so with minimal electricity.
There’s some semantics here.
Large neural networks have been shown to build parallel models of the world that they learn through their training data. A chess model trained only on chess move notation does have an interior board and state that it tracks, despite never being told about a board, the number of pieces, how they move, etc.
LLMs do something similar with language, mapping tokens to ideas in ways that can produce meaningful and grammatically correct output without needing to be formally taught any of this.
Our brains also do things like this - we observe X and then try to intuit Y, the truth of the world and reason for X.
Where this breaks down is humans can then take our guess for Y and propose Z, and try to find it out in the world. If we can’t, then Y is probably wrong, and our new understanding of the world is Y* that is consistent both with X and the absence of Z.
I cannot stress enough how far we are from machines that can do this. They can learn to approximate Y through training, but they have no means to extend that to propose an unobserved Z, no means to update their internal state given the results of that observation (without fully retraining the model on new data - which makes it a new model), and no way to question X given Z - they can talk at great length about data biases but can literally never excise bad data from their training weights. There have been experiments with training models on time-boxed datasets (e.g. stuff up to AD 1880) and asking it about modern concepts - predictably hilarious results.
All the current ‘AI’ stuff is smoke and mirrors around this very hard problem. The machines can talk, but they can’t think (Z) and can’t learn (Y*).
And you can’t excise 20 years of Fox News from some boomer conservative. They can literally never back down from their biases.
Training data is a lot more permanent than context, but at the same time, it’s not absolute, either. I’ve seen plenty of thinking blocks where it knows about X from its training data, but perhaps it’s out-of-date and should check other sources. It has the web tools, finds newer information, processes it as part of its problem set, and solves the problem with newer information.
The core issue is we throw all of this data away when we’re done with the problem! A LLM only has so much context, so it starts to compact after a while. And even if it doesn’t, we voluntarily chuck that shit out the window, as soon as the session ends.
It’s like Lenny from Memento. You got a baseline person who constantly believes he’s got to find his wife’s killer. And even when he tries to supplant his lack of long-term memory with photos and tattoos, it’s a shit system with shittier outcomes. Also, he doesn’t get to choose when his memory disappears. It just happens with some kind of stress trigger. Because of how flawed Lemmy is, do we call him somebody who can’t think and can’t learn? Is he intelligent? Is he even sentient?
Though, in the case of an LLM, it doesn’t have to be set up this way. We can still stash a long-term version of memories as rules and other permanent files. In fact, this is encouraged. Everything starts with a system prompt, a set of rules, and the user prompt. All of that is adjustable.
And yes, you can train out biases. Rather easily, actually. You don’t even have to adjust weights. You post-train through rules files.
But, in terms of philosophical questions, can we call LLMs intelligent if we can edit their memories at a whim? Is a brainwashed boomer intelligent because Fox News hacked their memories and biases? Is Lenny intelligent because he can’t form new memories?
You have to remember that just about every argument for or against LLMs can also apply to humans, and you can’t really arrive at definitions without clearly separating what doesn’t apply to LLMs and does apply to humans, in all cases.
The assertion is not about whether ‘you’ can do it. It is that the model cannot do this itself.
A human recognizing a bias in the model and correcting it is not evidence robot intelligence, it is further evidence for human intelligence.
Training is absolute, it determines the set of relationships the model can represent between input and output. Context, rules files, system prompts, etc. operate only on the conditional probabilities within this set and are impermanent. This is ‘priming’ in human psychology, and why a person in a herpetology class will talk about Pythonic features differently than in a coding class.
As an example of how this is not real learning and is inadequate to solve real problems, here is a paper where the authors challenged multiple models to solve unfamiliar games where they need to discover the rules through play and feedback. Even the ‘best’ model performed pitifully compared to humans.
Well, it knows how the knight moves, but it might forget that’s it’s an important rule after it starts to lose context. That’s also how LLMs start to lose track of what’s actually on the board.
There are definitely differences between LLMs and Neural Networks, but at the end of the day, they are both just gigabytes of weights. And really, both types are similar to what the human brain possesses. People tend to forget that the human brain isn’t just some monolithic thing. There’s a lot of difference specialized pieces. If you ask Magnus Carlsen how he calculated the best move on the chess board in a blitz game, he literally cannot tell you! It just appears, because that part of the brain is trained and hardwired to calculate good moves, based on the current board state.
A LLM uses logic and reasoning to arrive at a solution. That’s not really compatible with intense pre-trained calculation. You can’t talk yourself into a chess move at the speed the game requires. You can quickly think about possibilities, but you have to already have pre-trained calculations to get at least 90-95% of the way there, to pick from a few different lines.
On the other hand, Neural Networks beat the shit out of the greatest grandmasters in chess. Hands down. On average, they are at least +500 ELO to the best GMs. And everybody in the chess world knows this, which is why everybody uses these tools outside of 1v1 human to human games. (Obviously, cheating is taken very seriously, and has made plenty of headlines over the years.) Every GM analyzes their own games using those engines, to pick apart mistakes.
They never comprehend. They ingest information and find close approximations for responses, but they don’t comprehend any part of the process. If they did, they wouldn’t hallucinate complete bullshit answers and spit them out like they were fact
Oh, so, like my coworkers then.
/s
Everyone use big word. U no big word. Small words funny. Upvote
To be honest human spew bullshit all the time when they do not understand something. Or people are confidently incorrect all the time as they are remembering incorrectly. The amount of bullshitting is more a meter for intelligence, not something that discriminate between intelligence or the lack thereof.
This gives me “are you smarter than a fifth grader” vibes
They have not comprehended anything to date, as they don’t actually understand anything, or have an intellect with which to do so. They are probabilistic next word guessing machines, that is what is happening behind the smoke and mirrors, not “thought”.
You’re pretty incorrect: LLMs don’t comprehend, period.
pretty sure it comes from latin “inter-lego” (collect or pick out information, like you would pick out the good fruit from a bush with good and bad fruit).
https://en.wiktionary.org/wiki/intelligent
so it’s about filtering what you want to have from what you don’t want to have. i.e. a mathematical function f : X -> R which maps an object x to a real number to figure out how desirable that object is, i.e. assign value to it.
technically you don’t even need neural networks for it, sorting algorithms employ this kind of function all the time to sort objects according to some criteria.