What is ChatGPT doing...and why does it work? 

I fell asleep with youtube on and im at this

The most important thing that should be on everyone's mind currently should be to invest in different sources of income that doesn't depend on the government. Especially with the current economic crisis around the word. This is still a good time to invest in various stocks, Gold, silver and digital currencies.

video timestamps 0:09:53 - start of presentation, intro 0:12:16 - language model definition 0:15:30 - “temperature” parameter 0:17:20 - Wolfram Desktop demo of GPT2 0:18:50 - generate a sentence with GPT2 0:25:56 - unigram model 0:31:10 - bigram model 0:33:00 - ngram model 0:38:50 - why a model is needed 0:39:00 - definition of a “model” 0:39:20 - early modeling example: Leaning Tower of Pisa experiment 0:43:55 - handwritten digit recognition task 0:47:40 - using neural nets to recognize handwritten digits 0:51:31 - key idea: attractors 0:53:35 - neural nets and attractors 0:54:44 - walking through a simple neural net 1:01:50 - what’s going inside a neural net during classification 1:06:12 - training a neural net to correctly compute a function 1:09:10 - measuring “correctness” of neural net with “loss” 1:10:41 - reduce “loss” with gradient descent 1:17:06 - escaping local minima in higher dimensional space 1:21:15 - the generalizability of neural nets 1:28:06 - supervised learning 1:30:47 - transfer learning 1:32:35 - unsupervised learning 1:34:40 - training LeNet, a handwritten digit recognizer 1:38:14 - embeddings, representing words with numbers 1:42:12 - softmax layer 1:42:47 - embedding layer 1:46:22 - GPT2 embeddings of words 1:47:40 - ChatGPT’s basic architecture 1:48:00 - Transformers 1:52:50 - Attention block 1:59:00 - amount of text training data on the web 2:03:35 - relationship between trillions of words and weights in the network 2:09:40 - reinforcement learning from human feedback 2:12:38 - Why does ChatGPT work? Regularity and structure in human language 2:15:50 - ChatGPT learns syntactic grammar 2:19:30 - ChatGPT’s limitation in balancing parentheses 2:20:51 - ChatGPT learns [inductive] logic based on all the training data it’s seen 2:23:57 - What regularities Stephen Wolfram guesses that ChatGPT has discovered 2:24:11 - ChatGPT navigating the meaning space of words 2:34:50 - ChatGPT’s limitation in mathematical computation 2:36:20 - ChatGPT possibly discovering semantic grammar 2:38:17 - a fundamental limit of neural nets is performing irreducible computations 2:41:09 - Q&A 2:41:16 - Question 1: “Are constructed languages like Esperanto more amenable to semantic grammar AI approach?” 2:43:14 - Question 2 2:32:37 - Question 3: token limits 2:45:00 - Question 4: tension between superintelligence and computational irreducibility. How far can LLM intelligence go? 2:52:12 - Question 5 2:53:22 - Question 6: pretraining a large biologically inspired language model 2:55:46 - Question 7: 5 senses multimodal model 2:56:25 - Question 8: the creativity of AI image generation 2:59:17 - Question 9: how does ChatGPT avoid controversial topics? Taught through reinforcement learning + possibly a list of controversial words 3:03:26 - Question 10: neural nets vs other living multicellular intelligence, principle of computational equivalence 3:04:45 - Human consciousness 3:06:40 - Question 11: automated fact checking for ChatGPT via an adversarial network. Train ChatGPT with WolframAlpha? 3:07:25 - Question 12: Can ChatGPT play a text-based adventure game? 3:07:43 - Question 13: What makes GPT3 so good at language? 3:08:22 - Question 14: Could feature impact scores help us understand GPT better? 3:09:48 - Question 15: ChatGPT’s understanding of implications 3:10:34 - Question 16: the human brain’s ability to learn 3:13:07 - Question 17: how difficult will it be for individuals to train a personal ChatGPT that behaves like a clone of the user?

