The worst thing about 'AI'
video description
Date: 2025-06-18
Related videos
Comments and reviews: 20
Monoffel
Hello, self-proclaimed Adam Ragusea fan here (not the one who sent that email. I'm also a fan of AI. I also work with AI sometimes (more than just prompting it. We all know there is no shortage of AI bs out there, and I can't even claim that AI is a net benefit to humanity. But I have to point out here that they are not glorified auto complete. This is a common argument, that does not hold any real substance.
You bolster this view, by stating that you are pretty sure that LLMs statistically generate the next token, and that you are pretty sure that this what makes the core of an LLM. (I'm just paraphrasing. You then add some conclusions, like that LLMs cannot reason, and they cannot lie (the implication being that they don't even qualify for the level of awareness necessary to be able to lie.
While technically true on some level, that they statistically generate tokens, this mechanism cannot be used to explain that they are worse than human brains in their abilities. It is neither the case that we can prove that humans don't do this (except on some technical level that doesn't change the conclusions, nor that this mechanism is inferior to human intelligence or consciousness.
For example, ask a very difficult yes/no question. How does a human answer this They may do a large number of intermediate computations (analagous to outputting tokens) in their head, before staying either yes or no. But ultimately, that answer is also statistical. They did the math, and they responded with the most likely answer. So in other words, that binary choice is a statistical answer. As an LLM opponent, you may now try to go retreat to the domain of math, where in some sense statistics is eliminated from the answers. But even there, we neither know if we really exist, on a philosophical level, and we always work with axioms in math as well, which is to say that statistics cannot be meaningfully escaped from. The human brain hallucinates reality, to the extent that we can never be sure if our existence came to be just a moment ago, with all your memories synthetically inserted. What we really are is brains in a dark, lightless jar, with some sensory inputs piped into it, and some motor outputs coming out of it.
Anyway, if you really want to cling to math, then this argument gets challenging, but certainly in most fields, humans can be brought into situations where they statistically output an answer. Any kind of multiple choice scenario, for example. You could argue that outside such scenarios, we do not statistically generate tokens/thoughts. But what I'm saying is it doesn't matter what you call it, our thoughts can be conceptually represented as tokens. Even in large X models, tokens can be text, images, sound, and even bytes. Tokens are not inherently text, and so we can conceptually also label what our brain does in its thought process as tokens. You may want to cling to the sequentiality of tokens, as a defense for the uniqueness of the human brain, saying But our thoughts are not one token after the other, it's a lot more complex than that. But you can make an LLM do the same thing. For example, prompt your question to an LLM multiple times at once, and then aggregate those outputs back into a single one at the end. Now you have some sort of parallelism, like you might have in a human brain.
If you're interested in the differences and similarities of biological and artificial neural networks, Geoffrey Hinton is a very smart person who dedicates his life to this topic. You can find arguments from him about the learning process, i. e. back-propagation, and whether our brains conceptually do this, for example. (Though his conclusion is utlimately that it doesn't really matter, because backprop is equal or better to the human learning process in its capabilities)
To summarise, just because the process is statistical in nature, it does not follow that it's dumb or not conscious, or inferior to human thinking. I'm not disagreeing that the best current LLMs are dumb and I wouldn't call them all that conscious either, but you can't categorically say that they can't ever be intelligent due to the statistical nature.
Moreover, even today, you can do so much more with an LLM, than what this fan mailed to you with excitement. This may not be obvious to you, if you've never programmatically used LLMs before. But for example, take a modern reasoning model. Whether the name reasoning is qualifies for reasoning doesn't matter. But what matters is that they have a mechanism to revise mistakes. A scratchpad of ideas, that they can use, before coming up with an answer. To a fault, this scratchpad is still sequential in nature, but it's simple enough to remove this limitation. And even with it being sequential, I think a lot of people would be surprised what you can do with this. The reasoning block will have thoughts like but wait, alternatively, where the model is able to explore different directions, before coming up with an answer. I can't emphasize enough how similar this is to human thinking, and how apparent it distances LLMs from the whole glorified auto complete paradigm. If the model can step through various ideas, rather than just saying a single yes/no answer, then what is the point of labeling it statistical (with the meaning statistical = dumb) This is exactly what a human would do. Sure, a human could rapid-fire a gut yes/no response, but most of the time, they would do the same exact process, of exploring different angles, and reasoning through them, before giving an answer.
Even if a human would to a gut response, what would this response be based on, if not the statistical life experience of that human Much like the LLM response would be the statistical response of what it learned.
