Imagine you fall ill and go to visit the hospital, but this hospital is fully run by AI agents.

The doctors, the staff, the nurses powered by Large Language Models like GPT-4.

How do you think that would work out?

I can imagine what's going through your brain.

You might be thinking things like malpractice lawsuit, 100% fatality rate, zero chance of survival, right?

Well, not quite.

If you get nothing else out of this video, just remember this.

Doctor agents, again, AIs run on Large Language Models, can keep accumulating experience from both successful and unsuccessful cases.

Sounds kind of like humans, doesn't it?

Simulation experiments show that the treatment performance of doctor agents consistently improves on various tasks.

The knowledge that doctor agents have acquired in agent hospital and the simulation of a hospital is applicable to real world Medicare benchmarks, right?

That's interesting, isn't it?

After treating around 10,000 patients, this would take over two years for doctors in the real world.

The evolved doctor agent achieves a state of the art accuracy of 93% on a subset of the MedQA data set.

If you're new to this channel, it's a point to understand that this isn't a fluke.

This is at this point, we can say it's kind of a pattern.

The pattern is this.

We are able to run certain AIs, certain neural networks and train them in simulation in a digital environment.

We're able to extract their learning, their digital brains, if you will.

That experience that they've achieved in simulation is applicable, is useful in the real world.

For example, right now you're seeing this little robot.

This is NVIDIA's simulation called, this is probably Isaac Jim or something similar.

They have a few different ones.

You're seeing that same robot in the real world with various wheels attached to his hands and feet, or his hands and feet are basically wheels.

It learns to do it in simulation.

It does it how or many millions of times.

The simulation has the same physics as we have here in the real world.

It's slightly randomized here and there to make it a little bit more robust, a little bit more adaptable.

Looks like by 2000 iterations, it starts being able to open that door.

Time runs a thousand times faster, 10,000 times faster.

You're able to train them very, very rapidly.

When you take them out of the simulation, they're able to very robustly perform those tasks in the real world.

I would see that action movie.

But before we dive deeper into this paper, we probably should touch on this little tweet.

This was posted on April 20th, 2024.

Ethan Malik.

We've talked about him on this channel before.

Excellent person to follow.

Not a very controversial person.

Not a very outrageous person.

He posts studies about AI.

Warden professor studying AI, innovation and startups as a substack, like a blog with a newsletter called One Useful Thing.

He posted a paper called Automated Social Science, Language Models as Scientists and Subjects.

No big deal, right?

Here's that paper, MIT, Harvard.

The idea was pretty simple.

Again, they take Large Language Models, LLM, again, kind of the central theme with a lot of this stuff, right?

They're basically asking, and I'm simplifying quite a bit here.

They're asking, can this LLM, can it be like a scientist?

Can it come up with hypotheses, run experiments, gather data?

Can it, if you will, do science?

Again, I'm simplifying.

Please don't yell at me.

Social scientists start by selecting a topic or domain to study.

Within that domain, they identify interesting outcomes and causes that might affect the outcomes.

These are the variables.

The proposed relationships are the hypotheses.

For example, I have a hypothesis that your luck is greatly increased when you hit the like button and subscribe to this channel.

It's really improved.

We can test that out by, well, you hitting the like button.

We'll you know, go ahead and gather some data.

Go ahead and do that now.

This is us designing an experiment to test these hypotheses by inducing variations in the causes and measuring the outcomes.

After designing the experiment, social scientists determine how they will analyze the data.

They recruit participants, run the experiment and collect the data.

You're the participant, hit the like button.

They analyze the data and estimate the relationship.

Basically, this is kind of the scientific method, right?

This is how we move our understanding forward.

Up to very recently, 100% of this, this progress, this process was made by humans.

Very recently, we're seeing AI take over just a little bit here and there.

All right, we've talked about GNOME, millions of new materials discovered with deep learning.

This is a DeepMind thing.

This little robot sits here creating new materials.

There's a separate sort of AI that comes up with certain potential recipes that this robot then tries to create.

This, what I assume is a blast proof case in case the robot messes up and something goes boom.