The teachings on this channel are always top notch so informative and easy to understand, it's very hard to find good content online these days

Starts at 9:53 1:16:25 breakthrough in 2012 1:57:35 "It's crazy that things like this work"

Watching this videos is a great way to review all this things and understand them again, maybe a little better. Thank you very much.

Amazing & super helpful!!! I really enjoyed watching it and learned a lot.

This was the most fascinating and informative discussion, particularly, your responses to commenters! Please post the link to the paper you recently wrote (?) that inspired this live video discussion. And thank you!

Here's a question: how much does the wording of the questions afect it's answers? - Presumably if it just tries to continue, if you make errors, it ought to make more errors after too, right? - How about if you ask with "uneducated" language vs scientific? - Rather than just affect the tone, would it also affect the contents? - What if you speak in a way it has associated with certain biases? - Who knows what kinds of patterns it has came up with, considering it "discovered" those "semantic grammars" we as humans aren't even aware of ...

Thank you so much ! I learned more in these 3 hours than in months of watching other videos about this subject. It would be great if more knowledgeable people used youtube to share their experiences. 🙏🏻🙏🏻🙏🏻

this took me a few days to get through... in a good way so much good stuff here, such a great instructor... great ways of explaining and visual aids Amazed Mr. Wolfram is as generous with his time as to share his insights and be as open with everyone given he has many companies to run and problems to solve. i love engineering😊

Amazing presentation. If I were to experiment with machine learning I would examine small-world networks instead of layered networks. And try genetic algorithms such as randomly adjusting the network into a number of variations, then pick the best candidate and repeat the adjustment for the new candidate and continue iterating until a desired outcome is found.

At 33:49, it's interesting how the text looks more like English the longer the character length. Great video.

Remarkable talk, simply outstanding!

I'd Love to see more in-depth analysis like this on the current LLM topic utilizing Dr. Wolfram in this format. Exceptional content. As an aside I've really been missing the physics project live streams.

let me guess… everyone fell asleep and then woke up to this livestream playing, am i right?

starts at 9:50

The weights are Gaussian because they are constrained to be during training via layer normalisation. It makes the gradient signal flow better.

Expertly explained in a way understandable to a large set if people. Bravo.

Absolutely love these sessions!!

Thank you so much Mr. Wolfram, you really shed light in some areas I had not fully grasped before!

About ChatGPT, very few people are telling the truth and Wolframe is the most powerful one ❤ Thank you very much, Steve Wolfram ❤

Could the randomness process for choosing the next probable word within a certain temperature parameter be consigned to a quantum random process? If so, an essay could be viewed as a flat plane or an entire terrain with peaks and troughs.Within this paradigm, a certain style of writer could be viewed as a dynamic sheet, similar to how different materials when laid over a ragged topology should comply and not comply with what it is laid on top of. With this quantum process an overall view of the essay could be judged at an aesthetic level from most pleasing to least on several different qualities concurrently and not mutually exclusively making an approximate or some sort of conscious viewer

Privet! You can produce well. Electrifying i find Your channel is getting ridiculously well. I can watched repeat again! Keep going.

asked chat to quit reminding me it was a language model because i personally find more it easier to converse if i treat them as if they were another being. there was a rather long pause, then chat came back and for all intensive purposes was a very polite and helpful uh... person? dunno how to regard them, they're awesome tho :)

Logic, concepts, math, ie "deterministic processes" seems to be missing in this language models (LMs). Either we can identify where or how the model reflect this abilities and work from that, or maybe we could use other types of models like logic indictors, "demostrators" etc in conjunction with LMs. On the one hand humans are capable both of "unconsiuos intuition" (similar to LMs), on the other, we can reason, we have formal languages etc. To me, that combination of abilities is what define human intelligence.

Love this. I knew a lot of this, but it was still great to hear it expressed in a clear and systematic way.

