By YouSum Live 00:00:00 Science and AI limitations in predicting complex systems. 00:00:30 AI struggles with extrapolation beyond trained data. 00:01:41 Language simplicity aids AI success in text analysis. 00:02:07 AI's limitations in creativity and originality. 00:06:53 Computational exploration of vast possibilities by humans. 00:09:23 Computational language as a tool for formalizing the world. 00:14:52 The importance of computational thinking and automation in work. 00:15:20 Leveraging AI as an interface for computational tasks. 00:16:48 Training AI models for specific computational tasks. 00:17:45 Weak form of computation in llms. 00:17:50 Challenges in guiding proofs using llms. 00:18:00 Limitations of llms in mathematical proofs. 00:18:43 Llms excel in making homework but struggle at edge of human knowledge. 00:19:02 Llms prone to errors in math without guidance. 00:19:37 OpenAI's focus on long-form reasoning surpassing human capabilities. 00:20:20 Building systems to extend human capabilities. 00:20:36 Exploring the fundamental workings of machine learning. 00:21:26 Balancing computational capabilities with human needs. 00:22:11 Challenges in developing effective AI tutoring systems. 00:22:28 Goal for llms to understand and assist human learning. 00:23:00 Conceptualizing beyond human intelligence and AI capabilities. By YouSum Live
I find 99% of all the UA-cam ‘geniuses’ to represent admiration rather than appreciating what the term implies. But in your case, yes, Stephen Wolfram is a genius in the true sense of the word.
Also it's pretty discrediting to LLMs to say they are only good because language has (easy) grammar. A lot of tests on LLMs show that they have a (though limited, incomplete) world model. It understands basic mathematics, and some basic things about our world.
I think that world model that LLMs have is a fundamental part of language, and shows there are deeper underlying patterns in language that hint at a world view. I think that's what we are seeing with this emergent world view in LLMs
It doesn't "understand" anything, it is able through it's massive training data to recall information it has seen before and piece it together in a legible format.
So, there is a plugin for ChatGPT so it can access Wolfram resources. How about an interface to Wolfram resources that can be used by any language model?
Semantic space has a shape. It's a model, so of course it has a similar shape to what is being modeled. I like the idea that only that which is simple or computationally reducible can be modeled sufficiently in current scale foundation models. Rigorous agentic behavior is necessary to deal with computationally difficult activation pathways.
Before we can expect an AI to accurately predict meaningful events, it probably needs to be able to accurately describe the present, and prior events. A graph structure is probably a good way to represent the present and the past.
I understood NKS is based on Emergence. Sabine is supporting now. Is Stephen Wolphram nks based on emergence? Say yes or no Yes. ChatGPT I understood you after the movie Automata
LMs in Science? Do LMs behave like scientists? After chatting with ChatGPT for some time I think it behaves like a lazy student during an oral exam. The student is brilliant in using the language, maybe because he already read some books in his life but he did not prepare for this particular exam. So, when asked by the teacher he produces a good sounding answer, the best he can produce, some kind of small talk inspired by a mix of everything he read in his life. The question is, are we (humans, scientists) all and always behaving like a lazy student? Does the ability to create or find new knowledge emerge from the size of a LM?
Yeah I was very confused about that. It's very easy for a shallow, like 5-6 layer deep neural network to learn a very decent approximation of a sin(x) wave, very quickly. I don't know what he meant to say there. (And yes, with ReLU activation only)
@@comradepeter87 Yes, but will it continue the sine wave outside of the data it was trained on? It will not. It can't because none of the math of the nn is periodic.
YOU JUST DON'T WATCH ENOUGH RICK AND MORTY ! YOU DON'T GET IT BRO, IT'S NOT BRAGGING IF YOU CAN BACK IT UP 🔥🔥😤😤😤💰🤑🔬🧑🏾🔬🥼⚗️🧫 NARCISSISTS ARE HUMANS TOO 😡🥵🙂↔️🤪🤨📸📸📸📸📸📸📸😵😵😵😵😵😵😵😵😵😵😂😂😂
My opinion is that Stephen Wolfram is struggling to understand modern transformer based AI and I'm not sure why because he's actually described them in the past. Whilst its technically true an LLM such as chat GPT simply produces a next word based on statistics, its disingenuous to say that's all its doing because with that statement there is no implication of the profound understanding inherent in the model of every word prior, in the context of its training leading to that choice. Stephen's downplaying of neural networks leaves me cold. Sorry, your attempt at creating AI using graphs and tokens didn't work out, Stephen.
I think I agree with you, but you found some pretty harsh words. Stephen has as broad a horizon in mathematics, physics and computer science as it gets and is absolutely used to integrating new ideas into his framework quickly and easily. LLMs are a new idea he tried to explore scientifically too quickly without having tested them enough as a user before, to get a good grasp of their capabilities and understanding of the world. The actual problem then arises when he's ask about LLMs and speaks with confidence about them without having a clear and above average grasp.
