Today in health, it we're going to take a look at a Gentek AI that is using agents in AI. We're going to take a look at a YouTube video that I just watched. See what channel it's on Matthew Berman's channel. And he takes a look at. Andrew Ning. Who I will give you some more detail on, and he talks about agentic workflows in a presentation to the Sequoia capital team.

Stay tuned. We'll get to it in a minute. My name is bill Russell. I'm a former CIO for a 16 hospital system and creator of this week health. I said, have channels and events dedicated to transform healthcare. One connection at a time. We want to thank our show sponsors who are investing in developing the next generation of health leaders.

Notable service now, enterprise health parlance, certified health and Panda health. Check them out at this week. health.com/today. All right. Hey, everything we're talking about, it's always on our news site. This week, health.com/news. Check it out. We are still raising money for Alex's lemonade. Stand.

Thanks to you. We are up over $40,000 for this year, 50,000 last year, 40,000. This year, we're up over 90. Our goal is 150 by the end of this year. If you want to be a part. Hit our website top banner. For Alex's lemonade stand, go ahead and click on that to end childhood cancer. All right. One last thing. Mentor someone share this podcast with a friend or colleague use it as a foundation for daily or weekly discussions on the topics that are relevant for you and the industry.

They can subscribe wherever you listen to podcasts. Alright, here we go. I continue to research on the the different frameworks and models for implementing AI. This was specifically a generative AI in this case. And what we can do.

A lot of the research is what we're doing for media here.

But at the end of the day, it's working with information it's generating. A new types of information. And in that scenario, it is applicable across healthcare as well as many other industries for that matter. This talk specifically is Andrew Ning. Prominent computer scientists known for his contributions to artificial intelligence co-founder and former head of Google brain ex chief scientist at Baidu. Co-founder of course SA Coursera. Educational tool online, which a lot of you are familiar with. His educational background, UC Berkeley, MIT, Carnegie Mellon. Suffice to say smart guy. Namings talk to Sequoia capital.

If you're not familiar with Cory capital leading venture capital firm in Silicon valley. Known for its successful track record and its portfolio companies constituting over 25% of the NASDAQ. Total market value. So very successful. Venture capital firm. NINGs talk to square capital focused on agents and their potential in AI. Advocating. The powerful capabilities of agents.

When powered by models like JBT 3.5, asserting they can outperform or perform at the level of GPD four with the use of agents. So there's a lot of different design patterns with regard to agents. I found this video fascinating. I will link to it in the show notes it's worth watching and listening to you will get ideas as he talks about this.

So let's talk about agents in AI Ning differentiates between non agent and agent based AI workflows. He talks about, he emphasizes the iterative nature of agent workflows. He argues that, that agent's. Are capable of performing distinct roles in collaborating. And that leads to a superior outcome than traditional AI approaches.

When you just ask a question of Chacha CPT, it's called zero shot. So he talks about zero shot versus this iterative nature of the design patterns of agent-based AI. And so the example he throws out is he illustrates this by with coding benchmark and. There's a, an agent-based approach using GPT 3.5 and it surpasses the performance of Four. In that zero shot coding test.

So if you go to two chatty. And say code this and versus doing that and using one of the design patterns of agent-based AI. You can actually use 3.5 with agent base that is recursive. A model of looking at the code over and over again, identifying its own errors and those kinds of things, and it will outperform GBT four. Hopefully, you're getting the picture that I got as I was listening to this, which is, by utilizing agents, either collaborative agents or adversarial agents, you can get a better solution.

And we do this in life all the time. We write a paper and then we give it to somebody to proofread when they're proofreading it, they are collaborative, but they're also adversarial. They're looking at it to say, Hey, you made a mistake, you missed a period. You missed a question, mark. Your flow of your sentence is incorrect, so forth and so on. So we use these patterns in life and it turns out they can be applied here as well.

So let's talk about the design patterns in agent based AI. So he outlines for them. Reflections. And this is where agents assess and improve their outputs. This is where. Anyway, but I'm going to, I'm going to stay short here and then come back to it. Tool use agents utilize predefined tools and functions for specific tasks.

We saw this in the chatbot beauty models. You can now. Go out and call specific tools and they will return. They will perform some. Some algorithms on the data and return to you a result and that returned results can then be plugged back into the general AI model. You have planning. This is where agents strategize and plan their actions. They're not just, zero shot again.

They're not just responding. There's actually a plan to do that. And then there's multi-agent collaboration. This is multiple agents working together each with specialized roles enhancing the overall quality output. And I think a lot of healthcare. Development right now is counting on this multi-agent collaboration. That as you might use the GPT. Three or four or even five when it comes out. To do some things, but you're also going to run that through very specific medical model.

So this multi-agent collaboration. Becomes a thing. Let me go to each one of these. In a little more detail, give you a little bit of an example. So reflection being the first one. Reflection is an agent based. AI mechanism. It involves an agent evaluating its own output and identifying improvements.