If you wanted to create a better dam Ragusea video script with AI, here are some pointers:
instruct the AI to have a childlike or not-so-serious persona - this makes it more likely to admit mistakes, rather than doubling down
use a reasoning model for idea generation, and use a writing model to imitate Adam's style writing
supply examples of Adam's writing, at least initially, if not periodically - it may have ingested YT video subtitles in its training, but even then the statistical process often merges it into some sort of average
generate more tokens - a so called 0-shot prompt, where you ask a single question, and take that output as final, will always give you something that is short in length - the most obvious workaround to that is to just ask it to generate a part, then another part, and so on
ask it to create a rough draft of the script first, with its chapters, then have it refine that script over multiple prompts
then for each chapter, ask it to generate the content as a draft first, then refine it
since we want to be fair to the process, in the sense that the AI has to do most of the work by itself, and we want to just sit back and watch, make heavy use of self-reflection - ask the model to revise its prior outputs until sufficient
All of the above can be mostly accomplished in a linear fashion, but if you wanted to go real serious, you would program an agent framework:
you assign multiple agents, which you can think of as instances of the same LLM, but each with its own goal and conversation history
one agent ideates, one writes, one revises or proceeds, and several simply validate with different criteria (e. g. match the Adam style, match the humor, match the desired length, verify narrative consistency, etc)
these agents are then arranged in a loop to work together collaboratively
if you assign the roles and define the goals and criteria sufficiently, in theory this agent framework can now spend e. g. all night to come up with a script, without any human invervention
(Note that the complexity of agentic frameworks greatly exceeds that of basic prompting techniques, and would require one or multiple people who really know what they're doing. It's not really practical today, but all big companies are researching in this area)
Now if you wanted to take it further still, you would post-train a model on Adam's writing samples directly. This is even more technically challenging, since now you have to deal with training models, where a lot can go wrong, it's expensive, and you are also locking yourself down to a particular model. Up until now, you could always plug in the latest models. But if you ever successfully train a model, it will now be very specific in producing output in the Adam Ragusea style. (It will also likely have forgotten almost everything else. This now unburdens you from having to provide writing samples, which in turn gives you extra room in your context. Context is analogous to short-term memory, and it is also a limited quantity, that has to be considered at all times, in each conversation history. Another practical workaround here is to offer the model a permanent storage as long-term memory. But this would likely be overkill for this project.
If you've done all this, and the scripts you get still suck, well then you still can't say that LLMs are fundamentally incapable, but at least you would have explored what current models are capable of, to the full extent. All of which is to constrast this with the 2023 video script you've got, and how much more AI can do.
reply
Hello, self-proclaimed Adam Ragusea fan here (not the one who sent that email. I'm also a fan of AI. I also work with AI sometimes (more than just prompting it. We all know there is no shortage of AI bs out there, and I can't even claim that AI is a net benefit to humanity. But I have to point out here that they are not glorified auto complete. This is a common argument, that does not hold any real substance.
You bolster this view, by stating that you are pretty sure that LLMs statistically generate the next token, and that you are pretty sure that this what makes the core of an LLM. (I'm just paraphrasing. You then add some conclusions, like that LLMs cannot reason, and they cannot lie (the implication being that they don't even qualify for the level of awareness necessary to be able to lie.
While technically true on some level, that they statistically generate tokens, this mechanism cannot be used to explain that they are worse than human brains in their abilities. It is neither the case that we can prove that humans don't do this (except on some technical level that doesn't change the conclusions, nor that this mechanism is inferior to human intelligence or consciousness.
For example, ask a very difficult yes/no question. How does a human answer this They may do a large number of intermediate computations (analagous to outputting tokens) in their head, before staying either yes or no. But ultimately, that answer is also statistical. They did the math, and they responded with the most likely answer. So in other words, that binary choice is a statistical answer. As an LLM opponent, you may now try to go retreat to the domain of math, where in some sense statistics is eliminated from the answers. But even there, we neither know if we really exist, on a philosophical level, and we always work with axioms in math as well, which is to say that statistics cannot be meaningfully escaped from. The human brain hallucinates reality, to the extent that we can never be sure if our existence came to be just a moment ago, with all your memories synthetically inserted. What we really are is brains in a dark, lightless jar, with some sensory inputs piped into it, and some motor outputs coming out of it.
Anyway, if you really want to cling to math, then this argument gets challenging, but certainly in most fields, humans can be brought into situations where they statistically output an answer. Any kind of multiple choice scenario, for example. You could argue that outside such scenarios, we do not statistically generate tokens/thoughts. But what I'm saying is it doesn't matter what you call it, our thoughts can be conceptually represented as tokens. Even in large X models, tokens can be text, images, sound, and even bytes. Tokens are not inherently text, and so we can conceptually also label what our brain does in its thought process as tokens. You may want to cling to the sequentiality of tokens, as a defense for the uniqueness of the human brain, saying But our thoughts are not one token after the other, it's a lot more complex than that. But you can make an LLM do the same thing. For example, prompt your question to an LLM multiple times at once, and then aggregate those outputs back into a single one at the end. Now you have some sort of parallelism, like you might have in a human brain.
If you're interested in the differences and similarities of biological and artificial neural networks, Geoffrey Hinton is a very smart person who dedicates his life to this topic. You can find arguments from him about the learning process, i. e. back-propagation, and whether our brains conceptually do this, for example. (Though his conclusion is utlimately that it doesn't really matter, because backprop is equal or better to the human learning process in its capabilities)
To summarise, just because the process is statistical in nature, it does not follow that it's dumb or not conscious, or inferior to human thinking. I'm not disagreeing that the best current LLMs are dumb and I wouldn't call them all that conscious either, but you can't categorically say that they can't ever be intelligent due to the statistical nature.
Moreover, even today, you can do so much more with an LLM, than what this fan mailed to you with excitement. This may not be obvious to you, if you've never programmatically used LLMs before. But for example, take a modern reasoning model. Whether the name reasoning is qualifies for reasoning doesn't matter. But what matters is that they have a mechanism to revise mistakes. A scratchpad of ideas, that they can use, before coming up with an answer. To a fault, this scratchpad is still sequential in nature, but it's simple enough to remove this limitation. And even with it being sequential, I think a lot of people would be surprised what you can do with this. The reasoning block will have thoughts like but wait, alternatively, where the model is able to explore different directions, before coming up with an answer. I can't emphasize enough how similar this is to human thinking, and how apparent it distances LLMs from the whole glorified auto complete paradigm. If the model can step through various ideas, rather than just saying a single yes/no answer, then what is the point of labeling it statistical (with the meaning statistical = dumb) This is exactly what a human would do. Sure, a human could rapid-fire a gut yes/no response, but most of the time, they would do the same exact process, of exploring different angles, and reasoning through them, before giving an answer.
