State of AI 2023: Highlights of 163 Page Report + Eureka Self-Improvement, MEG, Suno AI and GPT F

The annual State of AI Report is one of the most cited documents in the field of artificial intelligence, endorsed by the likes of Andrej Karpathy, led by Nathan Benech this year. The original co author, Ian Hogarth, is now one of the key players at the Bletchley Global AI Summit in November. Yes, I read all 163 pages of this year's edition and thought that giving you just the highlights, interspersed with developments from just the last few days like Eureka would be a good way of keeping you guys up to speed. But let's start out with a modality that the report didn't cover. Notice on page seven we have text, image, video, music, robot states, and we'll cover all of those. But what we don't have is thoughts. And why am I bringing that up, you might ask? Well, because a few days ago we had this from Meta Towards a real time decoding of images from brain activity. Now I have showcased thought to image and thought to text before on the channel, but never real time thought to image. It's not perfect, but progress is astonishing and I think it points to the explosion of modalities that we are seeing at the moment in AI. Now, if you're willing to sacrifice a real time flow of images from the brain, this is the state of the art using fmri, much more accurate but not real time. Now many of you might wonder if it works if someone imagines an image, and apparently it doesn't work nearly as well. And also at the moment it tends not to work if the participants are distracted by other things, as the paper says. In other words, the subject's consent is not only a legal but also a technical requirement for brain decoding. Back to the report and many pages are dedicated to covering GPT4. But to be honest, anyone who's followed this channel for any length of time will know a fair bit more than what is covered in these few slides. So I'm going to Skip to page 14 where I have a slight difference with one of the conclusions. They cite a paper which describes the false promise of imitating proprietary LLMs large language models, basically saying that you can imitate the style but not the content of a smarter model. But I wish that they had cited Orca, which despite being more than 10 times smaller than the original ChatGPT reaches, according to the technical paper, parity with ChatGPT on various benchmarks, showing only a four point gap in professional and academic examinations like the SAT, LSAT, GRE and GMAT. That's maybe why the information reported on Microsoft trying to substitute for GPT4 with or claiming to mimic the quality of OpenAI software at a fraction of the cost. On page 38 of the report they describe open ended learning with large language models describing how they can explore and gain skills in a game like Minecraft. And they say that the best example of this kind of self improvement and iterative prompting is Voyager, which is again GPT4 getting better and better at Minecraft. But just after the report came out we got Eureka and yes it deserves its own video, but as Jim Fan wanted of the key authors said, it's like Voyager in the space of physics simulations and essentially what they did was feed the source code of the environment to GPT4 and then they asked it to write the code for the reward function. And then using the sheer power of parallelization inside Nvidia's Isaac Gym, they could then test out those reward functions, see which ones in simulation work the best. Thanks to GPU acceleration, you could speed up reality a thousand X each time. The best grading systems would be used, refined upon, and then test it again, all in simulation. Ultimately, Eureka Rewards outperformed expert human written reward functions on 83% of the tasks performed. The average margin was 52% and Eureka was able to learn even pen spinning tricks, which are notoriously difficult even for CGI artists. This marrying up of large language models and iterative feedback from the environment is probably the future of AI. As Sheetal Shah from Microsoft said, the proverbial positive feedback loop of self improvement might be just around the corner that allows us to go beyond human training data and capabilities. And to those wondering yes, they are planning to connect it to a real robotic hand soon. I was lucky enough to get early access to the report and one of the most convincing and amusing demonstrations was Eureka learning how to run efficiently. Back to the report though. And if Eureka was all about simulated embodiment getting better with language feedback, how about language tasks being improved with robotic data? As this page demonstrates, the positive transfer seems to go both ways. Gaining vision data and embodiment or robotics data seems to make a model like Palm E better at pure language tasks. And while the report went on to mention RT2, watchers of my channel will know about RT2X that showed how data from a multiplicity of different robotic setups could complement each other and improve the base model. The basic message across the board is the same more data from more modalities improving performance performance across the board. And so when we get Eureka improving via feedback from GPT4, it reminds me of a previous video where I talked about Dall E3 improving from feedback from GPT4 Vision. The same themes are reoccurring improvements in one modality, ricocheting across two others. And remember, the original Palm language model and GPT4 weren't even designed for this. I am genuinely curious about the kind of recursive improvement that might occur with a model designed to be multimodal from the start, like Gemini. Speaking of robotics, though, the report did mention something I missed on the channel that's the first time win for a robot in a competitive sport which was first person view drone racing. Here is a glimpse of why the Swift system beat all the human champions.

