Navigating the Clinical AI landscape

Summary

In this dynamic webinar, hosted by Endpoints News and sponsored by Mendle, we delve into the evolving landscape of clinical AI and its pivotal role in advancing precision medicine. The discussion features insights from Mendle's co-founder and CEO, Karim Galil, and Unlearn AI's founder, Charles Fisher. Both leaders share their vision on harnessing AI to revolutionize healthcare, from making medicine more objective to enhancing clinical trials. The conversation navigates through challenges, innovations, and the potential of AI in healthcare, aiming to bridge the gap between technology and medical advancements for a more objective and efficient future in medicine.

Takeaways

• 🛠 The webinar, sponsored by Mendle, focuses on navigating the clinical AI landscape and charting the path to precision medicine.
• 👨‍⚕️ Speakers include Karim Galil, MD, co-founder and CEO of Mle AI, and Charles Fisher, CEO and founder of Unlearn AI, both sharing insights into the integration of AI in healthcare.
• 🧪 Mle AI's mission is to make medicine objective by creating the world's largest index of patient journeys, leveraging AI that understands medicine like a physician.
• 📚 Unlearn AI focuses on solving AI for medicine, particularly in making clinical trials more efficient and accurate through the creation of digital twins.
• 🖥 Both companies emphasize the importance of AI innovation specifically tailored to the clinical domain, beyond general AI applications.
• 📈 The discussion covers the AI market's segmentation into innovators, tool builders, and wrappers, highlighting the roles and impacts of each in advancing clinical AI.
• 📍 The critical role of data in AI development is underscored, especially the challenge of training AI models with complex, real-world clinical data to avoid inaccuracies and biases.
• 💻 The webinar explores the potential of AI to revolutionize clinical trials, with Unlearn's digital twins exemplifying how AI can simulate patient outcomes to streamline trials.
• 📢 Talent and culture within AI companies are deemed crucial, with a focus on the blend of engineering and research expertise necessary to drive genuine innovation.
• 📰 The dialogue encourages skepticism towards overhyped AI claims, advocating for a grounded understanding of AI capabilities and limitations in healthcare applications.

Q & A

What is the mission of MLE AI as described by Karim Galil?

MLE AI's mission is to make medicine objective by enabling the world's largest index of patient journeys, leveraging AI that understands medicine like a physician.

How does Charles Fisher's background contribute to his work at Unlearn AI?

Charles Fisher's background as a scientist with interests at the intersection of physics, machine learning, and computational biology, combined with his experience in machine learning engineering and computational biology, contributes to his work at Unlearn AI, focusing on inventing new AI technologies for medicine.

What distinguishes Mendle's approach to clinical AI from traditional methods?

Mendle's approach is distinguished by its use of a clinical NLP platform that allows command over the full spectrum of data, integrating both structured and unstructured data into a computer-ready knowledge representation for sophisticated clinical reasoning queries.

How does Unlearn AI aim to transform clinical trials?

Unlearn AI aims to transform clinical trials by using AI to create digital twins of individual patients, allowing for the forecasting of medical outcomes under various scenarios, thereby making trials more efficient and accurate.

What challenges do innovators face in the AI landscape according to the webinar?

Innovators face challenges such as the need for long-term R&D investment before product development, the need to differentiate amidst a saturated market of AI companies, and the difficulty in educating potential clients on the nuances of AI technologies.

What is Hypercube, as mentioned in the webinar?

Hypercube is a product launched by Mendle designed to enhance clinical analytics through an AI platform specifically built for the clinical domain, improving accuracy and scalability in clinical trial matching and other applications.

Why do both speakers emphasize the importance of publishing peer-reviewed papers in AI?

Publishing peer-reviewed papers is emphasized as it provides scientific validation of the AI technology, helps differentiate genuine AI innovation, and facilitates engagement with scientific and regulatory communities.

According to the webinar, how does AI technology face limitations in healthcare applications?

AI technology faces limitations in healthcare applications due to challenges like context window limitations, hallucination risks, and the need for explainability, especially in complex clinical reasoning tasks.

What is the significance of domain-specific AI technologies as discussed in the webinar?

Domain-specific AI technologies are significant because they address unique challenges and requirements of specific fields like healthcare, potentially leading to breakthroughs and advancements beyond the capabilities of general domain AI.

How do the speakers view the future of AI in healthcare and clinical trials?