Even if a human would to a gut response, what would this response be based on, if not the statistical life experience of that human Much like the LLM response would be the statistical response of what it learned.
If you wanted to create a better dam Ragusea video script with AI, here are some pointers:
instruct the AI to have a childlike or not-so-serious persona - this makes it more likely to admit mistakes, rather than doubling down
use a reasoning model for idea generation, and use a writing model to imitate Adam's style writing
supply examples of Adam's writing, at least initially, if not periodically - it may have ingested YT video subtitles in its training, but even then the statistical process often merges it into some sort of average
generate more tokens - a so called 0-shot prompt, where you ask a single question, and take that output as final, will always give you something that is short in length - the most obvious workaround to that is to just ask it to generate a part, then another part, and so on
ask it to create a rough draft of the script first, with its chapters, then have it refine that script over multiple prompts
then for each chapter, ask it to generate the content as a draft first, then refine it
since we want to be fair to the process, in the sense that the AI has to do most of the work by itself, and we want to just sit back and watch, make heavy use of self-reflection - ask the model to revise its prior outputs until sufficient
All of the above can be mostly accomplished in a linear fashion, but if you wanted to go real serious, you would program an agent framework:
you assign multiple agents, which you can think of as instances of the same LLM, but each with its own goal and conversation history
one agent ideates, one writes, one revises or proceeds, and several simply validate with different criteria (e. g. match the Adam style, match the humor, match the desired length, verify narrative consistency, etc)
these agents are then arranged in a loop to work together collaboratively
if you assign the roles and define the goals and criteria sufficiently, in theory this agent framework can now spend e. g. all night to come up with a script, without any human invervention
(Note that the complexity of agentic frameworks greatly exceeds that of basic prompting techniques, and would require one or multiple people who really know what they're doing. It's not really practical today, but all big companies are researching in this area)
Now if you wanted to take it further still, you would post-train a model on Adam's writing samples directly. This is even more technically challenging, since now you have to deal with training models, where a lot can go wrong, it's expensive, and you are also locking yourself down to a particular model. Up until now, you could always plug in the latest models. But if you ever successfully train a model, it will now be very specific in producing output in the Adam Ragusea style. (It will also likely have forgotten almost everything else. This now unburdens you from having to provide writing samples, which in turn gives you extra room in your context. Context is analogous to short-term memory, and it is also a limited quantity, that has to be considered at all times, in each conversation history. Another practical workaround here is to offer the model a permanent storage as long-term memory. But this would likely be overkill for this project.
If you've done all this, and the scripts you get still suck, well then you still can't say that LLMs are fundamentally incapable, but at least you would have explored what current models are capable of, to the full extent. All of which is to constrast this with the 2023 video script you've got, and how much more AI can do.
reply
paul5849
GPT 4. 5 version of an Adam Ragusea carbonara video:
[Opening shot: Close-up of creamy carbonara being twirled around a fork. Upbeat music fades in. ]
ADAM:
Carbonarafour simple ingredients, but endless possibilities to screw it up. Let's make it right, together.
[Cut to ingredients displayed clearly on a countertop: spaghetti, guanciale (or pancetta/bacon, eggs, Pecorino Romano, and black pepper. ]
ADAM:
First up, ingredients: spaghetti, eggs, Pecorino Romano cheese, black pepper, and guancialeauthentic Italian pork cheek. Can’t find guanciale Pancetta or bacon will do just fine.
[Cut to Adam slicing guanciale into cubes. ]
ADAM:
Start by slicing your guanciale into chunky little cubes. Big enough to crisp up beautifully without losing their soul.
[Sound of sizzling as guanciale cooks. ]
ADAM:
Now, into a cold pan, gently render that fat until crispy but not burned. Medium heat and patience are your best friends.
[Cut to boiling pasta in salty water. ]
ADAM:
While that's going, spaghetti into aggressively salted boiling water. Cook just shy of al dente because it’ll finish in the sauce.
[Close-up of Adam whisking eggs and cheese. ]
ADAM:
Now, the magic mix: eggs and freshly grated Pecorino Romano. Whisk until creamylike a savory custard. Lots of black pepper, please; after all, carbonara means coal miner style.
[Shot of pasta transferred to pan with guanciale. ]
ADAM:
Transfer your pasta directly from the pot to your pan of guanciale goodness. Splash in pasta waterthat starchy elixir helps everything emulsify beautifully.
[Slow-motion pour of egg mixture into the pasta pan. ]
ADAM:
Now, critical moment: heat off, egg mix in, and stir vigorously. The residual heat creates creamy perfection. If too hot, it scrambles; if too cold, it’s soup.
[Close-up, glossy creamy carbonara in pan. ]
ADAM:
This is your ideal carbonarasilky, glossy, and luscious. Taste and adjust pepper or cheese if necessary. Trust yourself, it's your dinner.
[Adam plating pasta, finishing with a final dusting of Pecorino Romano and black pepper. ]
ADAM:
Carbonara done right. Simple ingredients, technique mastered.
