Thanks for watching, would love your thoughts below. NO MUSIC version: ua-cam.com/video/DFDOyMZw9Q4/v-deo.html Watch Full AI Series: ua-cam.com/video/YulgDAaHBKw/v-deo.html Sponsored by Brilliant | Use brilliant.org/artoftheproblem for 30-day free trial and 20% discount
Love your videos! Just a little suggestion. Background music is good, but too loud. It should never cover your speaking, like now. I suggest 15 - 20% less music volume, and you are good!
I love this video. I remember watching your videos like 10 years ago on Khan Academy about compression, entropy, Claude Shannon, etc. All timeless. I have always loved this style of documentaries. We need to protect you at all costs.
I love your video aesthetics, how you blend retro video clips with your explanations. I think you'd really enjoy retro-futuristic concepts and games like Bioshock and Fallout.
Albert Einstein said: "If you can't explain something in a simple way so anybody can understand it you don't fully understand it yourself". Perhaps you are one of the few LLM experts we have!
anothr great video. Understanding the "world model" and the algorithm that makes the decisions in it was very expansive. Also adding the self training/emulation of dreams is a powerful analogy to the human. seeing how thinking longer, blended in with intuition to make better chains of thoughts is also fantastic. Every time I reflect on machine learning, I learn more about myself. Which kind of makes you think its more sentient if it reminds me of myself? or the best emulator ever!
It is a great explanation of how current AI models reason. I liked the video a lot! 1. Simulation of future states. 2. LLMs that can give kind-of accurate answers with step by step reasoning. 3. RL approach that makes LLMs to give multiple answers, then evaluate them to select the best one. (Required more time) It would be nice to see wether a model that wasn't trained on the internet data, could learn how to reason by interacting with an LLM, and practicing on its dreams, but maybe we'll see that in the future. For the awesome review, history explanation and divulgation: Thanks! 🎉
@@Farligefinn You know what, I just wrote a paragraph disagreeing with you but I reread the initial question and deleted it. Rereading and reinterpreting "Does it think actually?", I actually see what you're saying. A clearer word would have been "reason." "Think" can have a few different interpretations and I was contributing it towards consciousness. But whether AI "reasons" is a very different question entirely and I believe what him and you mean. Does it go through a sequence of logical steps from premises to a conclusion? Does it use deduction? This is what was meant.
@@whitb62 Do you all forget that an AI is literally just a mathematical algorithm that does things as it is told? No amount of complexity will change that. On the other hand, consciousness has been proven to be non-computable.
THANK YOU for publishing a no-music version of this video (see pinned comment by ArtOfTheProblem). It is such a clear and informative video that I hated to see it loose views due to the competing sound track. I'm going to watch it again right now to see if I missed anything the first time around. Thanks again for being so responsive to your followers.
@@ArtOfTheProblem I for one was looking up some "how does AI work" stuff yesterday and some of your vids came up a couple of times, I watched multiple authors with their own unique takes (3Blue1Brown and Nottingham Uni's Computerphile also good channels). This video made me follow tho. I think you earned it :3
Having read a bit about the AI safety arguements, learning about these arguably incredible developments into artificial minds is now accompanied by a sense of dread as well as the sense of awe.
Amazing explanations, visuals, and historical context! IIRC MuZero trained the policy and value networks (used to rollout the MCTS tree) also on the output of the MCTS tree. This seems super useful because search can be used to improve the the training of the networks (not just the results at inference time). I wonder if this also works for CoT/ToT in LLMS where the pretraining could include ToT to boost training performance?
yes it did, and yes it seems to help. Look at inference time training, just a few days ago a group got a new record on the ARC test doing this kind of thing (i haven't had time to go deep). x.com/akyurekekin/status/1855680785715478546
Great video as always, cant wait for the next ones! Top research quality. I think world models deserve more focus rather than llms, which are probably a dead end to true understanding of the real world. Yann lecun has very interesting ideas about these, in his JEPA and V-JEPA architectures and some of his lectures. I also think neuroscience can provide incredibly interesting and valuable insight into ml architectures as why not take ideas from a model undergone hundreds of millions of years of optimization for the same very abilities we are trying to model. Maybe memory is an interesting pathway (perhaps for a video), both working memory and long term (episodic, semantic)... Anyways, just some of the ideas I've been thinking about recently.
