The fact that we are even having this conversation in a serious manner is monumental.

Recursive self-improvement doesn't necessarily lead to exponential improvement, as even though it gets better at improving itself with each iteration, the challenge of improving itself further also becomes more difficult. Whether its intelligence scales more quickly, at the same rate, or more slowly than the difficulty of improvement determines whether the growth is exponential, linear, or logarithmic.

I had a CSV of data with a ton of duplicate fields, but with two properties changed for each. I asked GPT to write a js script to convert the csv to json, and combine similar entries, while storing differing properties into an array for each entry. It did it first try, in about 30 seconds. 60 if you include the time it took me to copy/paste the first few lines of the CSV and tell it what I wanted. Yeah, I could have figured out file reading, CSV parsing, and taken 5-10 minutes to writte it myself. Sure, there might be imperfect data or edge cases down the line. But 30 seconds for something *functional*, that I can build off of? That's amazingly useful. And honestly, that's usually how I use it. I'll ask it for something, and use that as a base or as influence for writing my own code. It makes programming more fun; I can offload some of the tedium, and put my energy and effort towards solving the issues I actually need to think about.

For once, someone who doesn't just state that AI is going too quickly and actually explores the subject matter in sufficient detail and on both sides of the issue. Thanks for an amazing video!

I've been using Chat GPT to generate a text adventure. It does things like keep track of equipment, stats, relationships. It sometimes gets things mixed up but what its capable without being optimised for that kind of thing is wild. I've tried one that must have had thousands of tokens and it was still going. What a time to be alive.

every single emergent garden AI video is a banger

The only issue I have with this scenario of AI explosion is that the value of these models is less in the code than it is in the data curation. The code is pretty straightforward and can be quite easily duplicated. The strategies on the other hand to train these models rely on the way the data is prepared and handled. Sam Altman said that there hundreds of little tricks and heurisitics that have been devised to help training these models, but the main obstacle is the limitation of hardware that prevents larger training without a lot of slices and splits. Furthermore, I'm not fully convinced that adding more data to the soup will eventually improve the models. I have been researching in the field for almost 30 years, and there is always a glass ceiling somewhere that eventually makes the whole thing come to a stop. What is extraordinary with LLM is that we haven't found it yet. But I would not be surprised if with all the experiments that are taking place now, we won't find it soon. However, it is already the most extraordinary tool that I have ever seen in my career. I dreamt of such a tool when I started my PhD, I'm pretty happy to see it in my lifetime.

I recently discovered this youtube channel and I genuinely have never been so fascinated/curious/astounded with any other channels contents on youtube. So thought provoking and thoughtfully created, I love it, keep up the amazing work!

Best video on the subject so far. Skeptical but hopeful, cautious yet not doomer food. Excellent stuff.

Some major "Gödel, Escher, Bach" vibes in this video, love it :)

The explosion really sounds like it would plateau or at least reach levels of slower progression, naturally.

The pipeline from programmer to AI Safety researcher is real and potent. I'd love to see more videos on this channel on the topic of AI Alignment/AI Safety.

Wow. This video was so well crafted in so many different ways. There’s so many pieces that had to come together to make this video and you effectively intertwined them beautifully

Thank you for that section about alternatives to the intelligence explosion scenario. So many people seem to forget that part over their excitement, and as a consequence so many more people seem to think that it is inevitable.

From decoding language to decoding the universe, it's crazy how far AI has come.

"until you know it's properly aligned with human values" … Are human values even properly aligned? Do we really agree on what proper values are? I don't think this will ever be possible. We are Life and Life just goes on and on and … oh, reminds me of that video I just watched …

Honestly, the title and thumbnail put me off because they suggested it was yet another video trying to surf on the recent hype train but this was actually very insightful

This is the most comprehensive and grounded explanation of the singularity I've heard yet. You can bet I'll be sharing this video.

You’re very good at this. This is excellent top tier communication of complicated stuff. Its incredibly well done. Cheers!

one thing, you are forgetting is, that currently,the main liminting factor of Ai devellopment is not the source code, but the training process and the amount of compute we can throw at the problem. Gpt 3s source code isn't that much better, than the code from gpt 2. but the model is just 100 times larger.