The trajectories flown by the humans and Swift. We notice that the autonomous system is more consistent across laps and is able to take tighter turns. This gives it a decisive advantage in a race. Swift won multiple races against each of the human champions and achieved the fastest recorded race time. This marks the first time that an autonomous mobile robot has achieved world champion level performance in a real world competitive sport.

What might come next? Well, these have received approval in China for fully autonomous flights of an electric air taxi. Let me know in the comments if you would fly in one of these Ehang drones. At the moment I'm undecided. Probably not right now. I'd have to see in action for a bit longer. On page 58 the authors describe another year of progress in music generation. But rather than just describe this progress, I want you to hear it. I'm not sponsored by them, but I've been honestly very impressed by Suno AI. Check out a snippet of the following rap all generated by AI AI explained.

Breaking through the mold Diving deep Discovering secrets untold Smarter than human It's a.

Whole new game Unleashing power A technological flame.

If you like that one, I'm going to end the video with another sample very quickly though. The way to test it out is to join their discord. Click on one of the chirp beaters and then just type forward slash chirp and then press enter. Then you have to press enter again for some reason and up this will come. At the top you can put things like rap or classical, whatever you want and then you can either enter four to eight lines of your own lyrics or have ChatGPT do it for you. Honestly, don't make more than around eight lines because it degrades performance quite a lot. Page 63 briefly describes the progress in medical performance of language models and I do honestly think that was worth highlighting. I just wonder if MedPalm 2 when it got 85% used a smart GPT like framework because if it didn't, things like self consistency, chain of thought, prompting and reflection could boost the score even higher than 85%. In other words, it is not out of the question that we could soon have in our pockets a model capable of outperforming all but the very best doctors at medical question answering Speaking of medicine, apparently it's the fastest growing field in terms of mentions of AI within papers. And just look at the breadth of fields now applying AI to achieve research breakthroughs and the increase in volume. And notice at the bottom how much mathematics is benefiting from this progress. And I bet many of you don't know about the Automated Prover and proof assistant GPTF. One of the co authors of that in 2020 was Ilya Sutskova of OpenAI. It was about proving mathematical theorems using a GPT model and again, something many might not know is that that was superseded in 2020. Their Neural Theorem Prover solved 10 international math Olympiad problems. That's five times more than any previous system. Their method was called Hypertree Proof Search and is way too complicated for me to cover in this video. And that's not even touching Alpha tensor from Google DeepMind, which improved upon decades old methods for matrix multiplication that has direct ramifications for the usage of GPUs in modern day AI. Very quickly what they did to automate algorithmic discovery was convert the finding of efficient algorithms for matrix multiplication into a single player game. And I love this bit. They say this game is incredibly challenging. The number of possible algorithms to consider is much greater than the number of atoms in the universe. And compared to the game of Go, which remained a challenge for AI for decades, the number of possible moves at each step of our game is 30 orders of magnitude larger. And it's these synergies that I'm interested in. The breakthroughs in one field that will then be applied to another. The state of the art results in one modality that push forward another another modality. And the number of people trying to do this is multiplying as well. Honestly, I hadn't even heard of Imbue. Apparently they have 10,000 H1 hundreds which are the most cutting edge GPUs. I don't know how I missed an AI company with five figure numbers of H100 GPUs, but here's the CEO describing.

Their approach at Imbue, we're training large foundation models optimized for reasoning. On top of those models we build agents that we use to accelerate our own research. And these do more than just output something. They also iterate and reflect and figure out what's the next step to do and then take that next step. We're starting with agents that code because it requires complex reasoning to be able to code well, and because that's the work we do every day.