One tweak that would help this video perform better is to decrease the relative volume of the background music, especially at the end right before the ad. But it may be too late for that on this one, idk how UA-cam works.
The information and animations are both excellent but the music overpowers your audio. Either lower the volume of the music or get rid of it completely, please.
I would love to see a video about the ethics of machine learning models and especially LLMs. There is a healthy body of literature out there to draw from about issues like intellectual property and copyright, enabling and obscuring bias, impact on marginalized communities, the resources used by model training and computation, etc
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
@artoftheproblem I agree! I love the music. But maybe if you just lower its volume compared to the narration, then you might appeal to more people without losing those of us who like the music (but not necessarily its intensity). I think ones who complain might just be easily distracted by the soundtrack’s loudness rather than hate the music choices.
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
thank you! I studied both in school, and naturally land somewhere in the middle....bad at both! I enjoyed algorithm design, but what Iove most is putting on a 'show' whether movie, play, product or haunted house :)
@@ArtOfTheProblem Thanks for reply. Well about AI - think we sould call it just statistical machines or dynamic patterns parsers. I am really skeptical about non text machine learning - we still have not solved fly brain problems - scientists have fixed 3d map without undestanding how its works - it like mapping intel cpu - and still having knowing nothing about ALU register memory, gates.
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
Here's a puzzle: Do all people reason or do many only memorize patterns? Even people who definitely do reason, do they always reason or do they also just memorize patterns most/much of the time?
That's a wonderful question Andrew. I'm a cognitive scientist who is watching the emergence of LLM-based AI with that very question in mind. The fact that LLMs can come so close to our own cognitive abilities is usually viewed as a sign that AGI is almost here. But it can also be viewed as a demonstration that human cognition itself is nothing more than the repetition of learned patterns with minor variations. In one case we'll be thrilled by how clever we are to have reinvented the awesome capabilities of human intelligence. In the other, we're more likely to be humiliated by the realization that we are, essentially, repetition/prediction engines. The reality almost certainly falls between the two, but as someone who has studied human intelligence his entire life (in and out of academia), my bet is that we are much closer to repetition/prediction machines that we'd like to admit. I'd love to find a deep discussion of this issue. Maybe a future video in this series (hint, hint)?
I'd argue humans don't tend to rely on either very often. Instead, humans tend to think very heuristically. Deductive reasoning and memorization/recollection are really only required for very precise tasks. Instead, our brains learn a very general feeling of how to do things by strengthening neural pathways that are used repeatedly. Even humans who try to act very logically are generally heuristically feeling their way through tasks, occasionally thinking through algorithms that have been "memorized".
I agree :) also If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help!
4:19 Could you elaborate on which hand-coded formulas used by Shannon with TD-Gammon in the year 1989 you are referring to? Also, when and how did Shannon work with TD-Gammon? "And so, the first key breakthrough in machines mimicking intuition for position quality came when neural networks replaced the hand-coded formulas Shannon used in 1989 with TD-Gammon"
Yes! I made a whole video on this you can check it out here: ua-cam.com/video/Dov68JsIC4g/v-deo.html - please let me know if you have questions after watching. Shannon didn't do TD Gammon Tesaruo did. enjoy
It seems like some of these developments regarding world models should have huge implications for robots that can function in a human centric world. I think we’ll see an explosion in development of robots that can help humans with everyday tasks and a robot that can be a useful household assistant will be a reality in the next 10 years!