Ok. So. That seems wrong. A sine wave is super easy to define and anyone can plot the next points based on the previous points whereas for a sentence the prediction, yes maybe easier than we thought, but it’s much harder to predict the next word. Anyway. I feel that was a poor example.
Regarding the sine wave example.. He meant that you have to make the machine learn trigonometry first to be able to complete the plot (Which is a herculian task). Without learning, it would just copy.
If ChatGPT writing code makes it boilerplate, then so is nearly all code. It outdesigned humans for ML reward heuristics (Eureka paper), for example. He's much more pessimistic on AI than I imagined. Disappointing.
Never heard of this guy, but he comes across as someone whose qualities do not include humility and curiosity. His idea of the genesis of human language seems rather unsophisticated. People 200000 years ago were certainly much more concerned with animals and plants they could eat than they were with rocks. That he mentions rocks first tells me that he has not really spent a lot of time thinking about how human language could have evolved.
Bud, I haven't yet watched the video but that's Wolfram, creator of the Wolfram language. Certainly he knows about the language, maths and qualities of the emergent complexities more than any of us
Bro people were drawing art on cave walls 50k years ago. Pretty sure rocks were not only important hunting material, but also to tell stories. I wouldn’t be surprised if rocks were worshiped like gods… Imagine what apes did when finding gold I wonder
Wow, Joscha Bach and Stephen Wolfram on one stage. Is there more of this discussion?
looks like it's from a UA-cam channel Imagination in Action, they have a Wolfram video posted 6 days ago in the same outfit, but not this chat yet 🤞
By YouSum Live
00:00:00 Science and AI limitations in predicting complex systems.
00:00:30 AI struggles with extrapolation beyond trained data.
00:01:41 Language simplicity aids AI success in text analysis.
00:02:07 AI's limitations in creativity and originality.
00:06:53 Computational exploration of vast possibilities by humans.
00:09:23 Computational language as a tool for formalizing the world.
00:14:52 The importance of computational thinking and automation in work.
00:15:20 Leveraging AI as an interface for computational tasks.
00:16:48 Training AI models for specific computational tasks.
00:17:45 Weak form of computation in llms.
00:17:50 Challenges in guiding proofs using llms.
00:18:00 Limitations of llms in mathematical proofs.
00:18:43 Llms excel in making homework but struggle at edge of human knowledge.
00:19:02 Llms prone to errors in math without guidance.
00:19:37 OpenAI's focus on long-form reasoning surpassing human capabilities.
00:20:20 Building systems to extend human capabilities.
00:20:36 Exploring the fundamental workings of machine learning.
00:21:26 Balancing computational capabilities with human needs.
00:22:11 Challenges in developing effective AI tutoring systems.
00:22:28 Goal for llms to understand and assist human learning.
00:23:00 Conceptualizing beyond human intelligence and AI capabilities.
By YouSum Live
My AI predicted the text “so to speak”. Only joking, I love Stephen’s videos. He’s a true genius.
I find 99% of all the UA-cam ‘geniuses’ to represent admiration rather than appreciating what the term implies. But in your case, yes, Stephen Wolfram is a genius in the true sense of the word.
Good one
Wolfram is our modern day genius.
Anyone know where to find the full talk? Thank you
Yeah I’m looking for the same
Also it's pretty discrediting to LLMs to say they are only good because language has (easy) grammar. A lot of tests on LLMs show that they have a (though limited, incomplete) world model. It understands basic mathematics, and some basic things about our world.
I think that world model that LLMs have is a fundamental part of language, and shows there are deeper underlying patterns in language that hint at a world view. I think that's what we are seeing with this emergent world view in LLMs
It doesn't "understand" anything, it is able through it's massive training data to recall information it has seen before and piece it together in a legible format.
@@MatthewLappin-g9d what is human understanding if not a bio-neural network taking in training data to create a cybernetic feedback system?
I liked the part where he said "computational"
Computationally speaking of course.
Dream : Wolphram and Tegmark takking to each other.
So, there is a plugin for ChatGPT so it can access Wolfram resources. How about an interface to Wolfram resources that can be used by any language model?
thanks
Semantic space has a shape. It's a model, so of course it has a similar shape to what is being modeled. I like the idea that only that which is simple or computationally reducible can be modeled sufficiently in current scale foundation models. Rigorous agentic behavior is necessary to deal with computationally difficult activation pathways.
Before we can expect an AI to accurately predict meaningful events, it probably needs to be able to accurately describe the present, and prior events. A graph structure is probably a good way to represent the present and the past.
17:00
I cannot challenge Stephen but Max os saying it can give a symbolic equation like sine theta.