The speakers view the future of AI in healthcare and clinical trials optimistically, believing that AI will continue to evolve and provide innovative solutions for previously unsolvable problems, though they acknowledge the current limitations and the need for continuous research and development.

Outlines

00:00

📢 Introduction to the Webinar on Clinical AI

The webinar, hosted by Car Bball from Endpoints News and sponsored by Mendle, focuses on navigating the clinical AI landscape towards precision medicine. It features two prominent panelists: Karim Galil, MD, the co-founder and CEO of Mle AI, who shares his vision of making medicine objective through AI-driven patient journey indexing, and Charles Fisher, CEO and founder of Unlearn AI, who brings a deep background in physics, machine learning, and computational biology. The introduction sets the stage for a discussion on how AI can bridge the gap between medicine's slow advancement and rapid technological progress, aiming to make clinical research scalable and medicine more objective.

05:00

🌟 Exploring Mle AI's Mission and Innovations

Karim Galil delves into Mle AI's mission to objectify medicine by leveraging AI to create a comprehensive index of patient journeys. He outlines the challenges of integrating technology with the slower pace of medical advancement and highlights Mle AI's journey from its founding to the launch of its first product in mid-2023. Galil discusses the innovative clinical NLP platform developed by Mle AI, which aims to revolutionize clinical analytics by enabling sophisticated queries on consolidated patient data, thereby addressing the inaccuracies and inefficiencies in traditional clinical data analysis.

10:02

🔍 Unlearn AI's Approach to Enhancing Clinical Trials

Charles Fisher introduces Unlearn AI's mission to pioneer AI solutions specifically for medicine, emphasizing the company's focus on creating AI technologies that tackle unsolved medical problems. He shares insights into Unlearn AI's digital twin concept, which forecasts individual patient outcomes in clinical trials to increase efficiency and accuracy. Fisher explains how this innovative approach can significantly impact the design and execution of clinical trials, particularly in neurology, by providing more precise and efficient methods for evaluating treatment effects.

15:03

📊 The State of AI in Healthcare and Clinical AI Innovations

The discussion shifts towards categorizing AI companies into innovators, tool builders, and wrappers, with a focus on how each contributes to the AI ecosystem. The conversation then pivots to the specific challenges and opportunities in applying AI within the clinical domain. Both speakers emphasize the importance of domain-specific AI technologies and discuss how companies like Mle AI and Unlearn AI are pushing the boundaries of clinical AI by addressing unique challenges in healthcare, such as the need for sophisticated reasoning and knowledge representation in AI systems.

20:05

🧠 Debating the Capabilities and Limitations of AI in Medicine

The dialogue deepens into a philosophical and technical exploration of AI's capabilities and limitations, especially in complex clinical reasoning. Both speakers share their views on the distinction between machine learning and deeper AI functionalities, such as reasoning and knowledge representation. They discuss the current state of AI, including the challenges of data context, hallucination, and explainability in clinical applications. The conversation highlights the importance of integrating machine learning with structured knowledge to address the nuances of medical data analysis.

25:06

💡 Challenges and Opportunities in Commercializing AI for Healthcare

The conversation turns towards the practical aspects of commercializing AI in healthcare, discussing the discernment problem faced by buyers in a saturated market and the importance of distinguishing genuine AI innovation from mere tool use. Both speakers share their experiences with regulatory approvals, the importance of publishing peer-reviewed research to establish credibility, and the strategic focus on innovation to navigate the complexities of the healthcare industry. They stress the need for AI companies to demonstrate real-world efficacy and ROI to survive and thrive in the competitive landscape.

30:06

🚀 The Future of AI in Clinical Trials and Ethical Considerations

The final part of the discussion focuses on the potential impact of AI on clinical trials, specifically the concept of digital twins and how they can streamline the trial process by reducing the need for control groups. The conversation then addresses criticisms of current AI methodologies, such as those by Gary Marcus, and debates the need for new AI theories to advance the field. The speakers also touch on the ethical implications of AI in healthcare, emphasizing the need for transparency, accountability, and rigorous validation in deploying AI technologies in clinical settings.

Keywords

💡Clinical AI

Clinical AI refers to the application of artificial intelligence technologies in healthcare, particularly in clinical settings, to improve patient care, diagnosis, and treatment outcomes. In the script, Clinical AI is discussed as a transformative tool for making medicine more objective and efficient by leveraging AI to index and understand vast amounts of patient data. The discussion highlights the potential of Clinical AI to bridge the gap between medicine's slow advancement and rapid technological progress, as exemplified by the missions of mle AI and unlearn AI, which aim to innovate in the clinical domain.