Context length is problem that's the main reason models needs to keep becoming bigger Or you could train a CNN inspired architecture where a model is shown some sliding window and they produce some token which is repeatedly given to it as input at last when the output is small enough to be taken as input for a full context model it is used like gpt Claude etc Or you could also use RL and mutate or find a js code capable of generating code, js is so abstracted it's perfect I made a small programing Language with hoisting such that sequence of process doesn't matter and simple Santax that local minimum escape problem is solved and I wanna train a model If I get a model I will than continue training else I'll do a dev log video eventually I'll get worlds first infinite context Model
Thanks so much for these, I had no idea about some of these approaches. I’m wondering now if anyone’s tried applying muzero to arc, since the challenge of arc is learning implicit rules from just a few examples
Constructive criticism: 1) The substance of the video was very good. Script was well written, delivery was ok. A bit monotone but not that bad. 2) Sound design was poor towards the end. The music drowned out your voice, and the lyrics were both distracting and discordant. 3) Your choice of clips, footage, and visuals was good. The video was informative when needed, and abstract/entertaining/interesting otherwise. 4) The narrative structure was okay. It was a mostly clear progression. At the end it became unclear which AI was doing what strategy. 5) Visuals were reused way too often. Visuals can be reused, but I think the brain wormhole clip was shown 6 times, way too many. 6) Beware over-using a metaphor image. The upwards shot at two trees was reused so many times as a visual for tree-like thinking that it just became annoying.
Please, invest in a decent microphone. It's brilliantly presented, but hard to hear well. The music track is not ducking either so your voice and the music compete for the same ears.
To me ai is just some linear algebra and some complex algorithm that follow order and the things is human only need few examples to learn meanwhile ai need a massive database of object and image to "understand the subject"
@ArtOfTheProblem edit: I'm pretty sure in the future a lot of people will be fired and replace by those "ai" And well literally the people that use the ai and also I get what you mean
This is a very hopeful video. There are billions of dollars being poured into bringing the resources to hand, to find an effective approach to AGI... Once AGI really kicks in, the acceleration of progress bounded only by our imagination will be something to behold. Absolutely awesome. I hope it leads to a world of abundance where we have no need for psychopathic power seekers. 🤞
@ArtOfTheProblem maybe something in response to the 5+ hours of Anthropic interviews on Lex Fridman... I'm sure that might inspire some topics? Sam Altman rarely gives any insights to what OpenAI are doing, Mark Zuckerberg is equally vague. I think that interview gives more of an insight to the direction of travel.
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
thanks for sharing, I feel same way. the music is part of the original idea for the channel...a feeling. but because people can get distracted I think i'll post music free as optional from now one.
Maybe it's not only the loudness, but also the choice of music that is distracting to some. E.g. at 2:00 I don't feel distracted, but at 15:00 very much so. Anyways, thanks for making these great videos
The methods used to train mu-zero seem to be conflated with the use of chain-of-thought methods for LLMs here, which tells me that this channel has gotten sloppy. Like, really sloppy. With self-play and world-models, the weights of the model are changed by some external trainer after each round. With chain-of-though in LLMs, there is *literally no learning happening.* No weights are changing. No reasoning from one problem will be kept for future problems. Maybe Mister Gippity can reason through one problem if you explain it, but you *will* have to explain it *again* in a new session. I expect "chain-of-thought" works because transformer-model LLMs have no internal feedback mechanisms in the way that older RNNs like LSTM models do. My understanding is that that fact is what has made transformer models so effective - they're easier to train at-scale when you don't need to take into account all previous states, essentially just giving it all previous inputs instead of training based on "what it might have been thinking at the time." But the result is that it is literally incapable of self-reflection, and the only way of recovering that feature is to give its own output back to it as input, which is what CoT does. CoT isn't some spectacular emergent behavior, it's just a workaround for some features that were removed to make training more efficient. But why should that feedback mechanism take the form of human-readable text? That sounds horribly inefficient to convert between "thoughts" (latent spaces) and English-text and back again especially when the "reasoning" that results cannot be applied to other problems. Because again... the weights are *not* updated after solving a problem. That's the "P" in "GPT." Sure, these "ai" companies will save your chat logs and use them to train and update their weights, but that's just training it on text that gets it wrong and has the corrections explained to it... which will lead to it continuing to get things wrong, expecting corrections to be explained to it. The "ai emperor" has no clothes, as far as I can tell.