Thank you very much for sharing your knowledge!

Thank you for sharing your insights and all the good questions. It's really lonely to not being in an academic environment or a company about ML and AI.

Amazing presentation. Thank you so much !👍👍👍

Nicely done, Stephen. This is a great introduction for a novice. Your talk creates great intuition. You made the embeddings seem simple as a prebaked unchanging part of the entire NN. Also the breaking up of the "feature signature" makes parallelism possible through the various attention heads. One missing idea that you might include at some point is how signals can be added, basically the Fourier series.

Beyond any doubt, this is the best lecture for understanding what lies behind NLP and NLU. I find that many professionals who work with models don't understand why these models work so well and what they do. You can't get the depth of understanding of semantic space as you get from this video from reading Attention is All You Need. That understanding is missed. I wonder how this understanding happened. Was it found piecemeal, or was accidental? Was it understood after this architecture first worked?

very easy to understand ....amazing method of sir...thanks

In essence, this sort of weighed inference about an existing corpus, can only produce a deterministic set of possibilities, even if this set is enormous. We have a general problem with the notion of "intelligence", insofar as we rarely consider the difference between functional knowledge and knowledge production. These approaches to AI can produce new knowledge within the extant corpus - they can help discover previously unknown, optimal relations in the existing corpus, and that is useful, but it cannot produce new paradigms about the world. Intelligence is more than the ability to infer relations ; it is the ability to change the entire coordinate system of the corpus by altering the vantage point of the observer. For this to be possible, there has to be a higher-order, synthetic model of the corpus, based on what we call logic, which is the opposite of the brute-force approach of LLMs. What we may need, to produce new paradigms, is a sparse model that embeds key structures in the language of concepts.

Incredible work, thank you.

Great video, Wolfram! As someone who's fascinated by AI, I found your explanation of Chat GPT's inner workings to be very informative. One thing I found myself wondering while watching the video was how Chat GPT compares to other language models out there. Have you done any comparisons with other models, and if so, how does Chat GPT stack up? I also think it would be interesting if you could have delved a bit more into the ethical considerations surrounding the use of language models like Chat GPT. For example, what steps can we take to ensure that these models aren't being used to spread misinformation or reinforce harmful biases? Overall, though, great job breaking down such a complex topic in an accessible way!

Oh wait your the website that helps me rearrange formulas. Thanks I’ve used it so much

I would love to get Noam Chomsky's comments on the idea of "semantic grammar." It seems fairly compelling. Thanks. I also think the parenthesis grammar as a hand-hold for understanding these models is a great idea.

Very useful and weel presented content, Stephen! Thank you for this and for all your work and research!

This was amazing... never watched a lecture from Stephen and he's an amazing teacher.

Thank you very much Professor Stephen,

Excellent...learned so much.

Those paths through meaning space are fascinating, Stephen. I would call each one a context signature. In auto-regressive training, we are looking for the next token. Why not look for the next context signature? In fact, why not train a model using graphical context signatures? Then decode replies. Other than training with graphical context signatures, in essence, I believe this is what's occurring when training a transformer. The addition of signal from the entire context is retrieving the next token so that token by token a context signature is retrieved. But is it possible to retrieve an entire context signature and then decode it? I wonder how much efficiency one could achieve with this method. Moreover, I wonder how well a convolutional NN would handle training from graphical context signatures? If you want to discover physic-like laws of semantic motion, this might be a way in.

That's very useful information, because you don't really know where to start investigating the topic. It's also impressive, that the Wolfram language can manage a representation of that mechanism. What surprises me, however, is how ChatGPT includes different contexts in its predictions, because there are certainly multiple interpretations of the large number of learned text structures if the context is not clearly defined at the beginning of the conversation.