💡Precision Medicine

Precision Medicine is an approach to patient care that allows doctors to select treatments that are most likely to help patients based on a genetic understanding of their disease. The webinar focuses on navigating the clinical AI landscape to chart a path to Precision Medicine, emphasizing the role of AI in enabling personalized patient journeys and treatments, thus making medicine more tailored and effective.

💡Innovators

Innovators, as discussed in the script, are companies or individuals who develop foundational AI models or approaches to address previously unsolved problems. They are characterized by their focus on research and development, employing highly specialized knowledge in mathematics, computer science, and other fields to create breakthrough AI solutions. Examples from the script include mle AI and unlearn AI, which are positioned as innovators in the clinical AI space, developing unique AI technologies for healthcare applications.

💡Digital Twin

A Digital Twin in the context of clinical trials, as explained by Charles Fisher of unlearn AI, is a digital replica of an individual patient. This concept involves using AI to forecast future medical outcomes for a patient based on their pre-trial data. It serves as a powerful tool for enhancing clinical trials by simulating control groups and reducing the need for patient enrollment, thereby accelerating the trial process and aligning trials more closely with patient expectations.

💡AI Ethics

AI Ethics concerns the moral principles guiding the development and application of artificial intelligence technologies. The script touches on ethical considerations, especially in the context of healthcare, highlighting the importance of transparency, bias mitigation, and the responsible use of AI to ensure patient safety and fairness. The discussion around AI hallucinations and biases in training data exemplifies the ethical challenges in deploying AI in sensitive areas like medicine.

💡Machine Learning

Machine Learning is a subset of AI that enables systems to learn from data, identify patterns, and make decisions with minimal human intervention. The script discusses machine learning in the context of developing clinical AI applications, such as patient data analysis and the creation of digital twins. It emphasizes the importance of machine learning in AI innovation but also acknowledges its limitations, such as hallucinations and the need for human oversight in healthcare applications.

💡Knowledge Representation

Knowledge Representation in AI involves the methods by which information is structured so that AI systems can use it to solve complex problems. The script references the use of knowledge representation in clinical AI to ground machine learning models in medical reality, reducing errors and enhancing the accuracy of patient data analysis. This concept is crucial for enabling AI to perform sophisticated clinical reasoning and make informed medical predictions.

💡AI Hallucination

AI Hallucination refers to instances where AI models generate false or nonsensical information as output. In the healthcare context discussed in the script, AI hallucinations pose significant risks, such as providing incorrect medical analyses or patient diagnoses. The conversation highlights the need for mechanisms to detect and mitigate hallucinations, ensuring AI systems' reliability and safety in clinical applications.

💡AI Adoption in Healthcare

AI Adoption in Healthcare concerns the integration of AI technologies within medical practices, research, and clinical trials. The script explores the challenges and opportunities of adopting AI in healthcare, including regulatory approval, cultural acceptance within traditional healthcare institutions, and the potential to significantly improve patient outcomes and operational efficiencies. It also addresses the discernment problem, where distinguishing between genuinely innovative AI solutions and those with limited value becomes challenging for healthcare providers.

💡AI Talent

AI Talent refers to the skilled professionals, including researchers, data scientists, and engineers, who specialize in developing and applying AI technologies. The script emphasizes the importance of AI talent in driving innovation, particularly in the healthcare sector, and discusses the competitive landscape for attracting and retaining these highly sought-after individuals. The success of AI projects, especially in complex fields like clinical AI, is often directly linked to the quality and expertise of the team involved.

Highlights

Introduction of webinar on clinical AI and precision medicine, sponsored by Mendle.

Karim Galil, MD, discusses his mission with mle AI to make medicine objective using AI to index patient journeys.

Charles Fisher's background in machine learning, computational biology, and his company Unlearn AI's goal to invent new AI technologies for medicine.

Discussion on the distinction between innovators, tool builders, and wrappers in the AI landscape.

Insight into Mendal's AI platform for clinical analytics and the challenges of bridging technology and medicine.

Unlearn AI's approach to creating digital twins for more efficient clinical trials.

Debate on the definition of AI and its implications for healthcare.

The importance of domain-specific AI technologies in advancing clinical fields.