Thanks for sharing. I'm not conflating MuZero's training with CoT. Rather, it's drawing an analogy between search strategies - both use systematic exploration of possible paths before committing to an answer. Also have a look at test time training, this does include weight updates! And transformers do have dynamic weight updates through attention. I'd argue using natural language for reasoning isn't inefficient - it's actually leveraging the model's core strength not to mention explainability....what do you think?
Thanks for watching, would love your thoughts below.
NO MUSIC version: ua-cam.com/video/DFDOyMZw9Q4/v-deo.html
Watch Full AI Series: ua-cam.com/video/YulgDAaHBKw/v-deo.html
Sponsored by Brilliant | Use brilliant.org/artoftheproblem for 30-day free trial and 20% discount
Thank you for the NO MUSIC version!
Love your videos! Just a little suggestion. Background music is good, but too loud. It should never cover your speaking, like now. I suggest 15 - 20% less music volume, and you are good!
This Guy just explained all the core concepts in AI on one shot ,Congrats man!
:) thank you, i cut a LOT out of the video in my edit - going to post a shorter bigger summary soon
I love this video. I remember watching your videos like 10 years ago on Khan Academy about compression, entropy, Claude Shannon, etc. All timeless. I have always loved this style of documentaries. We need to protect you at all costs.
Thank you, I love hearing from og’s! support future work: www.patreon.com/c/artoftheproblem
I love your video aesthetics, how you blend retro video clips with your explanations. I think you'd really enjoy retro-futuristic concepts and games like Bioshock and Fallout.
Love this , I definitely know the style you are talking about
Albert Einstein said: "If you can't explain something in a simple way so anybody can understand it you don't fully understand it yourself". Perhaps you are one of the few LLM experts we have!
THANK you this means a lot to me.
@@ArtOfTheProblem I did but reddit really hates it, it got removed on 4 subs. The internet does not like to get educated anymore man.
is this the last video in the series? regardless, can't tell you how valuable and enjoyable i've found them all. thank you for them.
anothr great video. Understanding the "world model" and the algorithm that makes the decisions in it was very expansive. Also adding the self training/emulation of dreams is a powerful analogy to the human.
seeing how thinking longer, blended in with intuition to make better chains of thoughts is also fantastic. Every time I reflect on machine learning, I learn more about myself. Which kind of makes you think its more sentient if it reminds me of myself? or the best emulator ever!
thank you! ai agree....also you are my "top commentor" according to YT. :)
It is a great explanation of how current AI models reason. I liked the video a lot!
1. Simulation of future states.
2. LLMs that can give kind-of accurate answers with step by step reasoning.
3. RL approach that makes LLMs to give multiple answers, then evaluate them to select the best one. (Required more time)
It would be nice to see wether a model that wasn't trained on the internet data, could learn how to reason by interacting with an LLM, and practicing on its dreams, but maybe we'll see that in the future.
For the awesome review, history explanation and divulgation:
Thanks! 🎉
thanks for sharing summary
Great vid. I love that it clearly explains the progression, like the pieces coming together. Can't wait to see the next steps!
thanks, next up i'm taking a detour into economics
So good! Loved it
Answering the question of "Does it think actually?" is as hard as the question "Are other people conscious like me?".
The hard problem of consciousness.
@@whitb62 Not really the same thing.
@@Farligefinn You know what, I just wrote a paragraph disagreeing with you but I reread the initial question and deleted it. Rereading and reinterpreting "Does it think actually?", I actually see what you're saying. A clearer word would have been "reason." "Think" can have a few different interpretations and I was contributing it towards consciousness. But whether AI "reasons" is a very different question entirely and I believe what him and you mean. Does it go through a sequence of logical steps from premises to a conclusion? Does it use deduction? This is what was meant.
@@whitb62 thanks for the forthright and civil answer :) was about to expect some harsher language that seems to be the norm online these days.