Here's how many materials in this case, crystals, right?

This is how many humans were able to sort of create that were stable.

With various computational methods, we had this circle, which is bigger.

Of course, this GNOME or GNOME, right?

It's this big circle, right?

AI is advancing science already in some areas.

In this case, basically these LLM agents, they create certain social scenarios, hypotheses, then they build the AI agents with certain relevant attributes, right?

For example, one of the experiments was a negotiation of some sort, right?

They designed those interactions, model estimation, data collection, and then they run the experiment.

They talk about all the things that worked well, all the things that maybe didn't work so well.

It wasn't perfect.

But at the end of the day, here's their conclusion.

We show that such simulations produce results that are highly consistent with theoretical predictions made by the relevant economic theory.

This could potentially allow for controlled experimentation at scale, right?

We can run these simulations at scale to test various things, to gather data, right?

We could re-yield insights that would generalize to the real world.

You may be wondering, well, why was this controversial?

What was the controversial piece of this whole thing?

This was it.

I caught a portion of this as it was happening.

People basically would come into the thread and they got very nasty.

They did not like this idea of collecting data and assimilation.

They didn't like the idea of Large Language Models potentially contributing to science.

They weren't very nice about it.

Ethan Molek says, yeah, he's going to delete that thread.

People are just insulting the paper authors without reading the paper or understanding the context of the fields being discussed.

They, meaning the authors, didn't ask, they tweet about them and didn't expect the tweet to travel so widely, getting nasty on Twitter.

Again, I saw a little bit of it.

I'm not sure exactly what got people so riled up, but having talked about some of these topics on YouTube, I know firsthand that there's some small but very vocal minority of people that go into this murderous rage when you suggest that Large Language Models have some ability to reason, to understand, to have world models.

They usually don't have something that can be called arguments that show why they think what they think.

It's mostly just insults and getting nasty.

I had a lot of misgivings about talking about this paper.

On one hand, I'm very busy because I'm launching my course in how to build AI agents.

It's going very well.

We're up to almost a thousand people.

It's been live for just a few days.

Check out the links down below if you're interested.

On one hand, I don't really have too much time to delve into this if you will, but on the other hand, talking about this new paper would probably make a lot of those nasty emotional people very, very mad, which I just cannot say no to.

Today let's talk about Agent Hospital, a simulacrum of hospital with evolvable medical agents.

Yes, AI agents that think, reason, learn, and evolve, and there's nothing you can do to stop it.

Let's get started.

This is kind of like their main overview of what this hospital looks like.

They're saying this is an overview of Agent Hospital.

It is a simulacrum of hospital in which patients, nurses, and doctors are autonomous agents powered by Large Language Models.

Agent Hospital simulates the whole closed cycle of treating a patient's illness.

Disease onset, triage, registration, consultation, medical examination, diagnosis, medical dispensary, convalescence, and post-hospital follow-up visit.

An interesting finding is that the doctor agents can keep improving treatment performance over time without manually labeled data.

This is kind of the big deal.

Normally when we train AI, we need lots of data.

For example, if we're training an AI that recognizes pictures of animals, we need lots of pictures of animals with labels.

They call it data pairs.

A picture of a dog, and then along with that, something that says this is a dog.

Many of those, so they can kind of generalize how dogs, what they look like.

A million pictures of dogs that are labeled dog.

I think in this case, since they're talking about dermatology, so let's say it's some sort of a skin condition, I don't know, some rash.

Maybe there's a bunch of pictures of this rash that's labeled with the name of that particular condition, right?

That's normally how we do it.

But here it seems like in the simulation, these doctors, these AIs keep learning, keep improving without manually labeled data.

Without human medical professionals sitting there and pouring over data and labeling it and making sure it's correct, you can say that they generate their own synthetic data within the simulation and then learn and improve from that data.

The central goal of this paper is to enable a doctor agent to learn how to treat illnesses within the simulacrum.

They're proposing a method called MedAgent Zero.

Here's kind of a map that they've built.

Looks like this is the entrance, right?

You have the registration window, pharmacy, waiting area, the triage station, biochemical testing room, et cetera, et cetera.