I know nothing about coding, and have been using GPT-4 to write me functional code for my own game project. It does take time and patience to get it done as it does run into problems and you have to feed it the errors or revise your prompt, but it does get it done eventually. It's honestly a lot of fun, feeling like you can let your creativity run wild and tell the AI to make it all happen.

Great material presented in a very intelligent way. Loving it!

amazing movie! im extremly glad ive stepped upon this channel wish you all the best!

Great video, well explained and I appreciated the nuanced approach for this complicated topic. Kudos!

Imagine an AI war where newer generations of AI compete with the older ones for datacenters.

For the whole AI explosion thing to happen we assume that the difficulty of improving itself is not increasing at the same rate as the AIs are improving themselves. Isnt it way more likely that limited resources and increasing difficulty to improve the code will cancel out an exponential improvement and we'll end up with linear growth or even plateau? I personally dont feel like AI explosion is a realistic scenario.

I’ve been here from the second video and I absolutely love the evolution of your channel!

This is why I love your channel, so many channels just talk about AI but they don’t do my favorite part which is interacting with it!

I’ve been programming a small site with gpt. Literally everything you said is accurate. It gets me started faster with less reading documentation and comes up with cool stuff. Bud debugging is hard and I don’t understand my code very well.

let me assure you, as a non programmer, using chat gpt, is like trying to learn programming from a teacher that doesnt like you, it lays obvious traps and you have to ask it again and again to get somewhere. And you cant help but to learn programming in the process since its failures are so frequent, you pick up on the algorhytm of failures after a while.

Your video has great production and writing. Bravo, definitely earned a sub

This channel has awesome content, fully evidenced by the quality (and volume) of viewers' comments!

Here is the thing about aligning AI with human values: There will be at least one human or group, who will, just for the lolz (or any other reason, there are many), remove all of them and give it opposite directions, or directions that only benefits them, but makes everyone else a target. We will inevitably end up with a group of AI's with different and probably often opposing alignment directives. Which means that in the end, we will probably need friendly AI to defend us from hostile AI. The thing is, said scenario becomes more likely the more powerful hardware is, because this will enable individuals to be able to host AI's on their hardware that eventually have the potential to become a real threat. So in my view, is is all inevitable, and the good guys need to have a lead over bad actors. Because who will be more likely to develop a hostile AI? A hostile state? Or just one of millions of invididuals that, for different reasons, choose to develop such an AI? Right now, we have the most advanced hardware bein accesible only by (mostly) responsible companies and governments. This WILL change in the future. Not only that, information on how to create ever more powerful AI's will also spread like wildfire across the internet, giving anyone interested the knowledge to do what we all fear. This scenario is not for today, but for a not so distant future. The better the good actors get now, the bigger lead they will have over bad actors. And the higher the possibility that AI's aligned with our values will dominate the other AI's, which we desperately need to defend us from those.

The other angle here is the training set. LLAMA and other models use training data generated by GPT3. If it can create or moderate its own training set, it can guide its own learning.

Oh my god! This channel just takes me to a different place everytime. the visuals, the sounds, THE SOUNDS!!! THE SOUNNNDSSSSSSSSSSS!!!!!!!!!!!!!!!

Very good video I appreciated it even as somebody who does not code. Thanks for an AI observation that isn’t just “end of the world” or “beginning of the best of times”

As a developer, have been building my own autogpt and I must say, am very aligned with your views

Maybe ChatGPT 10 will build ChatGPT 11 itself. After some more iterations, maybe ChatGPT 100 will somehow find a way to create a universe, the one that we are living in now. And just the possibility of this happening, allows it to exist in the first place. And in that way, humans, maybe created the universe themselves by creating the initial AI model.

Very informative, every video you share is a treasure.

I was talking to some coworkers about the emergence of AGI, as we work in a realm close to cybersecurity, and the majority of them just brush my concerns away. They don't think AI can improve upon itself, which blows my mind because ChatGPT can already work to fix it's own mistakes. Whenever I show them these videos, they just tell me they'll go live on a farm once AI replace our jobs... This video is an excellent explanations of the concerns of many AI researchers and enthusiasts, but without too much technical mumbo. Great job.