@@whitb62 Do you all forget that an AI is literally just a mathematical algorithm that does things as it is told? No amount of complexity will change that. On the other hand, consciousness has been proven to be non-computable.
"Charging down a path that often lead to the wrong conclusion." Yep, sounds human to me.
@@DisProveMeWrong so very human
Finally. I live for these videos. They are the most fascinating vids ever made. Thanks for keep on educating us further, you are a hero!
thank you I appreciate it
Keep u up the great work!
appreciate it
This is an absolutely amazing video!!!!!
thank you! I was so in the weeds with it i hope it comes across as clear? I tried to strike a balance...
THANK YOU for publishing a no-music version of this video (see pinned comment by ArtOfTheProblem). It is such a clear and informative video that I hated to see it loose views due to the competing sound track. I'm going to watch it again right now to see if I missed anything the first time around.
Thanks again for being so responsive to your followers.
Thank you for saying that , I find the music keeps me interested as I take sooo long to edit
Best Channel I've ever followed.
Thank you! when did you join? Please help post to your networks
@@ArtOfTheProblem I for one was looking up some "how does AI work" stuff yesterday and some of your vids came up a couple of times, I watched multiple authors with their own unique takes (3Blue1Brown and Nottingham Uni's Computerphile also good channels). This video made me follow tho. I think you earned it :3
@@notbfg9000 great to hear, i've been working to try and fix my thumbnails to make them interesting to click on. always open to feedback
@@ArtOfTheProblem No particular criticisms there :)
I don't really pay great attention to thumbnails, but maybe that's not true for most people lmao
Thank you so much. I wish I had the time to give feedback thanks for being willing to open it up
Appreciate the feedback! happy to share
Having read a bit about the AI safety arguements, learning about these arguably incredible developments into artificial minds is now accompanied by a sense of dread as well as the sense of awe.
I love to hear this...well said
I wish i could learn how to think 🤔
Exactly. Reasoning is a skill.
Amazing explanations, visuals, and historical context!
IIRC MuZero trained the policy and value networks (used to rollout the MCTS tree) also on the output of the MCTS tree. This seems super useful because search can be used to improve the the training of the networks (not just the results at inference time). I wonder if this also works for CoT/ToT in LLMS where the pretraining could include ToT to boost training performance?
yes it did, and yes it seems to help. Look at inference time training, just a few days ago a group got a new record on the ARC test doing this kind of thing (i haven't had time to go deep). x.com/akyurekekin/status/1855680785715478546
Great video as always, cant wait for the next ones! Top research quality.
I think world models deserve more focus rather than llms, which are probably a dead end to true understanding of the real world. Yann lecun has very interesting ideas about these, in his JEPA and V-JEPA architectures and some of his lectures. I also think neuroscience can provide incredibly interesting and valuable insight into ml architectures as why not take ideas from a model undergone hundreds of millions of years of optimization for the same very abilities we are trying to model. Maybe memory is an interesting pathway (perhaps for a video), both working memory and long term (episodic, semantic)...
Anyways, just some of the ideas I've been thinking about recently.
appreciate you sharing these thoughts i've been follwoing LeCun as well and hope to do another update once I see more results
One tweak that would help this video perform better is to decrease the relative volume of the background music, especially at the end right before the ad. But it may be too late for that on this one, idk how UA-cam works.
yeah i wish I could, it's locked after upload...i do have a no music version (unlisted link above) thank you for feedback
Thanks for video ❤
appreciate the comment please share with anyone in your network who is interested!
The information and animations are both excellent but the music overpowers your audio. Either lower the volume of the music or get rid of it completely, please.
Music free version in top comment and description
Ah Yoo, I see "Art of The Problem", I click. Easy like that.
:)
PLEASE make a video on memory augmented AI (neural turing machines/differentiable neural computers)
thanks for suggestion, noted! currently watching the field
I would love to see a video about the ethics of machine learning models and especially LLMs. There is a healthy body of literature out there to draw from about issues like intellectual property and copyright, enabling and obscuring bias, impact on marginalized communities, the resources used by model training and computation, etc
thanks for sharing, noted!
