Artificial Super Intelligence Might be Here Already....

So the idea of artificial superintelligence might seem like a far fetched dream at the moment, but currently there are a few things that show us that we might not be as far away as we think. Recently we had this research from MIT come out and say artificial intelligence, scientific discovery and product innovation. And essentially what this paper discusses is how crazy artificial intelligence is at when it comes to accelerating scientific discovery overall. And I'm pretty sure all of you guys recently remember where Sam Altman himself was talking about how when we do have incredible levels of scientific discovery, this is going to lead to a complete transformation in society.

If we can make an AI system that is like materially better at all of OpenAI than doing at doing AI research. That does feel to me like some sort of important discontinuity. It's probably still wrong to think about it that way. It probably still is the smooth exponential curve, but that feels like a good milestone.

And the reason that we know that this is really important for OpenAI is because the next level for OpenAI after agents, which is quite likely to be next year, is of course AI inventors and innovators, which is what they dub sci fi stuff. And this paper coming out of MIT basically shows us the glimpse into the future with as to how advanced technology is going to really transform our society and, and of course speed up one of the most important bottlenecks with as to technological progress. And that's of course scientific discovery, the Level 4 Innovators is going to be completely crazy once we do manage to get there, because that is going to change absolutely everything. Now in this paper they essentially talk about the entire process of doing inventions, which is where basically you can see that you have the idea generation. This is where scientists come up with new ideas for materials that could potentially meet specific needs or have desirable properties, for example strength, flexibility, conductivity. Then of course you've got the candidate materials, and this is of course scientists create possible material designs. Then you've got prioritization. And since it's pretty expensive to decide which designs are the most promising and should be tested first, this step is pretty tricky because they have to figure out which materials will actually work, which actually leads to many false positives. And then it moves to testing, then it moves to testing viable materials. And then you can see after filing patents you get improved products. And then of course you then get commercialization. Now this entire process is actually quite long, it's actually quite extensive. You know that from invention, from literally idea to final market release, it's probably something on the line to like 10 to 20 years. But when we look at how AI is changing the game, this is going to be something that is remarkable for us because we're now going to have the ability to create products and inventions in a much shorter loop. So from idea to market release, that time frame is going to be condensed quite a lot, which is why many AI researchers and those at Frontier Labs are talking about the fact that there's this idea that we are probably going to have a compressed 21st century. Remember guys, from the Dario Amade, the CEO of Anthropic, at his article, what he called the compressed 21st century. And basically this is the idea that after powerful AI is developed, we will, in a few years make all the progress in biology and medicine that we would have made in the entire whole 21st century, which is absolutely incredible. Think about that, guys. So when you think about that, that's basically stating that, look, there's going to be 75 years of progress in five to 10 years, which is absolutely insane. So it's quite like we're going to be living, let's say, maybe two or three lifetimes, all in a short span. And this is basically because if you can master automated AI research and discovery, you can have remarkable breakthroughs across science that transition into a remarkable different way of living for the average person. Essentially, this just talks about things like the elimination of most cancer, prevention of basically nearly all infectious disease, and the crazy things like doubling of the human lifespan, and the idea that the compressed 21st century could once again double it to age 150. I don't know about you guys, but I find the idea of living to the age 150 to be rather incredible. Granted that you're still in, of course, peak physical condition, because there's no point being alive if your life isn't great. So this is where they also show us the rise of deep learning in material science. And this shows a significant advance in the use of deep learning in material science over the past decade. And in the context of this paper, this graph helps us illustrate how the field of material Science has embraced AI technologies like deep learning since 2015. The graph basically has two components here, the blue circles, which is the material science publications. And this line represents the cumulative percentage increase in materials since publications that mentioned deep learning. And since 2015, there's been a really rapid increase which reflects the growing influence of AI in research and innovation. And this surge corresponds with technological advances in AI as well. And we can imagine how crazy that is going to get once we Start to get things like AGI, that's able to accelerate this development even further. Overall, this should show you guys that things the rate of material science publications and other discoveries are increasing year over year. And this is where they basically talk about graph neural networks. And these networks are designed to understand materials at a very detailed level, kind of like creating a super detailed 3D map of atoms and bonds in a substance. And it's basically trained to predict how new structures based on certain characteristics that scientists are looking for. So one of the first steps is inverse materials design. This is where the goal here is to find new materials that meet specific requirements. Scientists input target features like strength or flexibility into the AI and then the gnn, which is the graph neural network. The layers process this information and generate a possible material