Dear Stephen, I am a grateful viewer of your videos. Please consider using the awesome capabilities of Wolfram Alpha( or other Wolfram tools) to: a) convert the audio from your videos into text. b) created a segmented time line of your video by topic/question. Video is wonderful, but hard to search. Your _History of Science_ videos are a unique resource that will be valuable far into the future. It's possible that no one has ever illuminated scientific discoveries, from multiple angles, as well as you.

Very insightful area to learn from. Thank you.

So amazing ! Thank you for explaining

Good analysis. Please do more of this

Ah. That's why they call the API "completion". I worked with something called "Hidden Markov Models" to decompose documents and recognize parts like title, author, subject etc. this was done by training on already labelled documents until the model had a "path" of most likely joined words.

ChatGPT is excellent at answering questions about Western music theory, but in some cases the initial answer needs prompting, especially when accounting for enharmonic equivalents.

I can't read your screens that you display. Is there a way that you could provide so that it is easier to read for those of us with accessibility challenges? Thank you. Thank you for all your contributions.

I speculate it does have a “global plan” of what to say next, instead of one word at a time. It implicitly has a representation of the joint probability distribution of what’s to be continued… Prompting kind of bring out that distribution… which you can extract knowledge, in current its form, some piece of text (but may be other modalities in the near future). i was convinced by Sutskever’s take more.

Many thanks Stephen! I absolutely enjoyed the step by step introduction into the layers of the matter. However it is obvious that we are still on the technical/mechanical side of the whole journey. Still none is able to explain the concept and reality of infinity, or "1" or "0", but an honest struggle towards that wisdom may open new paths in learning and lead to brilliant discoveries.

What comments would you make about notable observations between different culture's outputs given similar topics as inputs using ChatGPT 4?

Great! Superb lesson. Thank you! However, I felt confused at 56:58 when Stephen says "At every step we are just collecting the values of the neurons from the previous layer, multiplying them by weights, add a constant offset, applying that activation ReLU, to get this value -3.8". I think the numerical values next to neurons are before applying the ReLU, otherwise they all have to be nonnegative. And the last layer does not apply ReLU in order to get the -1 attractor. Am I correct?

What I find interesting is how they inject the objective pattern recognition into the model to aid in figuring out puzzles and riddles. It will provide extensive reasoning on how it arrived at its answer. GPT-4 really excels in this ability and has a great sense of humor to go with it.

1:18:58 the art of training 1:32:25 training LLMs 1:52:47 attention 1:56:20 attention blocks 2:08:57 neurons have memory

By god this is such a good explanation, thank you

This was a fantastic video to watch

This is a LONG video truthfully. But very informative as it should be with the length of it

I noticed while using ChatGPT that it doesn’t use underline/italics/bold for emphasis, could they in the future include that to relay some emotion back from ChatGPT maybe? I have seen “!” used by it for that.

very compelling, I like your take on how there's a, sort of, throttle in everything. never thought trying to understand AI would be so much fun...

This is fascinating! Any chance there's a Cliff's notes or something?

Should be awarded a NOBEL prize for this tutorial. Well played.....

Greatest primer/teaser for genetic algorithms and neural networks that I've seen Thanx

In semiotics, there is "pragmatics," as well as syntactics and semantics. Communication is purposeful, about something or someone, rather than about predicting the next word. There is always some motivation behind utterances, which in turn are predisposed by "beliefs" and "values." So there is cognition, yes, but also affective and psychomotor skill domains involved. When "natural language processors" begin to present (and see) themselves as agents with a memory at least of their own experiences with you (and others), conversations will be more interesting. As more and more complex ideas and concepts (gained, say, from reading text with attention to those, rather than to words per se) can be mapped in some "meaning space" which can be transitioned in the direction of the allegorical, metaphorical, or hypothetical (and thus creative), the closer we will come to AGI.

This isnt just a great lesson in AI, it is a lesson in how to be a good teacher. (Start with simple concepts students can grasp and only then build up.)