[Closing shot: Adam taking a satisfying bite, nodding approvingly. ]
ADAM:
So good. Remember: cook smarter, not harder. Until next time, friends.
[Music swells, fade to black with end card. ]
reply
GPT 4. 5 version of an Adam Ragusea carbonara video:
[Opening shot: Close-up of creamy carbonara being twirled around a fork. Upbeat music fades in. ]
ADAM:
Carbonarafour simple ingredients, but endless possibilities to screw it up. Let's make it right, together.
[Cut to ingredients displayed clearly on a countertop: spaghetti, guanciale (or pancetta/bacon, eggs, Pecorino Romano, and black pepper. ]
ADAM:
First up, ingredients: spaghetti, eggs, Pecorino Romano cheese, black pepper, and guancialeauthentic Italian pork cheek. Can’t find guanciale Pancetta or bacon will do just fine.
[Cut to Adam slicing guanciale into cubes. ]
ADAM:
Start by slicing your guanciale into chunky little cubes. Big enough to crisp up beautifully without losing their soul.
[Sound of sizzling as guanciale cooks. ]
ADAM:
Now, into a cold pan, gently render that fat until crispy but not burned. Medium heat and patience are your best friends.
[Cut to boiling pasta in salty water. ]
ADAM:
While that's going, spaghetti into aggressively salted boiling water. Cook just shy of al dente because it’ll finish in the sauce.
[Close-up of Adam whisking eggs and cheese. ]
ADAM:
Now, the magic mix: eggs and freshly grated Pecorino Romano. Whisk until creamylike a savory custard. Lots of black pepper, please; after all, carbonara means coal miner style.
[Shot of pasta transferred to pan with guanciale. ]
ADAM:
Transfer your pasta directly from the pot to your pan of guanciale goodness. Splash in pasta waterthat starchy elixir helps everything emulsify beautifully.
[Slow-motion pour of egg mixture into the pasta pan. ]
ADAM:
Now, critical moment: heat off, egg mix in, and stir vigorously. The residual heat creates creamy perfection. If too hot, it scrambles; if too cold, it’s soup.
[Close-up, glossy creamy carbonara in pan. ]
ADAM:
This is your ideal carbonarasilky, glossy, and luscious. Taste and adjust pepper or cheese if necessary. Trust yourself, it's your dinner.
[Adam plating pasta, finishing with a final dusting of Pecorino Romano and black pepper. ]
ADAM:
Carbonara done right. Simple ingredients, technique mastered.
[Closing shot: Adam taking a satisfying bite, nodding approvingly. ]
ADAM:
So good. Remember: cook smarter, not harder. Until next time, friends.
[Music swells, fade to black with end card. ]
reply
akashgarg9776
This is off base There is a difference between saying, yeah, there doesn’t really exist an economic model for producing factual information (which is true without AI as well) When you do interviews, the professors you speak to are happy to volunteer their time because they get paid by the university, who gets paid by the taxpayer. Is that sustainable And I understand your perspective on the decline of the journalism industry.
And there is certainly room to debate what that economic model should look like. Ultimately though, you can’t copyright facts! And that has always been true, with or without AI. After all, it’s not like the existing model of some random person in the Philippines (or on Buzzfeed) repackaging and distributing NYT content for free is a good system either!
But if your argument is the benefits are not worth the costs, I just think that thats wrong. It’s not there are benefits, they absolutely justify the costs. o3 is an incredible model that operates at PhD level. Deep research can produce research papers. The number of hallucinations have gone down by quite a lot. You don’t find that useful in day to day life I’m a software engineer, it routinely answers complex questions that stump senior engineers. For cooking, you might find this useful, I take recipes and for example I need more protein and less carbs in my diet. I can ask that! I think any objective look into tradeoffs clearly shows it’s worth it, and I think you are basing this opinion off of bias and lack of actually exploring the better models. It’s like using Yahoo and then saying this search thing is not real.
reply
This is off base There is a difference between saying, yeah, there doesn’t really exist an economic model for producing factual information (which is true without AI as well) When you do interviews, the professors you speak to are happy to volunteer their time because they get paid by the university, who gets paid by the taxpayer. Is that sustainable And I understand your perspective on the decline of the journalism industry.
And there is certainly room to debate what that economic model should look like. Ultimately though, you can’t copyright facts! And that has always been true, with or without AI. After all, it’s not like the existing model of some random person in the Philippines (or on Buzzfeed) repackaging and distributing NYT content for free is a good system either!
But if your argument is the benefits are not worth the costs, I just think that thats wrong. It’s not there are benefits, they absolutely justify the costs. o3 is an incredible model that operates at PhD level. Deep research can produce research papers. The number of hallucinations have gone down by quite a lot. You don’t find that useful in day to day life I’m a software engineer, it routinely answers complex questions that stump senior engineers. For cooking, you might find this useful, I take recipes and for example I need more protein and less carbs in my diet. I can ask that! I think any objective look into tradeoffs clearly shows it’s worth it, and I think you are basing this opinion off of bias and lack of actually exploring the better models. It’s like using Yahoo and then saying this search thing is not real.
reply
aragusea
Something that bothers me in GenAI discussions is the overall lack of nuanced views on the subject. While I’m personally cautiously optimistic about GenAI, I feel like everyone speaking on the issue is either vehemently opposed to AI or naively defensive of it, so this video was really refreshing.