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
you'll never please 100% of any audience. 2nd law of conquest is a thing. keep doing your thing, your music is as iconic as vsauce's is to theirs
:) thanks
@artoftheproblem I agree! I love the music. But maybe if you just lower its volume compared to the narration, then you might appeal to more people without losing those of us who like the music (but not necessarily its intensity). I think ones who complain might just be easily distracted by the soundtrack’s loudness rather than hate the music choices.
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
I love your channel. Are you programmer or more like mechanical engineer?
thank you! I studied both in school, and naturally land somewhere in the middle....bad at both! I enjoyed algorithm design, but what Iove most is putting on a 'show' whether movie, play, product or haunted house :)
@@ArtOfTheProblem Thanks for reply. Well about AI - think we sould call it just statistical machines or dynamic patterns parsers. I am really skeptical about non text machine learning - we still have not solved fly brain problems - scientists have fixed 3d map without undestanding how its works - it like mapping intel cpu - and still having knowing nothing about ALU register memory, gates.
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
@@ArtOfTheProblem "yes the fire rises" Bane =)
Here's a puzzle: Do all people reason or do many only memorize patterns? Even people who definitely do reason, do they always reason or do they also just memorize patterns most/much of the time?
That's a wonderful question Andrew. I'm a cognitive scientist who is watching the emergence of LLM-based AI with that very question in mind. The fact that LLMs can come so close to our own cognitive abilities is usually viewed as a sign that AGI is almost here. But it can also be viewed as a demonstration that human cognition itself is nothing more than the repetition of learned patterns with minor variations. In one case we'll be thrilled by how clever we are to have reinvented the awesome capabilities of human intelligence. In the other, we're more likely to be humiliated by the realization that we are, essentially, repetition/prediction engines. The reality almost certainly falls between the two, but as someone who has studied human intelligence his entire life (in and out of academia), my bet is that we are much closer to repetition/prediction machines that we'd like to admit.
I'd love to find a deep discussion of this issue. Maybe a future video in this series (hint, hint)?
I'd argue humans don't tend to rely on either very often. Instead, humans tend to think very heuristically. Deductive reasoning and memorization/recollection are really only required for very precise tasks. Instead, our brains learn a very general feeling of how to do things by strengthening neural pathways that are used repeatedly. Even humans who try to act very logically are generally heuristically feeling their way through tasks, occasionally thinking through algorithms that have been "memorized".
Reason takes effort and the brain doesnt like to do that often it switches to pattern recognition and intuition as much as possible
@@sulemanmughal5397 I would go further and say going from reasoning to this is one kind of learning and is also akin to 'muscle memory'.
I agree :) also If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help!
Great video, as always:)
The perceptron is a universal approximation machine. Ai cannot think it can only approximate thought. Ai = approximate intelligence.
Excellent video! You are a born professor! 👍
thanks mom
absolute cinema
4:19 Could you elaborate on which hand-coded formulas used by Shannon with TD-Gammon in the year 1989 you are referring to? Also, when and how did Shannon work with TD-Gammon? "And so, the first key breakthrough in machines mimicking intuition for position quality came when neural networks replaced the hand-coded formulas Shannon used in 1989 with TD-Gammon"
Yes! I made a whole video on this you can check it out here: ua-cam.com/video/Dov68JsIC4g/v-deo.html - please let me know if you have questions after watching. Shannon didn't do TD Gammon Tesaruo did. enjoy
@@ArtOfTheProblem Thank you. I'll watch it.
Thinking is for fools lol, now KNOWING….. knowing is Cool AF😎!