I do this within Zapier, so I put a prompt in one step. Where the data goes through something and then it comes back out with a result. And then I put it right back through. Chat GPT to take a look at it and it's an iterative process and I try to, wow. Anyway, I can get into the specifics, but at the end of the day, every interview I do goes through Jaci GPT in multiple. Layers and it pulls out key quotes.

It pulls out a, an outline of our discussion. It can then create an article based on the key quotes and the discussion that we had and those kinds of things, but it's multiple steps and it's the same model looking at the information over and over again, generating more higher quality and more accurate output with each. With each iteration. The example that Nin gave was a an AI agent tasked to write an essay right after completion of the draft, the agent uses reflection to analyze the work, identify weak arguments, repetitive sections, or unclear statements, and then revise the essay accordingly. And that process may occur multiple times each iteration improving the essays. Coherence in depth. All right.

So reflection reflections one you can do today with whatever model you're using. Tool use. So to use referrals refers to the ability of agents to employ specific tools or functions to accomplish tasks. These tools can be prebuilt functions or external API APIs that the agent can InfoQ as needed. Enhancing the capabilities.

The example I gave is an AI agent designed to assist with data analysis. It can use. Statistical analysis tools to compute metrics. Whatever happens to be standard deviation. If the agent needs to visualize that data, it can call upon the graphing tools to create the charts and whatnot. So he talks a lot about that.

That computer vision was the reason that this was created, this ability to call other things, do something. Cause it. In early versions of generative AI. It was essentially blind. Like it couldn't see a, an image or make an image out. And so they created these API tools that could go out. Use the third party who said this image is of a pier in an ocean.

And then it would feed that back to the generative AI tool to do with it, whatever it was going to do. So tool use is a very powerful. Mecca is a multi-agent tool. The third is planning and I'll be honest. This is the one I understand the least, but let me so planning involves the agent's ability to strategize and outline steps to achieve a goal.

This process allows the agent to consider various approaches. Predict potential outcomes and select the most attractive path to success. And as an example, an AI planning agent in logistics could analyze different routes. Considering factors like traffic, weather conditions. Delivery windows to plan the most efficient delivery schedule. It assesses various scenarios to devise a plan. That minimizes travel time and costs while ensuring timely deliveries. And again, I guess it's multiple agents, it's saying. Hey, you D you tell us what the best path is to develop the best. Result for this. And then it will iterate a couple of times and then come back and say, this is the best path to get there.

So that's planning. And then the last one is really powerful, man. I just, it just, it gets my mind going and it's multiagent collaboration. So multi-agent collaboration is the interaction and cooperation between multiple agents, each with specialized capabilities or knowledge. These agents work together, sharing insights and resources to solve complex problems more effectively than they could individually.

And the cool thing about this is you're going to have agents working with agents. And so you have this ability, depending on your token size and your ability and budgets and those kinds of things, you can have these models working with each other very rapidly to come up with. Just fascinating results. The example he gave was actually medical one and medical diagnosis system. A different AI agents could specialize in various aspects of patient care.

One agent might focus on analyzing symptoms, another on medical history, a third on diagnostic imaging, and yet another on treatment options together, they can provide a comprehensive diagnosis and treatment plan with each agent contributing its expertise to the collective effort. And you get an idea for these are really powerful design models. I always come back to this idea of architecture, design models.

It's important. If you're going to be somebody who's just buying AI, it's important to understand this. Are you going to buy the concept of one model to rule them all with single shot or zero shot kind of inputs and outputs? Or are you looking at in asking the right questions around these models to say, look, does it allow for multiagent and how do you facilitate multiagent?

Because we believe that multiagent will be the way that healthcare, especially on the clinical side is delivered. He talks about the, the practical applications of agent-based AI in the future. He emphasizes the practicality and potential of agent-based AI and improving productivity and problem solving. He speculates on the evolution of agent-based AI highlighting the importance. Of fast token generation and the ability of agents to perform complex tasks efficiently.

One of the things about grok is just the fast token generation and the. Massive amount of tokens you get, because you have the ability to use grok as that second. Tier to look at the information and to iterate on it because it can do it almost instantaneously. When you're talking about the fast token generation and machine to machine interaction you can get results very quickly. He concluded this, that agent AI based workflows represent a significant advancement in AI, potentially achieving performance levels of future AI models through current technologies.

And he mapped that out with 3.5 and four of chatty PT. But he believes this is going to be a more effective model moving forward. And I think it's fascinating again, I'm going to put the link in the show notes for you so that you can watch this video. As well. I think we should all consider doing more research, listening to more experts like like Dr.

Ning and and considering the various models, especially when we are signing contracts, suspend. Hundreds of thousands, if not millions of dollars for these various things. Anyway, that is all for today. Don't forget. Share this podcast with a friend or colleague. You said as a foundation for mentoring. We want to thank our channel sponsors who are investing in our mission to develop the next generation of health leaders.

Notable service now, enterprise health parlance, certified health and 📍 Panda health. Check them out at this week. Health. Dot com slash today. Thanks for listening. That's all for now.