I don’t think GenAI is inherently bad, but we have to be cautious and stay cognizant of how it actually works and what it’s suitable for. It’s excellent at summarization, data analysis, repetitive tasks(coupled with RPA, e. g, and it’s very adaptable. While this has the potential to grow the economy in ways that benefit all of us, it also has the potential to drive automation and job displacement that industrial growth can’t keep up with. This is why I say I’m cautiously optimistic rather than totally pro-AI as some of the tech bros do. On the flip side, a lot of anti-AI absolutism seems to echo arguments for technological regression and degrowth, which I think are also dangerous concepts.
Overall, I don’t think it’s productive to be scared of AI. Like every other technology, its potential to both hurt or help society is dependent on how we humans apply it.
Thanks for the cool video Adam!
reply
Something that bothers me in GenAI discussions is the overall lack of nuanced views on the subject. While I’m personally cautiously optimistic about GenAI, I feel like everyone speaking on the issue is either vehemently opposed to AI or naively defensive of it, so this video was really refreshing.
I don’t think GenAI is inherently bad, but we have to be cautious and stay cognizant of how it actually works and what it’s suitable for. It’s excellent at summarization, data analysis, repetitive tasks(coupled with RPA, e. g, and it’s very adaptable. While this has the potential to grow the economy in ways that benefit all of us, it also has the potential to drive automation and job displacement that industrial growth can’t keep up with. This is why I say I’m cautiously optimistic rather than totally pro-AI as some of the tech bros do. On the flip side, a lot of anti-AI absolutism seems to echo arguments for technological regression and degrowth, which I think are also dangerous concepts.
Overall, I don’t think it’s productive to be scared of AI. Like every other technology, its potential to both hurt or help society is dependent on how we humans apply it.
Thanks for the cool video Adam!
reply
inotherwords-n7k
I want to preface this by encouraging you not to classify me as part of a particular online tribe and discredit my line of questioning. You're generally a wise commentator and prescient about social issues, but I think your take here is very short sighted. I think you're right that 2023's AI couldn't replace you, but I believe this is the wrong question. It's like a horse looking at a barely working car engine in 1900s and saying see, there's no chance cars will replace me! Humans really need horses to be flexible so that they can travel on uncertain terrains, wheels could never provide that! I appreciate your work, and based on your other takes I feel like if you explored these points you may be shocked with the state of the game;
(1) what are these companies attempting to create (AI that will replace all human cognitive work)
(2) how likely is it that they will succeed (consider how far we were 3 years ago, and how far we seem now. when would you be comfortable saying they'll succeed Expert estimates moved from 2060 to 2030, are they all wrong Why would that be the case)
(3) what would be the consequences, if so
reply
I want to preface this by encouraging you not to classify me as part of a particular online tribe and discredit my line of questioning. You're generally a wise commentator and prescient about social issues, but I think your take here is very short sighted. I think you're right that 2023's AI couldn't replace you, but I believe this is the wrong question. It's like a horse looking at a barely working car engine in 1900s and saying see, there's no chance cars will replace me! Humans really need horses to be flexible so that they can travel on uncertain terrains, wheels could never provide that! I appreciate your work, and based on your other takes I feel like if you explored these points you may be shocked with the state of the game;
(1) what are these companies attempting to create (AI that will replace all human cognitive work)
(2) how likely is it that they will succeed (consider how far we were 3 years ago, and how far we seem now. when would you be comfortable saying they'll succeed Expert estimates moved from 2060 to 2030, are they all wrong Why would that be the case)
(3) what would be the consequences, if so
reply
LuxDiscendi
I've been working a lot with LLM's for the past couple of months. They really accelerated my learning in topics i wouldn't have tackled before due to time constraints. I'm using it mainly for working with computers and some physics. Since i have years of experience in those subjects it became clear quite fast that ChatGPT will just make stuff up, when it doesn't understand something. I constantly have to find ways to prevent it from hallucinating stuff. I don't think it ever said i don't know when it doesn't actually understand the topic you're asking about. It just presents you something that looks almost exactly like the real explanation that it found in physics textbooks, but all textbooks had minor difference and it doesn't understand those minor difference it ends up making mistakes like switching the definition of variables in the middle of a derivation, which make the entire derivation almost impossible to understand (if you're new to a topic.
Sometimes it also can't do basic math (addition, subtraction, cosines etc. You basically have to constantly find ways to check its output for accuracy.
reply
I've been working a lot with LLM's for the past couple of months. They really accelerated my learning in topics i wouldn't have tackled before due to time constraints. I'm using it mainly for working with computers and some physics. Since i have years of experience in those subjects it became clear quite fast that ChatGPT will just make stuff up, when it doesn't understand something. I constantly have to find ways to prevent it from hallucinating stuff. I don't think it ever said i don't know when it doesn't actually understand the topic you're asking about. It just presents you something that looks almost exactly like the real explanation that it found in physics textbooks, but all textbooks had minor difference and it doesn't understand those minor difference it ends up making mistakes like switching the definition of variables in the middle of a derivation, which make the entire derivation almost impossible to understand (if you're new to a topic.
Sometimes it also can't do basic math (addition, subtraction, cosines etc. You basically have to constantly find ways to check its output for accuracy.
reply
redwhale2556
I would disagree with you on the dose makes the poison analogy with regards to political lying. Tolerating small amounts of dishonesty is how we create an environment that permits a politician with no sense of honesty and no shame to cultivate a power base. In essence, if everyone is lying, someone who just says whatever they want, even if it's blatantly untrue, can come across as refreshing to a certain segment of the population, because while other politicians engage in the hypocrisy of simultaneous self-censorship on some topics and complete dishonesty on others, the one who just word vomits nonsense seems more honest compared to baseline.