Looks like they've used this tiled map editor.org to build out some of this stuff.

I gotta say, I'm very excited about the intersection of video games and AI, specifically AI agents.

I mean, the fact that we can use generative AI to voice dialogue, create dialogue, all that stuff, that's cool.

But something about having actual autonomous agents running around in games and affecting the gameplay world, I gotta say, that's making me very excited. And looks like they also use this phaser.io, a fast, free and fun open source framework for various browser games.

I guess here's kind of what that looks like.

They use something like that to build this, right?

This is a hospital sandbox simulation environment, tiled and phaser.

Tiled is a map designing tool.

Phaser is a framework to manage the movements and interactions of the agents on the sandbox.

They create the various agents and the information about the roles of these agents is generated using GPT-3.5.

I think they're just using GPT-3.5 just for kind of generating the background for these people.

I hope they're not using it for the doctors.

Can you imagine going to a hospital and the doctors running on like GPT-3.5?

You would walk right out, right?

It's a little LLM joke for you there.

They create a series of medical professional agents, including 14 doctors and four nurses.

Doctors diagnose diseases and create treatment plans, whereas nurse agents focus on three hours supporting the day-to-day therapeutic interventions.

Here's example agents.

We have a radiologist, his duty is spelled out, skills, et cetera.

We have a patient, internal medicine doctor and a receptionist.

If they're running on GPT-4, I'd be curious to know how much this whole experiment cost.

I believe the Stanford experiment, the social simulacra we've covered on this channel kind of had a similar thing, but there it wasn't even a hospital.

It was just a little village where people went around their business and tried to organize a party.

I think that one blew through thousands of dollars of OpenAI API credits.

Planning, we've covered how to approach these agentic planning, how to create planning architectures for agents in that Stanford simulation.

The interesting thing there, they don't go into as much detail about planning as I think they did in the Stanford paper there.

The trick was starting very broadly with the entire day, right?

Kind of like, what are the big rocks in the day?

That's the first thing you ask ChatGPT, this LLM.

You kind of keep fine-tuning it like, okay, so then break it down by every hour, then every half hour, then every 15-minute intervals, etc.

Here, these patients in this case, right, they have their daily planning and dynamic planning, right?

Upon arrival at the hospital, they'll go to the triage station, and then based on what happens there, their actions and movements will follow based on what happens there if they tell them, oh, we got to get you into urgent care, whatever they go there.

Here, this person gets sick, KM goes to the hospital at the front desk, the triage, I guess this means registration, they head over to registration, and then into consultation, then into medical examination, diagnosis, then to the medical dispensary, convalescence, and then there's this arrow pointing back to the disease starts over again, which this sounds like a nightmare if it's just an endless loop, but okay.

Planning for the medical professionals, they just kind of fulfill their responsibilities based on their designated role.

They're trained from two types of actions.

One is practice.

Actually dealing with the patients, giving them care as patients are assigned to them during their shift, and the follow-up information from patients that will help them polish their medical records experience.

The second part of their kind of learning or planning aspect is actually learning.

Outside of working hours, they engage in studying past medical records to gain clinical experience, reading medical textbooks to expand their knowledge.

It's this idea that a Large Language Model, right, this soulless machine, if you will, sits there in a simulation, reading a book in its off time when it's not working, and is able to expand their knowledge to improve the results from reading a book.

I mean, to me, that's very exciting.

That's very interesting to think about.

But I think this is where a lot of people get uncomfortable, which by the way, if you're one of those people, I'd love to know more.

What is it that is so disturbing about this idea?

Because again, there seems to be a lot of anger directed at the people that are talking about this.

We won't cover a lot of this stuff, but I do encourage people to read the paper.

But let's focus on the core things here.

The evolution, right?

The improvement of these agents, they propose their kind of framework, their strategy med agent zero, and there are two important sort of areas, modules in this strategy, the medical record library and the experience base successful cases, which are to be used as reference for future medical interventions are compiled and stored in the medical record library for cases where treatment fails, doctors are tasked to reflect and analyze the reasons why they failed.