The challenges of AI adoption in healthcare and distinguishing between genuine AI innovation and superficial applications.

The significance of peer-reviewed publications as a metric for assessing AI companies.

Discussion on the ethical and practical challenges of AI in medicine, including data biases and model explainability.

The critical role of talent in the success of AI companies, emphasizing the need for a mix of domain experts and engineers.

Perspectives on the future of AI in healthcare, emphasizing continuous innovation and ethical considerations.

The potential of digital twins to revolutionize clinical trials by reducing the need for control groups.

The importance of maintaining an open dialogue and collaboration within the AI and healthcare communities to foster innovation.

Transcripts

hi everyone I'm car bball with end

points news I'm your moderator for

today's webinar navigating the clinical

AI landscape and charting your path to

Precision medicine we're sponsored by

mendle and I'm excited to welcome our

panelists joining us we have Karim Galil

MD co-founder and CEO of mle AI mle's

mission is to make medicine objective by

enabling the world's largest index of

patient Journeys leveraging AI that

understands medicine like a physician Dr

Gil's experience as a physician

demonstrated that medicine does not

Advance at the same rate as technology

with mendal he aims to bridge this Gap

facilitate clinical research at scale

and make medicine truly objective Kareem

is an entrepreneur by Spirit his first

company Krypton Works LED healthtech in

the Mina region with customers including

Fortune 500 companies we're also joined

by Charles Fisher CEO and founder of

unlearn AI Charles is a scientist with

interest at the intersection of physics

physics machine learning and

computational biology previously he

worked as a machine learning engineer at

leap motion and a computational

biologist at fiser he was a Philipe my

fellow in theoretical physics at a CO

normal superier in Paris France and a

post-doctoral scientist in biophysics at

Boston University Charles holds a PhD in

biophysics from Harvard University and a

BS and biophysics from the University of

Michigan if you have any questions

during today's webinar our panelists

have reserved some time at the end so

please hit the Q&A button at the bottom

of your screen when you think of your

question this webinar will be available

on demand tomorrow to re-watch or share

with colleagues and now I'll pass it off

to Kareem to get us started

Kareem um thanks Gary and um Charles and

I met through an investor like few weeks

ago and uh we started talking about the

state of AI in healthcare and we found

that we have a lot of uh common pain

points and we thought like the

conversation

should get extended so we decided to do

an endpoints webinar so today is more of

a casual conversation about how we view

the state of AI in general state of AI

in health care what are the challenges

and we both have a two startups who are

in the space of AI but not necessarily

competing to the same product so it's

going to be interesting to see two

points of view from two angles um uh

around the clinical AI space but before

we get started we thought we give quick

intro about Manda quick intro about

unlearn um mandal started in 2017 but

actually we launched our first actual

product in around mid 2023 and

essentially the rest of the journey has

been around R&D and we see that as the

Journey of almost AI company today it's

different than building a software

company you have to kind of have

long-term greed re-invest in a lot of

R&D before you can actually have a

product we were very fortunate to have

um investors who believed in the mission

from the very goodg go the company is

the most funded clinical NLP company

today um but around um end of 2023 we

were very excited to launch hypercube

and this year we are growing our

customer base uh we're going to share a

couple of use cases maybe today um what

we do at mandal is if we can go to the

next slide is we view the state of um

clinical analysis today like everyone is

using standard stack think of like the

snowflake the big queries of the world

keyword search we do standard ETL work

and that process is is is is pretty

difficult it cost a lot of money but it

is very prone to inaccuracies we've seen

a lot of cases where someone would come

and say we have 10,000 callon cancer

patients um running our platform we only

find 2,000 um but if we go to the next

slide what we have done here is we've

built a clinical NLP platform that

allows you to command the full spectrum

of your data so it's not only your

structured data that being said we're

also able to ingest structured data like

claims and and and fire and XL7 we

consolidate all of that in a computer

ready knowledge representation something

that the computer can reason over and

then we provide a platform that allows

you with no code to query this data and

do very sophisticated uh clinical

reasoning type of questions like find me

patients who have X Y and Z but has

never went through a treatment for

example um

we're very excited about the early

results there is a paper that we just

submitted uh for the chil conference and

uh we see that the product if we also go

to the next slide

um um next slide

please has been uh taken to test against

uh Chad llama for a use case like

clinical trial matching so group of

researchers at tupen has been

interrogating three systems for certain

uh number of clinical trials and we're

able to see that our uh approach has

been uh showing better accuracy but also

has been significantly better and

designed to actually scale in the

clinical domain from a cost and and from

a processing time so that's essentially

