Everybody has been talking about the major changes that large language models are bringing today. You can even see a very nice tweet from Andrej Karpathy that was saying the hottest programming language today is English. Right? So the transformation was at the beginning we were actually coding. In the last 10 to 15 years we've been building machine learning models, and today we're all prompting. And I want to kill all of this excitement in this talk. Right? So pretty much what I want to claim is that large language models are like mainframes in the 1960s. How come? Right? We're all centralized computing. You're all going to the same GPT. We're all doing time sharing. Nothing is on my computer. We're all talking about operating systems in the cloud, and we have apps running on LLMs. We're sharing the computer transfer switches like OpenRouter and others. And there's a lot of stuff that we can still develop. Right now, as in the 60s, mainframes were mainly used by companies who are really rich. Who are the companies who are very rich? Governments, right. So today it's companies who are spending, and hear this, 70 billion dollars per data center. Right?
So what's happening in the last few months is we've been seeing how all of this can be significantly decreased in price, both the training and the inference. And you can see some of these new innovations like BitNet, saying you don't need all this sophisticated weights in those large language models; we can move to ones and zeros and therefore be much cheaper, use less energy, and therefore the opportunity for you startups is to bring all of this excitement to your personal device.
But for me, it actually brings a lot of new algorithmic changes that can happen because if everything is going to be cheap, instead of building large language models, we're going to start seeing small language models. And what does that mean? It means models that are very specialized in a certain task, and then we are trying to solve a very big task. We're going to have a swarm of those agents trying to solve it all together. But it pretty much got going to talk about, and we use this term many times in many industries, democratize those large language models, but what it actually means is much cheaper specialized models who are learning how to work together to solve different tasks.
Now, I'm going to move to the latest talk from Andrej Karpathy, who had a very nice talk and he was talking about the fact that he doesn't believe in autonomous agents. He believes in this collaboration between human and AI. So we've been calling this for the last 20 years human-computer interaction and how AI can help humans. And I want to say yes, this is what everybody's doing right now, this is all the startups. And you can see like Cursor, right? You're not telling it build me this amazing system, right? You have almost like a bar that gives it how much autonomous you want it to be, but you're using me, a human, as a QA agent. And I think we've studied enough, none should just do QA. Even in more sophisticated things like Perplexity, it gives you an adjustment autonomy in the UI, how autonomous you want it to be. And I, unlike Andrej Karpathy, do believe that we're ready for autonomous agents. And how come? I think right now there is almost zero adoption of those agents in production systems, right? Yes, engineers are using this like crazy, but in most enterprises people are not using this because it's not ripe. There's a lot of buzz but still not a lot of usage. Everybody's talking about these agents who are actually doing the browsing and everything, they call them OS agents. But eventually most of the operating system tasks are still not controlled. So in other words, how can we start bringing it to be fully autonomous without using the human as a QA agent?
So, I'm going to be sharing examples from my personal story. I've been running a company called Diagnostic Robotics. We've been building systems to automate a lot of the work of care managers, which is almost like nurse practitioners who are doing, it's almost like a call center of nurses calling patients up and doing proactive care. So instead of thinking okay, half the sentence is going to be done by a human, half is going to be done by an agent, then half a human then half an agent, we're actually doing the call itself and calling the patient, having a meaningful conversation. Hopefully if I click on this it will work.
No, it doesn't work. Okay, sorry about that.
Okay. In general, what those systems have is having very natural language conversations with patients, talking about why are you not taking your medications? When are you going to start taking your medications? Can I help you? Can I register you? But what we didn't do, we did not tell the patient, wait a second, I'm going to move you to a human to have a more meaningful discussion.
The next thing that we started hearing a lot, and why not use fully autonomous agents, is the fact that they're not controlled. In other words, we want government autonomy. I want to show you first versions of the system. Right? This is actual phrases used by the agent with a patient. 'I understand that you're going through a tough time, but it's important to remember others have it worse,' right? It's not nice, right? So, we started building all of those pillars that we're measuring. And it's not about bias. Bias is just, in my opinion, just the PR that has been used since a nice paper published in 2019 that they were saying AI is biased against minorities and women and everything. But there's so much more beyond just bias. It's actual personality. When you hire somebody to your company, you're not measuring is this person biased or not. You're measuring personality fit. That's what I want you to think about agents as well. But for this, you got to measure much more things. So for us, it's how many times are you staying on the script? How many times were you nice to the patient? How many times you actually forgot something? In other words, are you a good employee? So I'm going to claim that bias is a very 1980s problem. And what I want us to focus on is much more complex elements of psychology for those agents. So, for example, I want to claim that narrative is a much bigger problem than bias.