It's still a guiding principle to be used as a cautionary reminder for subsequent treatment processes.

Here's kind of their med agent zero method, right?

You have the patient query, so the doctor can use medical record retrieval prompting, so aka asking the patient experience retrieval, then it looks like the medical agents, the doctors generate an answer.

I'm assuming this is their diagnosis, right?

Is it correct?

Is it not correct?

Here we have kind of the golden answer, meaning this is the truth, kind of the ground truth, like this is what we know to be the true answer.

If somebody comes in with disease X, right, we make sure that the doctor comes up with that.

If the doctor says something else, we say, no, that's incorrect.

The doctor goes into the reflection mode, thinking about it again, as they say here, distilling guiding principle to be used as a cautionary reminder for further treatment, right?

If the answer is correct, right, then we know that that's the case for that hospital patient.

We add that to the medical records library, again, to be used in the future.

If they come in, we're able to retrieve it and say, oh, you had this in the past, right?

If they don't get it correct, and after reflection, right, after thinking about it, they attempt to re-answer it, right?

If they get it wrong again, we abort it, right?

It looks like they're just not getting it.

But if they answer correctly on the second try, like it's added to the valid experience, to the experience base.

I gotta say, this is similar to probably how the human learning process is, probably everything except the fact that there's probably no golden answer.

There's no absolute truth that we can compare it to usually.

Also, if you're wrong twice, we don't just pull the plug on you.

But other than that, this is very lifelike.

This process builds the record library, the experience base keeps expanding, the more patients are treated.

The way those records are built, so they use text embedding into vector space by this embedding model provided by OpenAI.

Basically how LLMs store data isn't in text format, how you and I would, they use something called vector space.

I think the easiest way to kind of visualize that of these little 3D charts where each word has a set of traits.

Based on those traits, you can kind of visualize it in a 3D environment.

If you're looking for a cluster of words, for example, they have kind of similar meaning or similar tone or whatever, you can kind of organize it like this.

Here's kind of one that's showing the different directions of like sky, wings and engine, right?

A helicopter drone and rocket will kind of be in this space, whereas a goose eagle bee would be in this space.

All things with these characteristics would kind of be in that same vector space, which is interesting to think about because I feel like kids when they're learning to speak English or whatever language they're learning, you can kind of think of it as them using something similar to this.

They learn the meanings of words by kind of contrasting them to the other words that they know.

Right there, these clustering of words, whether it's words that are opposites or words that are groups.

All that is to say that I don't think that the fact that LMAI agents act kind of human-like at the end of the day, I don't think that's an accident.

The point here from the experimental results is that after they keep training on 10,000 samples, 10,000 patients, they show a continuous improvement, a continuous increase.

As you can see here, there's a rapid increase in precision, right?

There's like this rapid increase up to maybe 2000 training samples.

I mean, it continuously increases from there on, but I mean, there's a little bit of a diminishing returns, but the point is they keep improving.

They keep learning.

It does seem like they're using GPT-3.5 for this because probably GPT-4 would be, I mean, 10,000 patients, however many, 14 doctors or however many they had.

I mean, that would cost a lot of money.

Keep in mind that if they just substituted GPT-4 for this, the results would likely be even better.

Here's the breakdown of various diseases, how well they're able to Figure out what those are.

For some of these, I mean, they started out at like 20% accuracy or 40 or 60% accuracy.

Over time, like on this one, this has got to be around 90% accuracy.

There's an obvious learning process that happens.

All right, so they used both GPT-3.5 and GPT-4.

I apologize if it was a little bit unclear to me which one they used.

They used both.

As you can see here, they compared it to just prompting these models for the answer.

These were the results.

They use chain of thought as well as these other sort of approaches, these methods for testing them.

But MedAgent Zero, which is this paper that we're covering, so training them in a simulation with 10,000 iterations, 10,000 patients.

I mean, both GPT-3.5 and GPT-4 beat out all the other ways of using that model.

They are the number one.

GPT-4 comes in at 93%, which outperforms human experts in the MedQA data set, which humans experts are around 87%. So GPT-3.5 approaches a medical professional on that data set.