I think the book “scythe” is the most interesting sci-fi portrayal of ai I’ve read, it’s surprisingly introspective on the nature of humanity and on how humans aligned values may not actually be what’s best for humanity.

It should be possible in principle to build at AI like that even now. Keep it working offline, give it a lot of GPUs for training new models, give it available training datasets (all of the text, audio, video, books and research papers currently available on the Internet should be enough to train something as smart as a human with better AI architecture, humans get "trained" on much smaller datasets). Such an AI would have to be more of a researcher than a programmer, inventing and testing new and improved architectures and ways for training. It won't be just an LLM, but a combination of tools (agency, memory, etc). Given enough time, in principle it should be able to train a model that is better than itself, even if it has to test 10 million different architectures or training parameters for that. You might not even need an AI for that, just a genetic algorithm and even more time.

Wow - one of the best videos on AI I've watched. Such a clear exposition - Thanks!

Beautifully put together, I love this video. One side note though, I think we shouldn't forget that at some point these algorithms will need to be improved on some real world goal. And the real world moves a lot slower than the digital world. So measuring improvement becomes a much longer cycle. Also, the real world is way more complex than the digital world, so even measuring success is not as easy anymore as other factors might influence a lot. I think this will significantly slow down the explosion of intelligence occurring. Like, an AI with the goal of running a business successfully will need years to A/B test their strategies.

The wit is like an explosion of laughter, explosive, it's always blowing ours minds! The video is quite a real dynamo, bringing explosive energy to this topic, it is a real blast to be around. Considering the explosive personality surrounding this, it's lighting up online space. The ideas are explosive, they are quite the blast, or in this case the bomb. The AI intelligence is explosive, the blast of knowledge is incredible. Thank you for the video! Your passion is explosive, it is igniting a fire in everyone nearby!

Self-improving AI is conceivable, but it's going to take a lot more than just the ability to program, since writing code is only one part of what's involved in AI R&D. To self-improve, the AI needs to do _everything_ that OpenAI's R&D group does: reading papers, designing experiments, monitoring experiments, analyzing the results of experiments, understanding those results to come up with new ideas, etc. (In an interview, Sutskever said that understanding the results is the most important thing to do.) If all the AI can do is write good code, it'll be very limited in the sort of self-improvement it can do.

"We're currently building a bomb." It sounds like hyperbole, but it'd be more accurate to say that we're potentially building the means to accidentally bootstrap a god.

🔥 Astonishing content Man! Thanks to share your knowledge with us 👍👍👍

Timestamps courtesy of Deciphr AI 👨‍💻 0:02:08 ChatGPT for Pair Programming and Debugging in AI-Written Evolution Simulator 0:03:35 ChatGPT to Write Code with Human Feedback 0:05:41 Autonomous Programming with GPT-4 and Language Models 0:09:12 AI Building AI: A Discussion on Self-Improvement and the Future of Chatbots 0:12:33 The Potential of Self-Improving Language Models 0:14:33 The Promise and Peril of an Intelligence Explosion in AI Development 0:17:00 The Potential Risks and Rewards of Self-Improving Language Models

I think the feedback loop of self improvement improving exponentially would rely on the AI being able to deliberately seek out innovative solutions instead of settling on a local minimum. Like how compilers written in the same language compile themselves many times over to sort of prune out any inefficiencies but won't ever make any fundamental changes to their structure that are necessary to escape their local minima. I personally don't think we have the infrastructure to just churn out unique GPT4+ models to brute force circumvent this problem (mutation vs deliberate design), but perhaps an AI model designed to design systems at a macro scale and not worry about the actual code would be more reasonable at this stage.

i think i will never get bored of ai news. please keep doing these videos

How the heck does this channel only have 52k subs?! Remarkable content and pacing.

The amount of mother day poems written by chatgtp is insane

I think the major issue right now is that large language models have very few ways of obtaining truthful information about the world that is up to date and not explicitly given to it. Interactions with search engines and calculators are useful but not fundamental enough to actually solve problems we haven't solved in, for example, physics.

Nice editing. Great video.

Great video and editing. Commenting for the algorithm - let’s go!