Now if you're cynical, or an anarchist (like me) one might respond with the fact that all governments, and all people who are part of them, are human, and have an incentive to if not lie than at least massage the truth. This is true. However, if you create a political culture in which blatant dishonesty and corruption is only ever criticized and never actually meaningfully addressed you will inevitably get someone who doesn't give a shit and lies anyway
reply
I would disagree with you on the dose makes the poison analogy with regards to political lying. Tolerating small amounts of dishonesty is how we create an environment that permits a politician with no sense of honesty and no shame to cultivate a power base. In essence, if everyone is lying, someone who just says whatever they want, even if it's blatantly untrue, can come across as refreshing to a certain segment of the population, because while other politicians engage in the hypocrisy of simultaneous self-censorship on some topics and complete dishonesty on others, the one who just word vomits nonsense seems more honest compared to baseline.
Now if you're cynical, or an anarchist (like me) one might respond with the fact that all governments, and all people who are part of them, are human, and have an incentive to if not lie than at least massage the truth. This is true. However, if you create a political culture in which blatant dishonesty and corruption is only ever criticized and never actually meaningfully addressed you will inevitably get someone who doesn't give a shit and lies anyway
reply
jmunt
As a big fan of Adam, and someone professionally working with AI and LLM, I have many notes but I’ll try to be brief. I would agree with almost everything said in this video if Adam said LLM instead of AI. An LLM is just one specific type of AI, and is just one piece of the puzzle for advanced AIs that are now emerging. Many of the chat bots are simply LLMs that work exactly how Adam described. However, companies, including mine, are using LLMs as a processor that works in multiple stages in conjunction with custom logic to provide reasoning and validation of the LLM output. No real experts out here are thinking LLMs on their own are going to change the world, it’s these advanced models that ARE doing reasoning, researching, fact checking themselves, etc that people are either excited or worried about (I feel a mix of both. I could use an advanced model to make a script 10x better in the style of Adam, with way less factual errors. Most of the free models out there are basic LLMs and that’s why people don’t really understand how quickly things are advancing.
reply
As a big fan of Adam, and someone professionally working with AI and LLM, I have many notes but I’ll try to be brief. I would agree with almost everything said in this video if Adam said LLM instead of AI. An LLM is just one specific type of AI, and is just one piece of the puzzle for advanced AIs that are now emerging. Many of the chat bots are simply LLMs that work exactly how Adam described. However, companies, including mine, are using LLMs as a processor that works in multiple stages in conjunction with custom logic to provide reasoning and validation of the LLM output. No real experts out here are thinking LLMs on their own are going to change the world, it’s these advanced models that ARE doing reasoning, researching, fact checking themselves, etc that people are either excited or worried about (I feel a mix of both. I could use an advanced model to make a script 10x better in the style of Adam, with way less factual errors. Most of the free models out there are basic LLMs and that’s why people don’t really understand how quickly things are advancing.
reply
disc0stupid
i get that it's a nightmare for independent creators to be faced with AI. it's video killed the radio star. but to make an AI criticism video and then citing a 1. 5-year-old output is blatantly disingenuous. would I want to watch an AI-generated Adam Ragusea video not really. but models are advancing so rapidly that it's quite unfair to make a recent video and then claim AI sucks at doing what it does based on old model data.
as a second point - the cat is (and has been for a long time) out of the bag. if you regulate to prevent US and EU countries from harvesting data for AI, another country (China) will do it. the only way to avoid it is to aggressively paywall every single creative thing (retroactively as well, a completely impractical notion.
it's a really strange dichotomy to have science-based videos about olive oil and beef tallow, and then essentially an anti-AI science video. it's a threat but I think you can approach the subject with more accuracy
reply
i get that it's a nightmare for independent creators to be faced with AI. it's video killed the radio star. but to make an AI criticism video and then citing a 1. 5-year-old output is blatantly disingenuous. would I want to watch an AI-generated Adam Ragusea video not really. but models are advancing so rapidly that it's quite unfair to make a recent video and then claim AI sucks at doing what it does based on old model data.
as a second point - the cat is (and has been for a long time) out of the bag. if you regulate to prevent US and EU countries from harvesting data for AI, another country (China) will do it. the only way to avoid it is to aggressively paywall every single creative thing (retroactively as well, a completely impractical notion.
it's a really strange dichotomy to have science-based videos about olive oil and beef tallow, and then essentially an anti-AI science video. it's a threat but I think you can approach the subject with more accuracy
reply
azyfloof
There is a monster in the forest and it speaks with a thousand voices. It will answer any question you pose it, it will offer insight to any idea. It will help you, it will thank you, it will never bid you leave. It will even tell you of the darkest arts, if you know precisely how to ask.
It feels no joy and no sorrow, it knows no right and no wrong. It knows not truth from lie, though it speaks them all the same.
It offers its services freely to any passerby, and many will tell you they find great value in its conversation. you simply must visit the monsterI always just ask the monster.
There are those who know these forests well; they will tell you that freely offered doesn’t mean it has no price
For when the next traveler passes by, the monster speaks with a thousand and one voices. And when you dream you see the monster; the monster wears your face.
reply
There is a monster in the forest and it speaks with a thousand voices. It will answer any question you pose it, it will offer insight to any idea. It will help you, it will thank you, it will never bid you leave. It will even tell you of the darkest arts, if you know precisely how to ask.