what mle does we are U clinical

analytics 2.0 through an AI platform

that has been built for for the clinical

domain um that being said I'll let you

Charles introduce un

learn

yeah thanks thanks Kareem um so unlearn

is uh basically an AI research company

um who uh we we're focused on maybe if I

had to say it really broadly I'd say

solving AI for medicine that would be

the mission of of the company um also

founded in 2017 uh now raised about $135

million um uh totally about 75 people

and you know as I said I think that our

Focus has always been that we're we're

an AI research company who's focused on

really inventing new kinds of of AI

technologies that allow us to solve

previously unsolved problems

particularly in medicine that's that's

our real goal as a company so we end up

on a lot of these kind of like AI uh AI

lists um but our main focus today is uh

in terms of our goto Market strategy is

that that we work with biotech and

pharmaceutical companies to help them

run uh more efficient um and more

accurate clinical trials so if you go to

the next

slide um and yeah just click just click

to the animations animations but you

might as well just click them so it's uh

effectively what we what we aim to be

able to do to create is a digital twin

of an individual person and what that

means is uh let's imagine that that

Charles participates in the clinical

trial I'll take data from myself uh

collected before the clinical trial

input that into some uh

uh AI train pre-trained model and I

wanted them to be able to forecast what

would happen to me in the future um and

we'd like to be able to do this for

effectively any measurement that you

want to forecast it at any point in the

future for anything you want under any

scenario for any person uh so it's a

very challenging problem uh but that's

the type of type of thing that we work

on and then we we partner with biotech

and pharmaceutical companies mostly in

phase two and three clinical trials

starting in neurology uh to to uh

include patients digital twins to to

make those trials

better and next slide I think goes back

to you Kareem but I could be wrong nope

it does um so we we thought just to set

the stage today um kind of uh we put

together this slide so just kind of see

like set the stage around how we see the

state of AI in general and specifically

around the clinical domain and we

essentially see there's three big

buckets of of AI related companies you

have folks who are actually building

foundational models or foundational

approaches towards U an AI problem and

we like to call those the innovators and

those are essentially folks who would

hire mathematicians PhD and computer

science more like a R&D sophisticated

approach to again an AI problem that

hasn't been solved before and needs a

different unique approach um but then

there is also a lot of companies to

today that are coming in the tooling

space and those make the job of the

innovators easier uh they provide you

with compute power they would provide

you with marketplaces for models they

would provide you with u uh uh levels of

abstraction that allows you to build AI

faster or to deploy AI faster then we

also see that there is almost now um we

were making the joke almost every couple

of hours there is a new AI company and

bunch of those come and sit into the

wrapping space and and those would

essentially leverage technologies that

are built by the innovators using tools

built by the tool Builders to put an

interface on top of a foundational

technology and essentially build a a use

case or or a product so there wouldn't

be essentially um a breakthrough in AI

it is more of like a productization or a

layer of of of ux on top of a model

built with the innovators on the

innovators side the challenge that we're

seeing is uh most of the work and the

attraction has been happening around

General domain Ai and became a term and

the idea is you want to build an AI that

can do everything under the sun and

there are companies that have been on

the Forefront of that such as open AI C

here anthropic Google deep mind but what

the premises of today's webinar is about

is more that there is a breed of

companies now that are doing the same

type of innovation they're actually

solving very hard AI problems but

specifically for the clinical domain or

specifically for a clinical use case um

and we believe that this is going to be

beyond the hype of AI this is going to

be actually the state where you're going

to find a lot of domain specific AI

technologies that are built say for

healthcare for fintech for for defense

kind of applications and we see Mandel

and unlearn to be one of those companies

that are in the clinical domain yet kind

of innovating and building breakthrough

type of Technologies and one question we

thought we could start off today with is

like what is AI to begin with I I think

there is a lot of uh there isn't really

like a strong definition there is a lot

of debate actually how would you how

would you describe how would you define

AI but again to set the stage here is

how we think of the market this also set

the stage maybe we can start with you

Charles kind of how would you define an

AI what's the difference between Ai and

just a piece of software sure um so I

actually take a very narrow view of what

AI is uh it in in in my view today um so

historically if you look back AI was

sort of a field of research about

creating computers that are able to

learn and act on their own without any

human oversight effectively and that was

what AI was and there's lots of ways

that people can do that um but what

really has become the dominant Paradigm

in the last

decade has been things that I would call

connectionist connectionist and so