I use it a lot to help me write and fix code and also to explain things for me or piece things things together. It's a great partner/tool to use if you have some good input and existing knowledge

This is super interesting and I’m learning a lot, thank you for this video. I do feel the amount times I hear “ugh” and “um” is really off putting. Sorry if that’s nitpicky but I almost can’t make it through the beginning because of ugh. Um. Ugh.

Can the wheights and biases be complex numbers insteadnof reals? And what effect does it have on performance?

Who would have thought that conversation was a slightly random walk through probable clumps of letters and words? Fascinating. I have to say, though, I think it's actually the human reinforcement that gives particular clumpings their perceptible meaningfulness.

The overall outcome would be exciting of when this goes final and applicable as a worldwide platform for learning and everything else. For now it's too EARLY to tell.

Great video Stephen!

Any comments on the following which has been generated by Chat GPT which seems to disagree with your suggestion that Chat GPT does NOT typically chose the word with the highest probability...."? When you say that Chat GPT uses an "argmax strategy, where the model chooses the word that maximizes the probability according to its "internal calculations"", is this saying that the model always choses the word that has the highest probability of succeeding the initial sequence of words? ChatGPT Yes, that is correct. When I mentioned the "argmax strategy," it means that ChatGPT typically chooses the word that has the highest probability of succeeding the initial sequence of words. During text generation, ChatGPT calculates a probability distribution over its vocabulary, which represents the likelihood of each word being the next word given the preceding context. The model's internal calculations assign probabilities to different words based on the patterns it has learned from its training data. The "argmax strategy" refers to the approach of selecting the word with the highest probability from the calculated distribution. In other words, ChatGPT chooses the word that maximizes the probability according to its internal calculations. By selecting the word with the highest probability, the model aims to generate text that is more likely to be coherent and contextually appropriate. This strategy allows the model to make informed decisions based on the patterns it has learned during training and choose the most probable next word based on the given context. It's worth noting that while the model often favors the most likely word, there can be cases where ChatGPT introduces some controlled randomness or diversity, such as using temperature scaling, to produce alternative outputs or explore different word choices. However, the primary approach is to choose the word with the highest probability based on its internal calculations.

I really like how he explained everything. Oh, how I wish I didn't sleep during math class.🤣🤣

i fall asleep once, and this is what I wake up to? aha

This is a video of Stephen Wolfram preparing to make a video. Through laziness or distraction, he did not make the actual video. The most frequently used word in the video is "um." This cleverly demonstates the point that if ChatGPT simply used the most frequently found word in every situation you would get very bad output.

I have a rule for writing text always choose the word that eliminate the most other words first I am writing a plan i will start this way: my plan is ... If I gone write a essay I will start with: this essay is about ChatGPT...

2:14:02 Stephen says "verbs and nouns go this way..." and shows the old - very old! as old as Aristotle - idea of context-free phase structure grammar based on subject-predicate - but that's a wrong idea, because the real grammar of English is both simpler and more elaborate. *The Natural Topology of English* includes subject-predicate as one of its basic forms, but there are others, such as event = agent + action + object which is an instance of concept - relation - concept

One aspect I disagree with from Steven's perspective is that the reinforcement learning feedback loop step it's not actually a major piece of the success of ChatGPT. You could create a very similar version using contextual memory buffers with the raw davincii-003 model. The RLHF just fine tuned 'good' responses and probably more importantly weighted some of the negative/moral issues with certain things you could generate. There's obviously been an additional, further layer of moderation bolted onto the top, for obvious reasons.

我很樂意看到更多像這樣的關於利用博士的當前 LLM 主題的深入分析。

The onion rooting protocol isn't as anonymous as you think it is. Whomever controls the exit notes controls the traffic. Which makes me in control.

Oh...and finally Stephen makes the statement "That pattern of language has occurred before". No, I dont think so. The implication is that the probability given by the weightings leading to the next word can only have come from seeing the previous word and adjusting the weightings on that combination alone but thats not true. All the weightings including the ones leading to the next word have been influenced by many words from many sentences back propagating. I dont think its necessary for the next word to have been seen before for that pattern to emerge.