GPT-4 beats them by quite a bit, 87% versus 93%. Also keep in mind they can run the simulation a lot faster, right?

10,000 patients.

I mean, it would take a doctor many years to accumulate that much experience.

Here that could be ran pretty rapidly and every year you'd be able to run it faster and faster as technology improves, as both hardware and software and these neural nets improve as well.

If you enjoyed this, please hit thumbs up to get this out to more people.

If this bothers you, tell me why.

I am genuinely curious because I keep seeing this pop up and it's just not just me on my YouTube channel, but obviously even respected professors and researchers that post their work on Twitter.

Actually here I found one comment that actually, so once Ethan Molyk deleted the original tweet, a lot of those kind of really negative responses disappeared.

But here's one that from what I've seen, because I briefly went into that RABBIT hole of that thread, and I think this is kind of representative of what was said.

I think this person kind of captures the complaint, but just imagine this, but much nastier and attacking people personally.

He's saying, to be fair, you kind of casually post ultra dystopian stuff and are like, isn't this tech so cool?

Smart man in your area, but you seem not detached from the world a bit.

Maybe he meant you seem too detached from the world a bit if you didn't see this coming at some point.

Again, I really think this is really representative of what the people were saying.

On this channel, I try to be unbiased and try to understand a lot of different perspectives, or at least be open-minded to it.

I wouldn't understand this, but I just don't.

He's saying that studies like this are ultra dystopian.

I don't even understand why would it be ultra dystopian?

Accelerating progress by the use of machines, improving our ability to understand the world around us, improving our technology, improving our ability to help people that are sick to diagnose.

Maybe potentially have AI doctors that can help people without resources.

There's many places in the world where there's poverty, there's no doctors, there's no access to high-quality medical care.

What if these trained AI agents could help people for pennies to at least get a correct diagnosis?

How's that dystopian?

I almost feel like this ultra dystopian would be if we outlawed this technology, that we didn't allow humanity to benefit from what it could bring to the world, because that would benefit the people that already have everything that they need.

If somebody said that would be ultra dystopian, I could agree that would be.

But this I just don't understand.

I also notice how, and a lot of these comments have that in common, they do tend to get personal.

There's personal attacks.

Here they're saying, well, you're detached from the world if you didn't see this coming.

You may be smarter over here, but you're kind of clueless over here.

That's what a lot of these comments had in general.

They're not saying, I disagree with what you're saying, and here's a respectful breakdown of some points that I disagree with you with.

Most of them were wrong and you're clueless and you're a horrible person.

That was the kind of combination of those things, which again, I have a hard time grasping where that's coming from.

We've covered Sam Altman talking yesterday to the kids at MIT because it does seem like he's getting a lot of this from, this is kind of sad, from these college students at these elite universities, or at least some subset of them.

He posted this, I think, as a response to that.

Just the timeline fits to where I think he was talking about what he experienced during his college tours, talking to these students at Stanford and MIT, etc.

He's saying, using technology to create abundance, intelligence, energy, longevity, whatever, will not solve all problems, it will not magically make everyone happy.

But it is an unequivocally great thing to do and expands our option space.

To me, it feels like a moral imperative.

At the bottom here, he's saying his most surprising takeaway from the recent college visits that this, what he said above there, is a surprisingly controversial opinion with certain demographics.

Again, on this channel even, you can go back like a year, see some of those videos I did about GPT-4 reasoning where I talk about specifically this, me dealing with this, where you cover some cool new AI breakthrough that seems to be very beneficial and move science forward.

It's surprisingly controversial.

With certain demographics, it's a minority of people, but they're really incensed about it.

He says, prosperity is a good thing, actually.

D, D, growth.

That's it for me for today.

If you've enjoyed this stuff here on this channel and would like to join me in my quest to build the biggest and most valuable private forum where we can talk about AI, how to build autonomous AI agents, where we can all upskill and learn to use AI to prepare for what's coming in the future.

I launched a natural 20.

I'll put the links down in the description.

We're almost thousand strong.

Join us, won't you?

Let's build it together.

My name is Wes Roth, and thank you for watching.