connectionist models are these are a

sort of class of models that are really

inspired by the way the brain works I

might some of us might say Loosely in

Loosely inspired by the way the brain

works but that's effectively that's

effectively the concept so to me today

I'm actually primar I primarily Define

ai as as the use of algorithms that are

inspired by the way the brain works

uh to to solve problems without uh a

tremendous amount of human oversight of

course there's still some human human is

the teacher effectively but that's

that's the way that I Define it so it's

actually relatively narrow so I exclude

like classical machine learning

algorithms like random forest or spms I

actually exclude those uh from from from

the definition of what would be AI in my

mind and I I I agree with that

definition I uh I think the simplest

definition again I I unlike Charles I

don't come from a technical background I

come more from a clinical background but

the definition that made most sense to

me coming into that space it's a

clinical model of the brain it's a

software sorry a computer model of the

brain it's essentially you're trying to

get the computer to model a function of

the brain or multiple functions of the

brain but to that extent it has been

interesting because uh the way our brain

works and there's this really

interesting book think fast and slow you

have two systems right you have a very

intuitive system that doesn't think much

that doesn't take take time to make

decisions but it would essentially like

when you're when you swerve your car

when you're driving when you're

breathing you're essentially using that

system but you also have another system

that is slower more arithmetic that's

the system that you will use to solve an

algorithm or to think of a big idea and

machine learning in many different ways

mimics that intuitive fast pattern

recognition system but there isn't a lot

of Technologies or companies that has

also been working on uh on on on the on

system on on system one the more

arithmetic slower ones and essentially

the way I view machine learning large

language models neural networks it's

it's like a parrot in many different

ways that is looking at a pattern and

repeating it and when it comes into text

it makes you to it makes you confuse

fluency with intelligence but very very

few technologies have also system one

what what's what's your take on that and

how does that affect the use case that

you are working on

yeah so

um I don't know where to start start on

the use Cas so I'll start on the use

case and now back into the to the

broader question um you know what we

focus on as our use case is really

different I think from what everybody

else focuses on um kind of in in

actually most of the AI research

Community um because we're not we use

some stuff that that involves like

language processing but actually we use

that more internally uh for our our own

like internal data cleaning efforts

that's not like our products are not

built off of those kinds of things

externally um we actually use a totally

different style of of neural network

architecture that no one else in the

world uses um and so the original idea

for what we were thinking about was um

not so much how do you create AI to

solve problems that humans are good at

but how do you create AI to solve

problems that humans don't know how to

solve and so I'm thinking about like to

me the the pro the the the one like

prototypical example it's not from us

where

where somebody did this is Alpha fold

right if you give me the sequence of a

protein if you write it down in front of

me and you say Charles what's the

threedimensional structure of that

protein my brain cannot compute that for

you I don't know how to do that um and

so but those are the kind of problems

that I'm really interested in it's like

how do we create new kinds of AI system

to solve the problems humans don't know

how to so basically solve scientific

problems um and but I I'd say that a lot

of the work that's been going that

that's gone on over the last decade has

been kind of more around how do we

create AI That's able to automate

problems humans are good at um and so

those would be things like language Tas

or same thing with images right you

think about uh recognizing objects in

images or now even generating images

those are things that humans are really

good at and we're trying to create

systems that automate that so um anyway

so so to me like our our use case is

super duper different but Kem I forgot

the beginning of your

question beginning of the question that

most of the like people basically use

machine learning and AI as if it's one

word it's like synonyms each other when

essentially machine learning is more of

like if we use the early discussion

about AI is a model of the brain the

brain has two systems there's the

learning type of a system and there is

more the arithmetic more system um well

yeah so so for the work that we are

doing today I so I agree with you I I

think that all AI right now all AI

methods are sort of um relatively simple

input output systems um and uh all of

them uh so so my view is you said kind

of at the beginning I I wouldn't go so

so um far as to say that things are just

parrots I think that they're

interpolators uh they're very good at it

um but I don't think that that the kinds

of things that we've worked at today are

are by themselves going to get us to

True sort of general or super

intelligence I think that there are

breakthroughs that are required to get

to that and um I know people are working

on it uh you know reasoning systems and

other kinds of systems like that but I I

think that today they're primarily going

to give you when I say they they're