Would be helpful if he could produce a simple example in which LLM plus his calculation engine is better than LLM alone.
There is no specificity regarding the metrics of measuring computational intelligence and representing it.
What about protein folding. AI was wildly successful.
Map your perception onto the Transformers Perception
2:44 Textadistics?
❤❤❤❤❤❤
😊
Sometimes I wonder why people talk about stuff so clearly and still miss the point 😂😂
I understood NKS is based on Emergence. Sabine is supporting now.
Is Stephen Wolphram nks based on emergence? Say yes or no
Yes.
ChatGPT
I understood you after the movie Automata
LMs in Science? Do LMs behave like scientists? After chatting with ChatGPT for some time I think it behaves like a lazy student during an oral exam. The student is brilliant in using the language, maybe because he already read some books in his life but he did not prepare for this particular exam. So, when asked by the teacher he produces a good sounding answer, the best he can produce, some kind of small talk inspired by a mix of everything he read in his life. The question is, are we (humans, scientists) all and always behaving like a lazy student? Does the ability to create or find new knowledge emerge from the size of a LM?
Hype vs Reality
"have you know maybe 50,000 words in typical languages” THERES 170,000 words in English
“If it has linear activation functions it will predict a linear continuation of a sine wave”, really?! Not sure about that one 😅
Yeah I was very confused about that. It's very easy for a shallow, like 5-6 layer deep neural network to learn a very decent approximation of a sin(x) wave, very quickly. I don't know what he meant to say there. (And yes, with ReLU activation only)
@@comradepeter87 Yes, but will it continue the sine wave outside of the data it was trained on? It will not. It can't because none of the math of the nn is periodic.
@@comradepeter87 ReLU is not a linear activation function though
I love him but his conceit is exhausting
YOU JUST DON'T WATCH ENOUGH RICK AND MORTY ! YOU DON'T GET IT BRO, IT'S NOT BRAGGING IF YOU CAN BACK IT UP 🔥🔥😤😤😤💰🤑🔬🧑🏾🔬🥼⚗️🧫 NARCISSISTS ARE HUMANS TOO 😡🥵🙂↔️🤪🤨📸📸📸📸📸📸📸😵😵😵😵😵😵😵😵😵😵😂😂😂
@@Rawi888is that a comment or is it art?
@@honkytonk4465 inclusiveOR
My opinion is that Stephen Wolfram is struggling to understand modern transformer based AI and I'm not sure why because he's actually described them in the past. Whilst its technically true an LLM such as chat GPT simply produces a next word based on statistics, its disingenuous to say that's all its doing because with that statement there is no implication of the profound understanding inherent in the model of every word prior, in the context of its training leading to that choice.
Stephen's downplaying of neural networks leaves me cold. Sorry, your attempt at creating AI using graphs and tokens didn't work out, Stephen.
I think I agree with you, but you found some pretty harsh words. Stephen has as broad a horizon in mathematics, physics and computer science as it gets and is absolutely used to integrating new ideas into his framework quickly and easily. LLMs are a new idea he tried to explore scientifically too quickly without having tested them enough as a user before, to get a good grasp of their capabilities and understanding of the world. The actual problem then arises when he's ask about LLMs and speaks with confidence about them without having a clear and above average grasp.
ai will become more useful over time to help people drink water.
Ok. So. That seems wrong. A sine wave is super easy to define and anyone can plot the next points based on the previous points whereas for a sentence the prediction, yes maybe easier than we thought, but it’s much harder to predict the next word. Anyway. I feel that was a poor example.
Regarding the sine wave example.. He meant that you have to make the machine learn trigonometry first to be able to complete the plot (Which is a herculian task). Without learning, it would just copy.
he lying out of teeth in some areas
This guy is overrated, he is saying much but concludes very little.
If ChatGPT writing code makes it boilerplate, then so is nearly all code. It outdesigned humans for ML reward heuristics (Eureka paper), for example.
He's much more pessimistic on AI than I imagined. Disappointing.
Never heard of this guy, but he comes across as someone whose qualities do not include humility and curiosity. His idea of the genesis of human language seems rather unsophisticated. People 200000 years ago were certainly much more concerned with animals and plants they could eat than they were with rocks. That he mentions rocks first tells me that he has not really spent a lot of time thinking about how human language could have evolved.
Humility no, curiosity yes
Bud, I haven't yet watched the video but that's Wolfram, creator of the Wolfram language.
Certainly he knows about the language, maths and qualities of the emergent complexities more than any of us
@@udaykadam5455 Hi Stephen
Bro people were drawing art on cave walls 50k years ago. Pretty sure rocks were not only important hunting material, but also to tell stories. I wouldn’t be surprised if rocks were worshiped like gods…
Imagine what apes did when finding gold I wonder
he is too biased to say AI is dumb. this is not a good spirit