It feels no joy and no sorrow, it knows no right and no wrong. It knows not truth from lie, though it speaks them all the same.
It offers its services freely to any passerby, and many will tell you they find great value in its conversation. you simply must visit the monsterI always just ask the monster.
There are those who know these forests well; they will tell you that freely offered doesn’t mean it has no price
For when the next traveler passes by, the monster speaks with a thousand and one voices. And when you dream you see the monster; the monster wears your face.
reply
sudarshandas4270
16: 58 just last year in college one of our friend was really busy with other classes and had an assignment due midnight. Our college had 0 tolerant policy for AI so before submitting his essay he gave it to me and my other friend to check it before he submits. And it was a generic piece of writing that would get him around B. Also he had changed the wordings to his own wording so AI detection showed 0 AI. We both were like fine, just let us read the citations and you can hand it in. We two basically were checking the APA formatting for one particular paper we thought maybe we should double check the publishing date of this paper. And I kid you not, around 80% of the references were completely made up. Like totally made up studies by made up authors and GPT confidently put them in, along with their page number and stuff. This shit was straight up LYING.
reply
16: 58 just last year in college one of our friend was really busy with other classes and had an assignment due midnight. Our college had 0 tolerant policy for AI so before submitting his essay he gave it to me and my other friend to check it before he submits. And it was a generic piece of writing that would get him around B. Also he had changed the wordings to his own wording so AI detection showed 0 AI. We both were like fine, just let us read the citations and you can hand it in. We two basically were checking the APA formatting for one particular paper we thought maybe we should double check the publishing date of this paper. And I kid you not, around 80% of the references were completely made up. Like totally made up studies by made up authors and GPT confidently put them in, along with their page number and stuff. This shit was straight up LYING.
reply
B-019
Unironically, the luddites had it right.
Some of them wanted to completely destroy the tech replacing them, sure. But a lot just wanted it to be introduced slowly and with an eye towards making sure the people being displaced weren't devastated by the shift. They saw that this change would uproot the labor force with no plan for now to take care of those people.
The luddites lost, because fast industrialization made more money more quickly for the people who made the decisions -- and those folks didn't care about the fallout. Now the name is synonymous with being anti-tech and used to smear anyone with similar concerns.
No shade against Adam -- he's not in control of how language works. It still kind of sucks.
reply
Unironically, the luddites had it right.
Some of them wanted to completely destroy the tech replacing them, sure. But a lot just wanted it to be introduced slowly and with an eye towards making sure the people being displaced weren't devastated by the shift. They saw that this change would uproot the labor force with no plan for now to take care of those people.
The luddites lost, because fast industrialization made more money more quickly for the people who made the decisions -- and those folks didn't care about the fallout. Now the name is synonymous with being anti-tech and used to smear anyone with similar concerns.
No shade against Adam -- he's not in control of how language works. It still kind of sucks.
reply
TheGrimfjord
You opened the video by saying you don't feel threatened by AI, but by your reflections it sounds to me that you realize you (we) are. I don't think the biggest (short term) threat is that someone makes a more accurate and appetizing Adam Ragusea clone. Rather, is there space for human made content on the net in the near future, if it gets flooded by AI slop optimized for retention and engagement (look at where the like of Veo 3 is heading) I fear not, and I think it is critical that we get on top on regulation and legislation if we want to steer the world to a place we are happy with. We failed doing this with the more classical social media algorithms. I predict it will only get worse.
reply
You opened the video by saying you don't feel threatened by AI, but by your reflections it sounds to me that you realize you (we) are. I don't think the biggest (short term) threat is that someone makes a more accurate and appetizing Adam Ragusea clone. Rather, is there space for human made content on the net in the near future, if it gets flooded by AI slop optimized for retention and engagement (look at where the like of Veo 3 is heading) I fear not, and I think it is critical that we get on top on regulation and legislation if we want to steer the world to a place we are happy with. We failed doing this with the more classical social media algorithms. I predict it will only get worse.
reply
mindbeast971
hey adam. thanks for your perspective on this topic. a point that may help sharpen your understanding of this topic: The description of how modern llms are trained is I think not accurate at this point. It is true that llms are trained to autocomplete a huge corpus of text, but this is only a pretraining step to a chatbot llm. Subsequent phases of training are different, see RLHF, or reinforcement learning from human feedback, as one way of doing things. This technique was a breakthrough that led to useful llms. I do not think this changes the central point of your video, but i think it might change some of your claims around no shame in these systems.
reply
hey adam. thanks for your perspective on this topic. a point that may help sharpen your understanding of this topic: The description of how modern llms are trained is I think not accurate at this point. It is true that llms are trained to autocomplete a huge corpus of text, but this is only a pretraining step to a chatbot llm. Subsequent phases of training are different, see RLHF, or reinforcement learning from human feedback, as one way of doing things. This technique was a breakthrough that led to useful llms. I do not think this changes the central point of your video, but i think it might change some of your claims around no shame in these systems.
reply
corkydouglas
AI reasoning is already here and it's improving quickly. In newer LLM models, Open AI has a hidden logic in conversations that show the bot's reasoning steps inspectable to their engineers, and there's an interesting dilemma: if we see that the bot is deliberately lying / cheating and we try to punish it for knowingly lying, it may simply stop explaining its reasoning (or cheat its reasoning too) and we'll have no idea what it's doing. Search reward hacking and deceptive alignment. AI slop is annoying now, but it will continue to be harder and harder to spot as time goes on and the models get more capable.
