FOR EDUCATIONAL AND KNOWLEDGE SHARING PURPOSES ONLY. NOT-FOR-PROFIT. SEE COPYRIGHT DISCLAIMER.

OpenAI CEO: “Super Human Coders By End of 2025” | Matthew Berman

by the end of this year it’s looking very likely that we’re going to have an AI agent that is the best coder in the world better than any single human coder and we’re going to be able to replicate it a million times over and right around the corner from that we’re going to have ai that can discover new knowledge cure all disease discover new science and fundamentally change the way Society works so I want to show you the proof for all of this and this video is brought to you by Polo AI an incredible AI image and video generator but more on that later Sam Alman just gave a talk in Japan where he highlights the insane progress that we’ve seen in AI especially since test time compute 01 03 deep seek have become mainstream so first Sam Alman is going to talk about the orders of magnitude increase in intelligence that these models have seen in such a short period of time and in parallel we’re seeing exponential decreases in cost deep seek shocked the world but for True AI leaders like Dario and Sam Alman deep seek actually fell right on the predicted curve of what the cost of intelligence will drop to so let’s watch this clip we used to be in a paradigm where we only did pre-training and each GPT number 1 2 3 4 each of those was exactly 100x of or not exactly but very close to 100x and at each of those there was a major new emergent thing the most important thing that happened in the field or at least to us in the last year is these new models that can do reasoning they are an incredible new compute efficiency gain and we can get performance on a lot of benchmarks that in the old world we would have predicted wouldn’t have come until GPT 6 something like that from models that are much smaller by doing this reinforcement learning um so we kind of have a sense now the trick is when we do it this new way it doesn’t get better at everything we can get it better in certain Dimensions so what he’s describing is from gpt1 2 three every single number version has been a 100x increase in intelligence and they see some kind of new emergent behavior from that increase and all of that came from pre-training kind of the traditional collect all the data throw it into a model and see what you get but now we have test time compute which means when you give a model a prompt it think longer about it that’s what the 01 and 03 family of models have done that’s also what deep seek has done and it’s a pretty incredible unlock and what we’re seeing is a faster rate of innovation and a faster rate of efficiency gains than any other technology that we’ve seen in history think about the transistor one of the most important Technologies in the history of humans the transistor is the fundamental unit of compute it is the thing that basically goes in every single device that you own it goes in things that you might not even think of toasters and refrigerators and everything and there’s this thing called Moore’s Law which was a prediction by someone named Gordon Moore who is a co-founder of Intel and it basically states that the number of transistors that can fit on a chip doubles every approximately 18 months and that means that the computing power we have essentially doubles every 18 months and a lot of people thought that Mo’s law was going to translate right into artificial intelligence but it turns out AI is actually far exceeding Moore’s Law and I’m going to talk a little bit more about that later when we go over Sam alman’s blog post that he just dropped now back to test time compute test time compute is interesting and deep seek really showed the world how it works you’re using reinforcement learning with verifiable rewards to train the model after the base model has been trained so what does that actually mean it basically means allowing the model to try a bunch of different solutions and it will get rewarded for the right ones now here’s the thing about rewards you can only reward a model if you know the answer to the problem so that’s why these thinking models are really good at some subjects but not others but the good thing is the subjects that they’re good at are actually really important in the world it’s things like coding and Math and Science and that’s because there are provable solutions two problems that these models can train on so 2 + 2 always equals four so that’s the verifiable reward with math if the model says 2 + 2 equal 4 you can say yes that is correct on the other hand if you say write a poem there is no correct answer to that so that’s why these models are getting really good at STEM Science technology engineering and math thanks to the sponsor of this segment Polo AI Polo AI is an AI video generation tool that allows you to use the most Cutting Edge open- Source models really easily to generate video that you want to see whether you need it for marketing b-roll or just having fun Polo AI is perfect for it it supports text to video image to video video to video and consistent character generation and they support a bunch of Cutting Edge video models including cing Runway Luma Vu huo and pix verse and you can generate video on all of them from a single API or through the UI for a single price too they have over 40 AI video effects and templates which allow you to create video and share it really easily so simply choose your model right here look at all these models they support then type in your prompt select your aspect ratio select visibility copy protection and then create they have a free version where you get obviously to try it out and then they have paid versions in which you get many more credits many more features and you can create private video so check out Polo aai try it for free I’ll drop a a link down below and let me know what you think thanks again to Polo AI for sponsoring this segment now back to the video all right next Sam Alman is going to talk about kind of the very clear path to having AI that can discover new knowledge and using all of the things that we’ve learned about pre-training combined with our ability to scale up test time compute is going to unlock incredible intelligence from these models and there doesn’t seem to be a limit in sight let’s watch but we we can now I think more intelligently than before say that if we were able to pre-train a much bigger model and do this where it would be and the thing that I would expect based off of what we’re seeing with a jump like that is the first bits or sort of signs of Life on genuine new scientific knowledge I don’t know how many times I’ve shown this graph but I keep coming back to it cuz it’s so important this is the graph of the intelligence explosion from the situational awareness paper