They seem to be flying completely under the radar with only about 300 subscribers on their channel. So to be honest, if we see the kind of manual dexterity that Eureka promises and a breakthrough in reasoning like Imbue and frankly, everyone else is working on and if AI can start to crack international Math Olympiad problems and compose Ratback battles, it does start to become a genuine question of what can't they do. I'm not saying that's true now, far from it. But in the not too distant future that may well be a very valid question. In the Industry section on page 82, the authors describe the chip export bans that apply to H1 hundreds. Basically, companies like Nvidia and Intel can't sell their cutting edge chips to China and select other countries, and those companies got around it by creating models like the A100 which were just outside of the export restriction range. Well, the report is already somewhat out of date in that respect. Just four days ago we hear that those chips are now also under export restrictions. Another development that's occurred since the publication of the report concerns copyright. The report detailed how Microsoft has moved to reassure the users of its tools that the corporation will assume any legal risks in the event of any copyright claims. And of course every day we seem to be reading reports of one company or another getting sued as it relates to AI data. But just a few days ago we had this Google giving the same kind of indemnification. Basically, if the use of one of its gen AI products triggers any copyright claims, it Google will assume legal responsibility, it says. My only prediction is that this battle over data and copyright is not going anywhere. In somewhat grimmer news, we had a series of slides on the use of AI in warfare and how in real time conflicts are being used as a lab to design the latest and most effective weapons. What came out since the report though is news about the potential first use of an autonomous killer drone. That's a drone that can identify and attack targets without any human control, with a human not even in the loop. There were rumors of this occurring in Libya a few years ago, but this seems to be more confirmed. Apparently it's only used autonomously when radio interference or jamming prevents direct operator control. The article in Forbes says going forward, if the system is deemed to be sufficiently reliable, large numbers of autonomous attack drones could be deployed simultaneously. Simultaneously without the need for trained operators. Dozens of bomber drones could attack simultaneously, immune to jamming and anti drone guns. And while many campaigners have sought to ban this kind of autonomous killer robot, there is still no international agreement on doing so. The pace of technology far outstrips the pace of diplomacy. We saw earlier in the video that AI is better at maneuvering drones, but apparently is also better at spotting targets. And other companies say that they are ready with an autonomous version of their Switchblade Kamikaze drone if there's a demand for it. Were quite a few pages on safety, but I worry that the debate is slipping into two camps. One that thinks we're all doomed and the other that thinks safety is a complete myth. Sadly, I could see the debate staying that way until there's an actual incident, either with autonomous drones or a bioweapon. On a lighter note, the report described the progress in text to video, but again, rather than reading it out, I thought I'd show you. And in this video, notice that you can pause at any moment and not discern anything other than a beautiful landscape. Now to end with, here is the predictions page of the report that it includes every year. And honestly I feel that the predictions are a little bit too conservative. Their first prediction is that a Hollywood grade production makes use of generative AI for visual effects. But we literally had reports of that occurring three days before the report was even released. It looks like the background to this low key promotional image was AI generated, created the clock face. And so far I don't think Disney have denied that. And what about the third one? Self improving AI agents crush the state of the art in a complex environment. They talk about games, tool use and science. But to be honest, Eureka is kind of already a demonstration of that. You could call physics simulation science or potentially tool use. So I would say this prediction will definitely happen, if not has already happened. And the fifth one about a group spending a billion to train a single large scale model. Well, multiple CEOs have already said that that's their plan and I seems like I'm being harsh. But the seventh one, we see limited progress on global AI governance beyond high level voluntary commitments. That's a very subjective prediction, depending on the definition of the word limited. I for one do think and predict that there'll be some sort of audit agency set up before the end of 2024. Now if I'm criticizing them, it would be remiss of me not to end the video with my own prediction. I predict that a model will be released in the next year that will break state of the art benchmarks across at least four different modalities simultaneously. Anyway, you can let me know what you think of that prediction and the video as a whole in the comments. Thank you so much for watching, as always. And as a tiny reward of watching all the way to the end, here is another rap about this channel.

Listen up y', all, I got something to say about a YouTube channel. AI explained today smarter than humans, that's what they claim Diving deep in tech, they ain't playing no games AI explained Breaking boundaries with Mike Unravel secret shining knowledge so bright they break it down Making complex things clear Teaching us all, evolving year after year.