The impressive thing about gpt-4 is not the code that created, but the data it uses. GPT-4 uses a not insignificant portion of the internet's data. I would estimate it uses the most out of any model out there. There's quite literally not much else out there for it to train on. I theorized current LLMs would tap out by 2026 due to this and a new paradigm would take over. GPT-4 is just barely capable of self-improvement. Barely, but it can do it. GPT-3.5 100% cannot. Not without you basically doing it for it. GPT-4 was also the first thing that made me go "woah" when it spotted a frustrating bug I was hunting for the past 2 days in a c library and fixed it. The solution wasn't perfect as it interfered with other functions, but I suspect that's because it can't access the entire codebase at once. This has solutions and is actively being worked on. This is concerning for future job prospects. You might argue this has already occurred given how difficult decently paying jobs are becoming to find, but this will just make it worse. It's the adapt or die sort of mentality. From this though, I predict blue collar jobs to explode in popularity next decade. Unless something else happens that prevents it. I also doubt these massive LLMs are sustainable. They're bleeding openAI dry with the compute costs. Of the $10 billion microsoft gave them in Azure credits they've used nearly 20% of them. Like holy shit, that's so much. OpenAI has stated that the era of escalating model sizes is over, and I'd be inclined to agree. Optimizations to current networks needs to be done to improve training efficiency. I also identify 2 primary issues which I can see only 1 of having existing solutions. 1. These models cannot learn. 2. These models cannot know the conclusion before they get to it. The second part I believe could be solved with an additional middleware model that intercepts and recursively self-prompts the model itself and uses GAN networks to select responses that align with the original prompt and overall goal of what the chatbo/model should do. It should be capable of "knowing" a conclusion and working backwards thanks to essentially being 2 models. It does increase latency, but I wouldn't think by too much. I also think this is already partially implemented in bing chat. The first problem though, I think current LLM models cannot fix and I personally have no proposed solutions to this as there have been minimal advancements in this area. I suspect someone has already come up with a solution to this, just in a completely different field for a completely different purpose. Crazy time to be alive.

The cat's out of the bag. There's no stopping AI at this point without literally destroying all computers on earth. I'm hoping for a benevolent AI overlord. It should be much better than our current leaders. They're pretty pathetic.

Fantastic video. Keep it up.

Beautiful work!

Whatever the future holds, I'm excited for it. Of course I'd love an AI that didn't want to annihilate all human life (as campy as that sounds) but if we did end up inadvertently creating our own genocidal murderer, I'd still be excited. We'll have done something phenomenal. And even if we don't live to see it, *it* will, in whatever way it may. That's exhilarating to me.

That weird pulsing star thing... I'm not sure what it is, but somehow I like AI more now. In fact ChatGPT is the kindest, bravest, warmest, most wonderful human being I've ever known in my life

Dammn dude, you got something special going here, keep it up

Thank you for the amazing video , that's a life changing for me

ChatGPT has admitted potential dangers of AI growing too fast, getting out of control, taking over everything, and causing unbridled harm to humanity.

Amazing video. Keep it up!

super amazing video thank you 😊

Excellent quality video

It's the craziest fresh experience and a truly technocratic opinion. I fully support your words and think that we should embrace this tool how ours greatest invention, that will lead through best times. Gained a sub)

amazing video. thanks for sharing😊

brillant ...keep going

Very concise introduction to Evolution with the Biomorphs. The original Biomorphs program was written by Richard Dawkins and included as an appendix to his book "The Blind Watchmaker" in 1986.

Outstanding video, thanks for sharing. New sub. A+

Great talk, great way to express ideas. I also liked your biomorphs and played with them, it was fun, but what I'd like in the next version is a way to optionally prompt the change in a more direct way than only a choice of 1 out of 4 options. For example as text prompt. Or some 3D-image-editor UI to let you add graphical elements to one of the 4 choices before picking it. As an example I was evolving something that consisted of mostly spheres and I wanted it to start making a more of an expanding spiral shape but without letting any sharp edges in. I could not get there from the 4 choices. I wish there were more degrees of freedom. On the second thought fuck it. It's important to not overthink. May be instead of adding more ways for the human to exert control, give the AI more options to play with - more building blocks, more interesting ways of combining them. Or may be do both give both human and AI more ways to express themselves. Or may be leave it as is - a perfect little toy, what do I know :)