interpolating what I mean by that is

that they they give you similar answers

to similar things they've seen before if

you ask it to create some brand new

thing that like it's never seen anything

like that before uh and you're

extrapolating to some really truly brand

new situation today's AI systems don't

don't know how to do that yeah um in in

our

use case here at Mandel our our number

one job is to synthesize a patient

Journey from complex data so we would

take thousand PDFs coming from an

oncologist some claims data coming from

a provide from an insurance company and

our job is to synthesize um coherent

patient Journey from many conflicting

facts so you would find in one page the

doctor saying the patient is uh

metastatic uh but then you look at the

pathologist saying the patient is stage

one and now you have two conflicting

pieces of information that you need to

reconcile which requires a lot of uh

clinical reasoning and our experience

with machine learning only approach or a

large language model only approach that

those systems as you said are really

good at pattern recognition and

repeating something that humans are good

at if they are fed with the right data

but it cannot do complex clinical

reasoning like a physician or or a

clinical expert would would would would

would be able to and the only way around

it is to couple it with some sort of a

knowledge representation or something to

ground the rules of medicine or the

concepts of medicine or ground the

machine learning model into it and it's

so interesting because um Yan laun which

is one of obviously the few people that

won the touring award for for neural

network newer architectures was saying

the only the neural network today is not

even at the level of a brain of a cat

and the only way we can come closer is

if we can couple it with symbolic Ai and

with knowledge representations and we

don't know how um so that's where we

have been spending a lot of time how can

you do complex reasoning with an AI

model

essentially

um yeah I was just gonna ask maybe um

and maybe this is too technical of a

question

but could you explain a little bit as to

how how you would View kind of that as

being different from fine-tuning just an

existing model on like say met text or

something like that like what how is

that going to be different so when when

when you're looking at like a lot of the

models today have limitations around the

context window and um a clinical record

by default will exceed that context

window easily which increases the

chances of of hallucinating the other

thing essentially how how those models

would work is it would bring terms that

are related to each other closer to each

other so an adverse event and fatigue

would probably be closer to each each

other to each other in the model but

once you move to another area you move

from uh oncology to Immunology fatigue

becomes more of a symptom rather than an

adverse event and how can you teach the

model to start understanding the nuan

differences between those two things um

I would say the other thing is no matter

what you do when you're are using a a a

especially a larger larger language

model the idea of hallucination becomes

incredibly dangerous when you are making

some sort of a clinical synthesis of a

patient journey and having like almost a

filtering layer that would tell you you

cannot have headache in your knees like

headache is probably not going to happen

in your knees ever and there is

impossible that you can have a stage one

patient who's also metastatic at the

same time and it's like kind of giving

you those like those things are

impossible to happen it's just like it's

against the the laws of medicine helps a

lot and decreasing the hallucination I

would say the third thing is

explainability in specifically in our

use case folks want to know why did the

AI why did the AI conclude that this

patient has adenocarcinoma and not any

other type of of of hisy and you have to

be able to trace back the logic and it

becomes incredibly hard when you're

having a machine learning only approach

to the problem I would say those are the

three things explainability

hallucination and the context window of

of of medicine in general is generally

High you know that's that's an

interesting idea because um I would bet

that those th well I would bet the um

those unallowable things if you actually

just trained a model on the raw data

like to just train a model on Raw data

that you would actually learn those

things because the data do have mistakes

I I absolutely worked on a project I

remember where a patient had had uh like

a leg amputation at some point and but

their medical records later on have them

having problems with that leg which was

just

amputated so if you're just training

entirely on the medical records you're

actually learning some of the mistakes

that are in the Raw data themselves yep

uh but then you still have to deal with

some Al some of the other problems like

the hallucinations and and and and

explainability but it's also very

interesting because the way folks

practice medicine in general is is has a

lot of intuition into it so if you start

training on like the only way you can

get to that is you have to train almost

on every medical record in the US coming

in from different states with different

regulations with different ways of of

approaching things to like take out the

bias but the other approach would be is

like hey no matter how you practice

medicine these are the like the Periodic

Table of medicine those are the laws

that would govern how we would how we

would approach um problem

um I think one one one topic we talked

about is as like both of us are

obviously commercializing an AI product

into the health care space and this is