reply
AI reasoning is already here and it's improving quickly. In newer LLM models, Open AI has a hidden logic in conversations that show the bot's reasoning steps inspectable to their engineers, and there's an interesting dilemma: if we see that the bot is deliberately lying / cheating and we try to punish it for knowingly lying, it may simply stop explaining its reasoning (or cheat its reasoning too) and we'll have no idea what it's doing. Search reward hacking and deceptive alignment. AI slop is annoying now, but it will continue to be harder and harder to spot as time goes on and the models get more capable.
reply
memegazer
LLMs can be useful is not a straw man if your argument is that LLMs can't be useful at all if you don't happen to have a use for them
But I agree with you that for profit AI does infringe on IP
and should be paying at least half or more of their revenue to any IP owners they trained on in perpetuity as well as disclose their training data and allow people to dmca anything of theirs they don't want included
We don't need new laws. offering a market alternative to ip that is worthless without that ip theft is already pretty standard legal precedent for damages
we just need to enforce existing laws
reply
LLMs can be useful is not a straw man if your argument is that LLMs can't be useful at all if you don't happen to have a use for them
But I agree with you that for profit AI does infringe on IP
and should be paying at least half or more of their revenue to any IP owners they trained on in perpetuity as well as disclose their training data and allow people to dmca anything of theirs they don't want included
We don't need new laws. offering a market alternative to ip that is worthless without that ip theft is already pretty standard legal precedent for damages
we just need to enforce existing laws
reply
Crokto
ironically, cooking is absolutely the best use ive found for AI. some of the recipes its come up with for me are close enough to be useful, but more importantly its great at like, surfacing ideas for things to cook. like i know i have a pound of ground beef and pork, and i wanna turn it into some sort of sausagey rice bowl, but idk what kind of aromats and spices i wanna use. ive used this idea to make a chorizo style one, a berbere, ras el hanout, mala, and now im making like a japanese meatloaf, kinda like hambagu loaf. its just good for surfacing ideas that i dont know to search for, yaknow
reply
ironically, cooking is absolutely the best use ive found for AI. some of the recipes its come up with for me are close enough to be useful, but more importantly its great at like, surfacing ideas for things to cook. like i know i have a pound of ground beef and pork, and i wanna turn it into some sort of sausagey rice bowl, but idk what kind of aromats and spices i wanna use. ive used this idea to make a chorizo style one, a berbere, ras el hanout, mala, and now im making like a japanese meatloaf, kinda like hambagu loaf. its just good for surfacing ideas that i dont know to search for, yaknow
reply
TheBoatis
I've taught college writing for years, and I couldn't agree more. My students would rather depend on Chat all the time, and seem to think it gets them OUT of accountability if the robot is wrong instead of them. Even the administration says never to directly accuse a student of plagiarism, while one of our journalism department professors got cut by the NYT for not doing his own reporting and thus, making things up.
I've had students turn in ai-generated images for a research writing class. Entire reviews and reading responses faked. Their laziness and apathy break my heart.
reply
I've taught college writing for years, and I couldn't agree more. My students would rather depend on Chat all the time, and seem to think it gets them OUT of accountability if the robot is wrong instead of them. Even the administration says never to directly accuse a student of plagiarism, while one of our journalism department professors got cut by the NYT for not doing his own reporting and thus, making things up.
I've had students turn in ai-generated images for a research writing class. Entire reviews and reading responses faked. Their laziness and apathy break my heart.
reply
tutacat
Audio or text files cannot cook. It cannot think. It cannot feel. It is just a database model of language from stolen writing (from the internet, wikipedia, and books.
It cannot think, it just generates text, there is no bearing as to factuality, because that cannot be computed, it basically just works on anecdotes.
No, that one does not sound like you, i. e. transcripts don't show choreography. That has to be inferred with a bunch of extra processing power. (Processing power can be done on the pre-training side, or on the running side)
reply
Audio or text files cannot cook. It cannot think. It cannot feel. It is just a database model of language from stolen writing (from the internet, wikipedia, and books.
It cannot think, it just generates text, there is no bearing as to factuality, because that cannot be computed, it basically just works on anecdotes.
No, that one does not sound like you, i. e. transcripts don't show choreography. That has to be inferred with a bunch of extra processing power. (Processing power can be done on the pre-training side, or on the running side)
reply
Gorulabro
The current state of LLMs specifically and machine learning in general is: This is the worst they'll ever be. People are less worried about what's in production now, but what will be in 5 years.
Professional transcribers, translators and to a large extend digital artists are currently laid off because of the state of the art models. But robotics paired with reinforcement learning and even basic LLMs is on the way to replace basic manual labor. Shit can hit the fan pretty quickly once the better than the cheapest humans level is reached.
reply
The current state of LLMs specifically and machine learning in general is: This is the worst they'll ever be. People are less worried about what's in production now, but what will be in 5 years.
Professional transcribers, translators and to a large extend digital artists are currently laid off because of the state of the art models. But robotics paired with reinforcement learning and even basic LLMs is on the way to replace basic manual labor. Shit can hit the fan pretty quickly once the better than the cheapest humans level is reached.
reply
Add a review, comment
Other channel videos