It seems like some of these developments regarding world models should have huge implications for robots that can function in a human centric world. I think we’ll see an explosion in development of robots that can help humans with everyday tasks and a robot that can be a useful household assistant will be a reality in the next 10 years!
thanks for sharing, yes I'm watching this very closely
wow, thank you for this!
appreciate it! stay tuned
Context length is problem
that's the main reason models needs to keep becoming bigger
Or you could train a CNN inspired architecture where a model is shown some sliding window and they produce some token which is repeatedly given to it as input at last when the output is small enough to be taken as input for a full context model it is used like gpt Claude etc
Or you could also use RL and mutate or find a js code capable of generating code, js is so abstracted it's perfect
I made a small programing Language with hoisting such that sequence of process doesn't matter and simple Santax that local minimum escape problem is solved and I wanna train a model
If I get a model I will than continue training else I'll do a dev log video
eventually I'll get worlds first infinite context Model
thanks for sharing
Commenting to help with the algo, and moving to the no-music one to do the same
@@shenrr6802 thank you! I have no music unlisted as to avoid splitting the momentum
Thanks so much for these, I had no idea about some of these approaches. I’m wondering now if anyone’s tried applying muzero to arc, since the challenge of arc is learning implicit rules from just a few examples
@@easlern yes this is happening right now with test time fine tuning !
Once we understand how we reason, making LLMs reason like us is possible.
The video is good but there are sooo many random sounds that make it difficult to focus on what you are saying, specifically towards the end.
Here you are! ua-cam.com/video/DFDOyMZw9Q4/v-deo.html
Goooood
thank you! curious what questions you have after watching this?
Constructive criticism: 1) The substance of the video was very good. Script was well written, delivery was ok. A bit monotone but not that bad. 2) Sound design was poor towards the end. The music drowned out your voice, and the lyrics were both distracting and discordant. 3) Your choice of clips, footage, and visuals was good. The video was informative when needed, and abstract/entertaining/interesting otherwise. 4) The narrative structure was okay. It was a mostly clear progression. At the end it became unclear which AI was doing what strategy. 5) Visuals were reused way too often. Visuals can be reused, but I think the brain wormhole clip was shown 6 times, way too many. 6) Beware over-using a metaphor image. The upwards shot at two trees was reused so many times as a visual for tree-like thinking that it just became annoying.
I like the music but it's too loud
thanks for note
The audio mix is horrific, it's not simply a matter of adjusting the levels
@@retrofitter no music version: ua-cam.com/video/DFDOyMZw9Q4/v-deo.html
Please, invest in a decent microphone. It's brilliantly presented, but hard to hear well. The music track is not ducking either so your voice and the music compete for the same ears.
thanks, I have a great mic, but I do need to mix the audio better which i'll do next time (btw, i have a no music version in top comment)
And that's kids why AI will replace programmers
it's interesting that the programmers I know are as divided as the field is
Thanks!
mucho appreciated!
Simply excellent video, your style reminds me of every frame a painting
appreciate this feedback, I also enjoyed that channel
does language think?
To me ai is just some linear algebra and some complex algorithm that follow order and the things is human only need few examples to learn meanwhile ai need a massive database of object and image to "understand the subject"
Lots of interesting research on learning with less , recent advances such as “learning to walk in 5 min” did u see my rl video ?
@ArtOfTheProblem edit: I'm pretty sure in the future a lot of people will be fired and replace by those "ai" And well literally the people that use the ai and also I get what you mean
The reply works lol 👍
This is a very hopeful video. There are billions of dollars being poured into bringing the resources to hand, to find an effective approach to AGI... Once AGI really kicks in, the acceleration of progress bounded only by our imagination will be something to behold. Absolutely awesome. I hope it leads to a world of abundance where we have no need for psychopathic power seekers. 🤞
thank you for sharing, would love to know what you'd like to see next
@ArtOfTheProblem maybe something in response to the 5+ hours of Anthropic interviews on Lex Fridman... I'm sure that might inspire some topics? Sam Altman rarely gives any insights to what OpenAI are doing, Mark Zuckerberg is equally vague. I think that interview gives more of an insight to the direction of travel.
