The Singularity is HERE? LLMS Are Now "Self Evolving"

So the LLM space doesn't seem to be dying down at all because recently there was a new type of LLM called a self evolving Large language model. And this is absolutely revolutionary because one of the key problems with current LLMs is that they can't update themselves. It's like a time capsule for knowledge embedded into a large language model. So apparently there is something called Writer, a 2 bit billion dollar valuation startup that's developing AI tools for enterprises that have now developed this new type of self evolving LLM that they claim can continue learning and updating its parameters. The implications for this are quite staggering because it means that they can continue learning and it allows them to change how it answers questions even after it is deployed. So basically one of the main problems with LMS is that they are really, really expensive to make. And the problem is, is that if we have to keep retraining models, it's going to continue to get more expensive. You take a look at this paper that dives into how expensive these things are going to cost. You can see it says that the largest training runs will cost more than a billion dollars by 2027, which means that only the most well funded organizations will be able to finance frontier AI models, which kind of leads to a poor monopoly on the situation. So the craziest thing about all of this is that most companies don't have a billion dollars to run these AI training runs. And of course, if we could solve this issue, it would be an enormous breakthrough for the AI community because it would reduce the amount of spending needed to fund these AI projects and thus speed up AI development. And not only is it the cost of these large language models that is the issue, we also do have the problem that current LLMs, when we are trying to ask them things that have happened recently, they always say without using their current search capabilities, their knowledge is accurate up to potentially a year ago. So right now, GPT4O, the version that I'm using, the knowledge cutoff, is actually up until 2023 and that is without search. So if I wanted to ask it anything really specific, I would have to get it to browse the web. And sometimes web doesn't have certain pieces of information. Now I hope you can all understand why having a model that is up to date is really important because we live in a really fast moving world. So if this model is still thinking in 2023 terms, it's going to have a severe disadvantage over a model that is able to dynamically update itself based on what recently happened. Now, they basically State that this actually works in a new and interesting way. It says traditional transformers, which are the model architecture that underlies popular LLMs like ChatGPT and Claude, are made up of layers or special filters that help the model learn different aspects of the input data. And within each of those layers, Rytr's self evolving LLMs also include a memory pool which stores important information from past interactions, said Rytr and co founder the cto. So basically you can see here they have this memory pool which is able to update itself from past interactions and is able to get better at certain responses by storing that somewhere inside of it. So each time the LLM receives new information it hasn't seen before, it's able to update the memory pool with that new information. And it does this throughout all its layers so that it can refer to the information in future interactions. And crazily, they also state that they do have some control over what the LLM does and doesn't learn, so you can't just troll the LLM with fake facts. And it's actually pretty interesting because I do remember when I was talking to ChatGPT, I oftentimes I would try to see what the model is able to believe from me. Like I would try to tell it, recently I was appointed the King of England. And sometimes it would try to say, nope, that's not possible. And it was pretty funny to see the kinds of responses that I was able to get from the models. But of course I know that this kind of thing, wide scale, is going to be really fascinating because I'm wondering how the LLMs are going to discern from what is true from what is false. And of course when things are on the Internet, people can tend to mess with them. So now when we actually refer to the training course, they say that developing a self evolving LLM and increases Training costs by 10 to 20% but doesn't require additional work once the LLM is trained. Unlike other methods to update models with new information like retrieval, augmented generation or fine tuning. So they're basically stating that this doesn't require any additional work, it just requires work on the front end, which basically means that once you have this model, it's going to be like the kind of model where you won't really need anything else, provided there aren't any major architecture breakthroughs like the Mamba architecture or those ones that other people are using. So it's going to be really interesting to see if this is widely adopted, but there is also some fascinating effects from this. So one of the fascinating effects as well is the fact that it leads to some interesting results on popular benchmarks. It says specifically, the model's performance on benchmarks actually improves every time it's tested, since it learns the information on the benchmark over time. For instance, the model was tested on a common math benchmark and it got less than 25 of the questions correct. And the third time it was tested, its accuracy jumped to nearly 75%. Now, I think this is a little bit controversial because some people would argue that if the model is able to remember what it was seen in the test previously, then the model isn't actually getting smarter, it's just able to understand what has gone on before. But when we look at the actual benchmarks, we can see that, yes, the score does improve. But I do wonder if that is just based on the memory of the model remembering the exact kind of questions, or it's able to remember the kind of reasoning that it has before. And because it isn't just something that you're asking questions to, it's able to dynamically understand the kinds of questions that it failed in the past. And of course it's able to get the future ones right. I think if that is the case, like the second scenario, where it's not just remembering previous questions and it's remembering the actual way to be smarter, then that would mark a huge, huge breakthrough because this would essentially mean that this is the kind of model that can literally get smarter. And, and if this is baked into something that's like a one level smart, then that would mean that you could technically get to AGI or ASI with a singular kind of system, which has truly stark implications if it's actually doing the reasoning part and not just remembering the answer to those questions. And I'm sure you guys can understand why this is one of the most important things to talk about, because self evolving LLMs are something that we realize is literally going to be the cause of an AI catastrophe, if that ever is one. Now, they also currently talk about how they do have this in beta with just two customers right now. And of course they're still ironing out all of the kinks. So that is going to be something that they will need to do. But one of the major problems that they are currently having with this model is that the information the LLM learns, the worse it gets at refusing to answer new and dangerous questions. So you can think of this as the new information that the model learns over time will begin to override the original data, such as safety instructions it was trained on and that's not great news for businesses that want to incorporate that kind of AI into customer facing products. And one of the main problems that we do have with today's AI is the fact that we always pay attention to the safety aspect because we don't want these models outputting harmful things. And if the model is able to learn new things or new ways of surpassing those kind of guidelines, then it's going to be really problematic for that model to be out in the public space. So it's quite likely that this is the kind of model that, that will only be used in the business sense because it's too obvious that a lot of people will just try and ruin the model by updating it with new and terrible information. Now they actually talk about this as well. They state that they have to limit how much new information that they can learn. And they argue that this isn't a big of a concern for businesses as they're just typically just trying to update an LLM with their own private information rather than all of the latest data on the web. And he added that if you make the memory pool around 100 billion to 200 billion words, that it's enough for the LLM to learn for at least five to six years for the typical enterprise. So that's pretty crazy when we think about it. This model is able to learn for about 5 to 6 years for the typical business. So I mean, I do wonder if this is going to be something that more people adopt as they're able to have these models that are continuously getting better rather than ones that they have to continually find new ways to prompt Engineer and of course do all these different hacks to be able to get the model to do exactly what they want want. Now of course this is something that has gone on in the industry because Mustafi, Suleiman actually recently spoke about how memory was basically going to be solved. And this is something that we do pay attention to because of course if you've seen on this channel actually spoke about Microsoft's recent breakthrough and how they spoke about this giant breakthrough coming through.