structure. Imagine it like having an advanced recipe where you input your desired dish, and then the AI suggests ingredient combinations that could make it. Then of course, we've got the three steps training process. I'm pretty sure some of you guys are familiar with these steps. So this is where it goes through pre training. This is where the first AI learns from a big collection of known materials. It understands different kinds of structures, similar to a student studying existing examples before trying to create their own. Then what we have is the fine tuning. Then the AI adjusts its learning for specific types of materials by focusing on certain properties, making it better at predicting materials for particular uses. And lastly, we have reinforcement. Finally, the AI is tested by the scientists who synthesize the materials it predicts. And based on how well the predictions turn out, AI keeps learning to improve. Then of course, we have the graph diffusion architecture. This is where the AI uses something called a graph diffusion process to create new materials. It starts with a known compound and adds some randomness, basically like adding a bit of noise to come up with new possible versions of the material. Then the AI works backwards to refine these noisy versions into something useful, trying to remove the randomness and end up with a stable, practical new material, which is somewhat similar to diffusion models in images. So when we actually think about how this all fits into the paper, this is basically just crucial because it significantly improves the materials discoveries process. So instead of scientists manually trying different combinations through trial and error, which, you know, takes a lot of time, the AI tool can quickly generate suggestions that are more likely to succeed. And this process allows scientists to be more efficient, focusing their efforts on evaluating promising AI generated ideas rather than generating them from scratch. Now, the crazy thing about this is that we also have this graph right here and Basically, on the first graph you can see that this one labeled new materials basically shows how the number of new materials discoveries basically changed before and after adopting the AI tool. Before adopting the AI tool, we can see that the rate of material discovery stayed relatively steady. And after AI was introduced, there's a clear upward trend indicating that scientists began discovering more new materials. Basically, the AI tool helped scientists generate more innovative materials, increasing their productivity significantly and automating the tedious part of coming up with ideas, allowing researchers to focus on the evaluation of and testing process. And we can also see the same in the patent filings. We can see an increase right after the AI integration and we can see the same in product prototypes overall. Showing you that right now, of course, this is just using DNNs, but imagine what happens once you get to AGI or even ASI level. AI that's able to increase the rate of discovery. It's going to be pretty crazy. Now, one of the craziest things that I saw about the study that isn't related to ASI but is somehow related to the economy was basically how these individuals had their attitudes changed after interacting with this AI tool, which highlights both the optimism and concerns regarding its impact. It basically tracks agreement levels across several statements measured on a 1 to 10 scale. Initially, scientists agreed that AI will make scientists in my field more productive. And this belief grew even stronger after they used the AI tool. However, there's concerns about AI replacing scientists who also increased after using the tool, which basically suggests that direct experience with AI led to greater awareness of its disrupting potential. And the agreement that AI will change the skills needed to succeed also rose, showing that scientists realized that the need for new abilities whilst working alongside AI is there. And of course some scientists plan to reskill to collaborate effectively with AI. And interestingly, satisfaction with their choice of field decreased after using the AI tool, which reflects the concerns over reduced creativity and increased automation in their roles overall, showing that once individuals start to use these tools, they realize how productive they are, but of course, potentially their diminishing roles for the future. Interestingly, what we actually did see here, and I'm just going to read this verbatim, basically it says, consequently, the number of researchers planning to reskill grows substantially. Finally, scientists report a small reduction in satisfaction with their choice of field, consistent with a decline in job satisfaction in the previous section and it says here, okay, these results show that hands on experience with AI can dramatically influence views on the technology. Furthermore, the responses reveal an important fact. Scientists did not anticipate the effects documented in this paper. And this fits a recurring pattern of domain experts underestimating the capabilities of AI in their respective fields, basically stating that, look, the majority of people that don't interact with AI tools will underestimate their capabilities until they use them. And what's crazy, okay, one of the statements from the actual researchers said that, look, while I was impressed by the performance of the AI tool, I couldn't help but the feeling that much of my education is now worthless. This is not what I was trained to do. And we also get to see comments from Rune. Honestly, I don't know who Run is, but there is some speculation that he's an OpenAI employee. There's some that he's an AI insider, but he basically said that obviously don't believe any economic studies at face value, but this is what it looks like when you've discovered superintelligence. The researchers are outsourcing idea generation tasks and running the experiments themselves, effectively making themselves lab robots. And basically this is just a provocative interpretation of the MIT study we just looked at. And essentially this kind of shows us an early indicator of superintelligence. And whilst yes, there is a crazy study statement, this is essentially what superintelligence will be used for. It's not going to be used to do your taxes or write essays or even create movies. It's going to be used for scientific discovery.