@@BrianMosleyUK yes I have been catching up on those
If you can help share my new video around any of your networks today it might catch fire and would help me support the channel. I appreciate your help! ua-cam.com/video/PvDaPeQjxOE/v-deo.html
Although I know I’m in the minority, I really enjoy the music. The ambiance created adds to the experience for me
thanks for sharing, I feel same way. the music is part of the original idea for the channel...a feeling. but because people can get distracted I think i'll post music free as optional from now one.
I generally like the music. But in the second half of this video the music is very loud and distracting.
Maybe it's not only the loudness, but also the choice of music that is distracting to some. E.g. at 2:00 I don't feel distracted, but at 15:00 very much so. Anyways, thanks for making these great videos
@@io9021 made a new music version too! ua-cam.com/video/DFDOyMZw9Q4/v-deo.html curious what questions you have after watching this
Hey, what's the name of the song at 16:05? Thanks!
these are all original tracks
16:19 is there a third side that is a bit nicer to the ai's and believes in them, that they are good enough as they are at reasoning? 😂😂😂
:)
I love your videos but the background music is just too loud
see no music version in top comment
I think people too difficult to conjecture computers’ thinking.
Thank you so much, love this perspective
inb4 the dozens of "stochastic parrot" arguments
take cover!
stochastic parrots are just autocomplete on steroids
The background noise is incredibly irritating.
sorry I should post silent versions, i'll do it and link here!
@ArtOfTheProblem good idea
@@mancroft Here you are! can you put this in your comment and then i'll pin it? ua-cam.com/video/DFDOyMZw9Q4/v-deo.html
is this 3blue1brown? 6:04
The network, not the graph
yes! i credited him, as well. if you watch his video it points back to mine
@ArtOfTheProblem really? I was just pointing it out, I found it interesting how you used a youtuber's media who I enjoy watching.
@@Blackhole.Studios yes I always loved that animation and credit to him for taking the time
@ArtOfTheProblem exactly
Grat Video but the annoying music makes it hard to follow
see music free version in top comment
Quantum computers + AI + satellites
its recording you simple as fart
The methods used to train mu-zero seem to be conflated with the use of chain-of-thought methods for LLMs here, which tells me that this channel has gotten sloppy. Like, really sloppy.
With self-play and world-models, the weights of the model are changed by some external trainer after each round. With chain-of-though in LLMs, there is *literally no learning happening.* No weights are changing. No reasoning from one problem will be kept for future problems. Maybe Mister Gippity can reason through one problem if you explain it, but you *will* have to explain it *again* in a new session.
I expect "chain-of-thought" works because transformer-model LLMs have no internal feedback mechanisms in the way that older RNNs like LSTM models do. My understanding is that that fact is what has made transformer models so effective - they're easier to train at-scale when you don't need to take into account all previous states, essentially just giving it all previous inputs instead of training based on "what it might have been thinking at the time." But the result is that it is literally incapable of self-reflection, and the only way of recovering that feature is to give its own output back to it as input, which is what CoT does. CoT isn't some spectacular emergent behavior, it's just a workaround for some features that were removed to make training more efficient.
But why should that feedback mechanism take the form of human-readable text? That sounds horribly inefficient to convert between "thoughts" (latent spaces) and English-text and back again especially when the "reasoning" that results cannot be applied to other problems. Because again... the weights are *not* updated after solving a problem. That's the "P" in "GPT."
Sure, these "ai" companies will save your chat logs and use them to train and update their weights, but that's just training it on text that gets it wrong and has the corrections explained to it... which will lead to it continuing to get things wrong, expecting corrections to be explained to it. The "ai emperor" has no clothes, as far as I can tell.
Thanks for sharing. I'm not conflating MuZero's training with CoT. Rather, it's drawing an analogy between search strategies - both use systematic exploration of possible paths before committing to an answer. Also have a look at test time training, this does include weight updates! And transformers do have dynamic weight updates through attention. I'd argue using natural language for reasoning isn't inefficient - it's actually leveraging the model's core strength not to mention explainability....what do you think?
Dislike man beceause I ask help from this "AI" to solve a problem=but crash and left the chat🤬🔪🪓