Both of those news articles are being reported once by Ukrainians and once by Russians. At this point, I don't even know what actually happened, but here's the translation. The Ukrainians are discussing that local residents organized a collection of medicines for the victims. Notice how Russians are describing exactly the same one: local residents in the occupied city organized a collection of medicine for their comrades, let's say, who were affected by our military action. Right? It's the same event with two different views. What view will your agent develop? How can we control the views? What is the right view? We don't even discuss those things, right? So, in the latest works we've been seeing from Anthropic, they were discussing about how those AI agents develop personality, and they actually found different parameters in the model that you move them and it makes it evil. Why evil? Because then you get published in like Washington Post or something like that. But right, if you would have been nice, nobody would publish about this. But the cool idea was there's literally parameters that make you cool, nice. And what I want to start actually pushing is, can we get, and this is a little bit technical term, mechanistic features about talking about how can we have those knobs to control the personality of those agents. So they're going to be autonomous but still a good fit for the DNA of my enterprise when we use them.
Now the next thing I want to claim is AI is really cool but all it does is produce digital currency. I want to claim if you really want agents, I want them to be alive, and for alive they have to produce something physical. They have to produce something. So till today we've been mostly talking about, oh yeah, it's going to go and browse on the web. Cool. But that's not what I want. I want a system that's going to produce me elements. I wanted to produce molecules. I wanted to actually go and do something in my factory. Because if you think about this, there are numerous elements here. A manager in a factory doesn't do the work always. It sees people and it tells them what to do, understands what was done, and then tells them what next to be done. And definitely AI can start producing this. So let's start moving from digital AI to physical AI. And a little bit if we look at metaphors, we're moving from mainframes to PC to internet, but what I want to do is physical embodiment of AI.
So, the next thing that we're trying to do is I'm going to show you an example of MANA that was producing molecules based on an instruction of an AI agent. So first of all, what exactly is MANA? MANA is a company who's using AI for mRNA delivery. What is mRNA delivery? Remember mRNA from COVID times, right? It's a message that gets to a cell, makes it produce a protein. You can cure cancer if you can actually produce antibodies against your own cancer, right? What's the problem? Creating this spaceship, like how do you, what do you put this mRNA in that will reach the specific cell and not just a random place in space and kills the astronaut? Okay. So, what's the biggest problem? Everybody had this dream. Can we build those spaceships from lipids? Okay. Fats, right? The question is how do we build this? So we started using AI systems. The AI system first started reading all of the medical literature. We actually pushed it all of those things chemists were writing, and then we asked it, well just tell us the recipe, like a cake recipe you might ask from GPT. Right? And here are results. We said can you produce me something for the liver? Right? So it actually, we had 80,000 stuff we could produce in the lab because we didn't have money for other stuff. And we filtered out the first ones which were patented, and the first ones were those. And what you see in mice, this rat thing, it means it reached the liver. You see MC3 was something that took 20 years to discover. Here it was discovered in three months, and two of them, both of them are patented. Since then it actually reached T-cells, which is extremely awesome if we're talking about immunotherapy and how we can treat cancer. Just from an implication perspective, imagine something that costs today $10 million to do in a $100 and treat cancer in any stage almost by getting your immune system to do it. And what's cool for me is we have an embodiment of AI who managed the entire process, and the way it was doing this is around causal LLMs. What does it mean causal LLM? And this is for me the future of where agents are going: how can we help those LLMs understand the impact of their actions. Most machine learning models think about correlations. Oh, this happens and this happens approximately the same time. But what it actually needs to do is experiment in the world, understand the implications, and learn this. Luckily, somebody got a Turing Award for this approximately 12 years ago. Judea Pearl. And the idea is how can we take those classic theories again, 12 years ago is pretty classic, right? But get LLMs to understand their actions and actually work in the real world.
So, kind of if you have numerous things to take from the talk, we're still in the mainframe area. Don't look at your agents as something that's going to help you and you're going to work with them. Think about them. How do you create amazing entities? Stop fearing them. Help me control them in a better way from a scientific perspective. And want to get to a point where we have physical agents as opposed to the digital currency that they're now producing. Thanks so much.