I mean, I'm really confident 2025 memory is done. Permanent memory. I mean if you think about it, we already have memory on the web. We retrieve from the web, you know, all the time, quite accurately now. Copilot has really good citations. It's up to date 15 minutes ago, knows what's happened in the news on the web and so on. So we're, we're just kind of compressing that to do it for your personal knowledge graph and then you can sort of add in your, your own documents and your email and calendar, stuff like that. So memory is going to completely transform these experiences because you will be. It's sort of frustrating to like have a meaningful conversation or go on a interesting exploration around some creative idea and then come back three or four or five sessions later and it's like let's start again. Completely forgotten what we talked about, you know. So I think that's going to be a big shift as well because you'll know not only does it lower the barrier to entry to you expressing a creative idea, but those things don't get forgotten too. So you can do this ambiguous cross reference back to something that you wanted. What was that thing I said like three weeks ago? And that is it's sort of like having a second brain in us. It's like an extension of your.

So it will be interesting to see how these memory models do change because I think on one side, yes, you have the model that is able to remember stuff specifically about you probably contained on a web page and the LLM is able to use that context every time it gets a response. But I think an actual self evolving model that manages to improve its own performance and is able to dynamically update its own parameters. I think that kind of thing is going to be more akin to AGI asi. And I do wonder that regardless of any of these enterprise uses, are there going to be some of the large labs working on this? For example, is Super Intelligence going to be working on this? And I do wonder what kinds of things they're doing behind the scenes to have these kind of models that are able to dynamically update themselves. Of course there is the security risk with, you know, these kinds of models. We don't really know what they can kind of do. But it is quite fascinating to take a look at how these experiments are changing the field. And of course one of the things that you might not remember is the fact that MA1 is going to be far smaller than any other model. And MA1 or MY1 is basically the model that is 500 billion parameters. It's being developed by Microsoft and apparently it's going to have some new things in it. So it will be interesting to see what kind of model that is and how they're moving forward with that model at Microsoft because there's a lot of things going on behind the scenes that happened when Microsoft acquired the company Inflection and there's an entire team working on a large language model and consumer products that are going to be developed into some amazing things. So let me know, you guys think about this crazy stuff, and I'll see you guys in the next one.