by Leopold Ashen brener and I’m going to say it again what we see automated AI research as soon as AI can discover new knowledge and then apply those new discoveries to itself we are in a self-improving loop and that is when we’re going to hit the intelligence explosion and it really truly feels like that is right around the corner and seemingly the first superhuman thing that AI is going to be able to do is coding coding by the end of this year is going to be completely different Sam predicts by the end of this year open AI is going to have a model that is the number one coder in the world better than any single human coder now let’s think back to Alpha go it’s using essentially the same technique as alphao did you have this AI model with really no prior knowledge of how the game of Go worked but through iter ative selfplay it got better and better until it figured out techniques that humans didn’t even discover the famous move 37 so imagine we now have this AI that is the best coder in the world we can replicate that hundreds thousands millions of times and these coders are writing code 24 hours a day and the only limitation is how much compute and energy we can throw at it these agent coders are joining the workforce right now through through projects like Devon but it’s not going to stop at coding this is going to happen with all knowledge work as AI gets better than any single human at any single category of knowledge work they’re going to flood the workforce and all of a sudden every single human is going to be insanely more productive because of the help of these AI agents in whatever field they’re in let’s watch this next clip let’s say like 03 our very latest best model that can program unbelievably well and if people have already done it it’s not so good at going to like invent totally new algorithms and that’s or new physics or new biology and that’s the thing I think you’ll get with the next two orders of magnitude so he said it right there although their current models are not going to be able to discover new algorithms and new science the next 100x is going to be able to do that that is not far away especially given we have a clear path on how to scale up to that with test time compute it’s really quite the the progress over the the reason scale is quite amazing our our very first reasoning model um was like a top 1 millionth competitive programmer in the world people thought that was very impressive it’s like wow in AI it’s you know the millionth best people that do this that’s pretty good um we then had a model that got to like a top 10,000 uh 03 which we talked about publicly in December is the 175th best program competitive programmer in the world I think our internal Benchmark is now around 50 and maybe we’ll hit number one by the end of this year that’s like an amazing rate of scale for more compute in this new paradigm and we don’t see any signs of that stopping all right so we said it right there they already internally have a top 50 programmer in the world I I can’t really stress how crazy that is top 50 of the entire 8 billion people on this Earth of course not everybody can code but of all the potential coders it’s top 50 and at the end of the year it should be number one and it makes sense AI is already better than us at chess it’s already better than us at the game go what’s the difference between that and coding there really is no difference besides the set of rules they are fundamentally the same in that it’s a nearly unlimited decision space but there are verifiable ways to say yes this is a win or this is the correct solution and no this is wrong or this is a losing decision and just to continue on the point of how rapidly AI is progressing Sam Alman is about to say by the end of this year we’re going to have ai that can essentially do anything better than humans just falling short of discovering new knowledge and again that’s the end of 2025 this year and maybe just in 2026 one year from now we’re going to have agents that can discover new knowledge let’s watch and overall I hope that by you know the end of this year we have a model that you can use if you have the pro tier you can turn the compute all the way up and you can ask it a really hard question not one that requires discovering new science but most things short of that and you kind of just get it to work it may have to go off and think for a few hours it may have to go use a bunch of tools but it kind of just does it for you so there it is incredible things to get excited about by the end of this year now Sam Altman just dropped a blog post in which he kind of talks about some of the the same stuff but really make some bold predictions about what agents and AI in general are going to be capable of very soon so let’s take a look all right so here’s the blog post it’s called three observations and listen to this systems that start to point to AGI and he adds a little star there so he’s going to Define AGI are coming into view and we think it’s important to understand the moment we are in so let’s look what he has to say about AGI down below by using the term AGI here we aim to communicate at clearly and we do not intend to alter or interpret the definitions and processes that Define our relationship with Microsoft so good because remember the original agreement with Microsoft was defined by the definition of AGI as soon as they hit AGI Microsoft could no longer get their research we fully expect to be partnered with Microsoft for the long term this footnote seems silly but on the other hand we know some journalists will try to get clicks by writing something silly so here we are preempting the silliness okay so basically he’s saying yes we’re going to say this is Agi but no it’s not going to change our relationship with Microsoft so he does talk about AGI and talks about the two ways to look at it one it’s just another tool in our tool set we humans have developed tools for as long as humans have been around from the wheel to fire to the steam engine the transistor and so on and he’s saying okay maybe it’s just another tool but on the other hand maybe it’s not in some sense AGI is just another tool in this ever taller scaffolding of human progress we are building together in another sense it is the beginning of something for which it’s hard not to say This Time It’s Different and I kind of agree this might be just another tool but more likely it’s the last tool we might not need anything after this AGI will be the definitive discovery of humans whatever you believe is the outcome from that whether positive or negative the fact that this is probably the final Innovation that we as humans need to make is probably true the economic growth in front of us looks astonishing and we can now imagine a world where we cure all diseases have