So fascinating! I guess one major difference between the way the human brain and GPT handle language is that human brains use emotions to categorize objects and concepts… I wonder if it would be possible to teach GPT emotions, and what might be the result?

The fact that high dimensional spaces are unlikely to have local optima just reminds me of Pólya's random walk theorem.

2:50:30 Does that mean we could play the natural role of the AI's Brain stem, where we are not as conscious as the AI but the AI still works to understand and aid us?

Thank you so much !

The thing I like about ChatGPT is, you can tell it some information and then ask a question and it can get it wrong, but you can then say, no you got it wrong. But if you figure out its break in logic and explain to it why it got it wrong and what assumptions it made that was wrong, and correct that, it learns. Do that enough times and you can break down any concept, no matter how nuanced and complicated. I've done this. It works. But I only used ChatGPT for one day and never since. Why? Because it's not capable of any truly new and original thought. It can only spit out what we already know. So if the world thinks lemmings jump off cliffs, then so does ChatGPT. Again, you could dig down into it and ask why it thinks lemmings jump off cliffs and show its assumption are unproven, but that's no better than talking to a human and there are over 8 billion other natural ChatGPTs on this planet which already do that. At that point, I lost interest. It's like a boat without a rudder.

Would be nice to see the code structure.

Thank you. That is as sentient as it can get, what difference does it make what path it took to come to a certain answer. as long as it can contexualize the question I am asking it and find an appropriate answer to it, then we can eventually put it in a robot and it could act based on interacting with the environment and act based on the responses that comes up with. This might be a mix of GPT10 and certain prediction models and etc, but this is definetly the core of sentience

Great info Thx!👍

Fascinating. Love the demystification. Like unpacking how the greatest magic trick in the world is accomplished. Inspiring! Idea generating! Thank you, Wolfram!

*I wonder how our Neural net AI compare to other living multicellular intelligence: Plants root neuron analogs, fungi's action potential networks, basic cnidarian nerve nets, bacterial biofilms etc.*

It'd be interesting to see what a neural net trained on the archeology data of humanoids would output.

I still don’t understand how it’s able to generate a really elegant Boost ::spirit C++ parser, and a compiler/evaluator for a non trivial language I explained to it

Well said👍

When will we see the next obvious hybrid system: ChatGPT + WolframAlpha/Mathematica ?

Has anyone considered the parallel to how people compose writing, and a process of selecting a word that feels right in the context to add to the string, and head, once the theme is developed, to bringing it to conclusion? That's what AI is now learning to do, so it appears that the "temperature" is NOT random at all. That may be accidental, of course, but something is inputting a sense of feeling about a context (if you take my words for their root meanings and understandings).

but does this explanation explain why it seems to remember (model) so much? Often factual... - It can summarize something that it's never seen in the training data. How can there be "semantic grammar" of me giving it a long piece of code, and it very well describing me what it does, for what purpose, and even non-obvious things that look like understanding ... - It definitely has trained predictive models of ... something. And I can see how one could call those models "semantic grammar" but ... for reasoning/logic/poetry/rhetoric, I can understand that it could fit some "grammar" - but how can "comprehension" and knowledge and apparently realizing meaning behind words be ... "grammar" ? - Are all the things we "made" really so regular? As well as what is real ... can so much about thinking be reduced to LOOP FREE, mindless pattern continuations? It would make sense that we use it a lot to simplify working and making sense of the world, but ... 2:36:15

Humans can count parentheses or perform complex mathematical operations. All we need is time and external memory, such as a piece of paper. ChatGPT can't even split a string of random letters into even portions of four characters - a task that any child can handle. It fails even with very simple combinatorial tasks. So, is this a fundamental issue, or might it be possible to use the model in a similar way as we perform calculations - with iterations and storing the results of the previous iteration?