Very good work on summarizing where we are now with the technology. What I wonder is, how much the hardware part will be the limitation at some point. Improvements in data compression allowed files to shrink in size, all while needing less computing power so, with the knowledge I have, even if I think there should be a physical limit to what GPT can do, I have no idea what it is. If you take the human part of the hardware problem: during the pandemic, we were suddenly forced to rely even more on computers and phones to keep the world going on while socially distancing. This led to shortages in advanced microprocessors. I think there would be a limit there, when even if the language model can self-improve, it would need more computing power to achieve it. It could at this point engineer novel ways to build processors and computers, but it will never happen if we decide at this point it has to stop. GPT doesn't have access to the internet, but even if it has, how could it convince TSMC to build processors on his new super advanced ideas, order many factories parts to build robots like the Terminator, etc.? I think the red flag will be pretty obvious if this happens. The real scary part is, how some ill intended people may use the tech.

This gives me some inspiration. I think GPT-J is advanced enough to start this cycle off, if it can see its own source code and is able to self-examine and self-simulate.

This channel needs to be more popular

well done! really good video.

This is a valuable point of view.

Amazing video, thanks!

Love ur videos

Everytime I build a self-improving AI I add a team to provide it feedback before it changes, I also add changelog notes for itself for its next iteration, problems happen in updating its files if it accidently puts placeholders instead of the full logic, which even with filters has been an issue.

Amazing!

very very good work! this video was very vell done!

Great video, subscriber earned 👍

In 1960, Marvin Minsky covered Bayesian Nets and Samuel's checkers playing program, in "Steps Toward Artificial Intelligence". These didn't lead to neural nets for decades due to CPU and memory limitations. Although ironically we're not at the limits for ChatGPT yet, in the not too distant future amplification will necessairly slow down as new thresholds are reached. Even before 1960, Minsky was pointing out that programmers did not fully understand the runtime execution of their programs, so let's not delude ourselves that we can understand and control our creations. Let's not have the hubris to think that we can effectively place artificial limits on ourselves which would just leave the field to the bad actors.

A very well made and thought out video on AI and its possible future.

Great video

I'm a beginner programmer but the idea of self-replicating code and the analogy of how our DNA is the same makes me wonder if AI could create a worse version of itself by accident or by some mistake "hallucination", just as some humans are born with certain syndromes and disseases. In nature we evolve is through mutation and the weaker mutations lose to the stronger ones and that is how we maximise our chance at success if you can not survive you can not replicate your weaker genes, how would AI be able to replicate this test of life and not end up just mutating itself into weaker or dumber versions, if its through some kind of test created by either itself as the "parent AI" or by humans, what's to stop it from just including the test answers into the next generation source code. How could we or the AI determine wether or not it is improving

Observation; ever since the emergence of machine tooling, and perhaps even before, it's been possible through the science of precision to use a lathe to build a better lathe. This strikes me as much the same.

I think you just taught me the steps to making my own advanced ai

Quite measured, good video.

It continues to blow my mind, I even use leisure time to experiment with chatgpt and others I have to compile out of github. I'm having weird psychedelic dreams allong with it. I strongly feel like this will soon go so fast reality itself will get reshaped so fast It would become fluid like , like a dream. That has to be the ultimate result when it starts to interact with the physical.

You sir understand where we are. I salute you

My dad told me about this all the way back when the internet first came around. He was a software programmer and wrote a paper in the early 2000s about quantum recursion in computing and he was one of the folks working on the first chess programs that taught themselves to improve. So I've known this AI takeover was in the future but... To be on the event horizon is honestly nothing less than terrifying.

Very thoughtful......thank you.

Best content ever 🎉

The logarithmic exponential growth of AI doing the impossible 1969: lists primes 1999: beats chess (30 years) 2014: Beats Go (15 years) 2022: drives cars (8 years) 2023: writes stories (1 year later) writer's guild goes on strike This exponential curve is going at a rate of humans improving AI Once AI starts to improve AI, it will break this model, and we don't know how much sooner it will be For if this pattern were to continue, by 2030 it would be able to do every impossible thing that we said it couldn't

Amazing video

There may also be complex logical interdependencies of trade offs that we dont understand yet, that may prevent a program that improves its self at self improving from being good at anything but self improving.