much more time to enjoy with our families and can fully realize our creative potential so I want to take a second and pause here because it’s kind of funny Sam Alman oscillates between being the hypest hyp pr ever about Ai and then also saying hey the hype’s a little bit out of control look at this tweet from January 20th Twitter hype is out of control again we are not going to deploy AGI next month nor have we built it please chill and cut your expectations by 100x and then just a few weeks later he’s talking about AGI coming right around the corner again I mean it’s pretty funny so now back to his blog post in a decade perhaps everyone on Earth will be capable of accomplishing more than the most impactful person can today now he touches on three what he calls observations here they are the intelligence of an AI model roughly equals the log of the resources used to train it and run it now training we kind of knew pre-training run it is the inference test time compute the thinking models that’s what he’s talking about here so what does he mean by that these resources are chiefly training compute data and inference compute it appears that you can spend arbitrary amounts of money and get continuous and predictable gains the scaling laws that predict this are accurate over many orders of magnitude now remember just a minute ago I showed you that he said within two orders of magnitude we’re going to have ai that can discover new knowledge at that point we are at the intelligence explosion and if the combination of pre-training and test time compute can get us to multiple orders of magnitude that’s the clear path that he has been talking about to AGI and B Beyond next and this was illustrated by Deep seek the cost to use a given level of AI falls about 10x every 12 months and lower prices lead to much more use so remember Mo’s law the number of transistors on a chip doubles every 18 months maybe alman’s law the cost of using AI at any given intelligence level Falls by 10x every 12 months that far exceeds Mo’s law here you can see this in the token cost from GPT 4 in 2023 to GPT 40 in mid 2024 where the price per token dropped by about 150x Mo’s law changed the world at 2x every 18 months this is unbelievably stronger and third the third observation the socioeconomic value of linearly increasing intelligence is super exponential in nature now that was a little bit of word salad but let me try to explain what that means now even as we get linearly increasing intelligence even though it’s much faster than that linear just means a straight line the actual value that comes from that intelligence is super exponential now he goes on to talk a little bit about software engineering agents in this post and he already covered it in that talk in Japan but let’s go over what he says in this post imagine a case where a software engineering agent and they think it’s going to be particularly important imagine that this agent will eventually be capable of doing most things a software engineer at a top company with a few years of experience could do for tasks up to a couple days long and he goes on to say but imagine you have a thousand of them or a million of them and it doesn’t stop there imagine that but now imagine it in every single field of knowledge again the intelligence explosion so AI may turn out to be like the transistor economically a big scientific discovery that scales well and that seeps into almost every corner of the economy and by the way I highly recommend the book chip War if you want to learn more about the birth of the transistor and really how it changed the world both from obviously a technology perspective but also a geopolitical perspective just a great book so we don’t think much about transistors or transistor companies and the gains are widely distributed but we expect our computers TVs Cars toys and more to perform Miracles and it’s just expected at this point the future will be coming at us in a way that is impossible to ignore and the long-term changes to our societ and economy will be huge and it is coming people’s jobs people’s lives are going to change dramatically there’s going to be a lot of friction in the coming years definitely but I think just like previous Technologies we’re going to adjust as humans to this new normal so an example is for a long time something like 90 95% of all humans worked on farms they needed to grow their food but at a certain point as Farms became more mechanized and automated I think only now single digigit percentage of all humans work on farms and what happened to all those other people are they just forever out of work well no they figured out other things to do for example my job talking into a camera if I needed to work on the farm all day I would not be able to do this and there’s a million other jobs just like it we’re going to discover new things to do with our time new ways to be creative and of course I tend to be very optimistic about the future but I do truly believe this and the thing that I’m most excited about is scientific progress the notion that cancer might be cured all diseases might be cured is something that is so exciting to me and he talks about that he says although some Industries will change very little scientific progress will likely be much faster than it is today the price of many Goods will eventually fall dramatically but he does say on the flip side which I hadn’t thought about the price of luxury goods and a few inherently limited resources like land may rise even more dram atically now there are some problems we need to solve and I’ve covered this in a previous video the balance of power between labor and capital is going to matter much more than it does today and he says it it could easily get messed up and this may require early intervention and so I covered this previously and what I said is that right now Capital has to pay labor to do work and that labor can choose to do it or not do it thus giving them a good amount of power however if Ai and robots are able to do everything humans can do but better and they don’t turn you down they don’t say I need a break they don’t say well this other company gave me a better offer I’m going to go work for them rather Whoever has the capital is going to be able to purchase labor infinitely and yeah that is going to change the balance of power between capital and labor so there’s a lot of things to think about a lot of exciting things obviously some unsolved problems in the very near future coming but I’m personally very excited about all of it and thanks again once more to Polo AI for sponsoring this video If you enjoyed this video please consider giving a like And subscribe and I’ll see you in the next one

FOR EDUCATIONAL AND KNOWLEDGE SHARING PURPOSES ONLY. NOT-FOR-PROFIT. SEE COPYRIGHT DISCLAIMER.