not an easy space to commercialize

anything in general let alone an AI

product so H H how is your experience

with adoption what how are you seeing

the buyers today on the heal everyone

wants to buy an AI product but do you

think everyone have the right criteria

can you can you talk more about that no

I mean obviously like you said it's a

difficult space uh to to work into um I

think for us we have found

um I think I I'm going to go with the

two the two things that have that have

been positive and I've for us and maybe

two things that I would say that are

that are barrier um uh so I'll start

with the positives um one of the biggest

things for us is is that I think when

you're deploying sort of AI based

systems into into any sort of medical

context um there's usually some kind of

extra layer of validation that people

are looking for um and so in our case we

work in phase two and three clinical

trials like we we are working as part of

a study protocol actually like our our

AI is like written into the protocol of

these studies and so we wanted to be

able to go through um and get regulatory

buying um and so going and getting uh we

got a regulatory qualification for one

of our methods uh in from the EMA in

2022 uh got a letter from the FDA at the

end of last year saying that they concur

with that and so um working with The

Regulators has actually been really easy

uh I would say um me I wouldn't say it's

fast like you know on a startup time

scale working with The Regulators hasn't

been fast but it has been a really

positive experience and that I think I

think the regulatory agencies in my

they're were thinking about this

correctly and that they're not actually

that focused on like you know what is

the underlying algorithm and and how

does it work and that they're much more

focused on the context of use and making

sure that however whatever AI system

you're creating when you deploy it into

some use case that it really work works

for that and so I think that that's

really good and then you know getting

those qualifications has really helped a

lot with uh with our sort of go to

market the other piece though that I'd

say has been a positive is just finding

Innovative people uh there are people in

the industry like uh who are uh they

want to do things differently they want

to see technology get adopted or they're

frustrated with the current you know e

rooms law situation where truck

development is just getting slower and

more expensive and they want to do

something different

I think finding those people has been

great now the barriers are um one of the

big barriers is I I've coined this

actually back my friend Austin Austin

Juan coin this back when we were working

at fiser together but he called it the

discernment problem and the discernment

problem is that uh there's a bajillion

companies that that say they are all

using

AI um and the people who we're trying to

sell to are not experts in AI um and

they have uh they to them all of these

companies look exactly the

same so how do you actually stand out uh

amongst all of these companies uh many

of whom are are not maybe doing anything

particularly Innovative but but you know

everyone's using the same language and

that would get to my second point which

is even inside the Pharma

companies uh I at today I feel like the

my definition of AI which I described as

very narrow which is like connectionist

based models that like neural networks

has been replaced by uh what I would say

the industry uses which is anything that

uses a computer this is ai ai so like

you ask like what's biostatistics AI

what's bioinformatics AI like all of

those things have just gotten rebranded

as AI um I hear lots of people for

example say they're like oh well we are

doing AI a we've been doing AI for 50

years but that doesn't make any

sense because the algorithms that people

are using for like modern AI systems

have been all like really developed in

the last 10 or 15 years so it doesn't

make sense if you say you've been doing

it for 50 years then you're doing you're

not doing uh you're not doing AI in my

opinion if you've been doing it 50 years

you're using other kinds of

computational methods so um so those are

that that's been my kind of broad

experience what what about you Karem um

I in my 's case we don't have uh I mean

there is pain in getting regulatory

approval and uh my heart goes to all the

effort that you had to go through that

uh but I think down the path to your

point it kind of gives you that stamp of

validation that can help a lot in your

goto Market in my use case we don't have

that so I have to only suffer from the

latter point that you made which is like

every AI come everything is an AI

company till proven opposite and uh to

your point also we are seeing the

innovators in in every org and those are

folks who are actually spending the time

um I always get a little bit

um I mean I'll put it lightly irritated

once I find someone with pure MBA

Investment Banking background who took

two AI corer courses coming in on a call

dcing and AI company um from on behalf

of a vendor and that essentially becomes

a big problem for us because um we're

not we're not looking at the same ground

bases and the example I always give my

team U when I wanted to buy a a a a ring

to my wife to propose I had no freaking

clue how to look to buy a diamond for me

it's a piece of glass still proven

opposite right everything just looks

shiny and it's only when I started

getting educated like here's the clarity

score here is this here is that it

actually started making sense to me why

would I pay five like 10K I'm not going

to say how much I paid for my wife's

ring but how much should I pay versus

like a 100 bucks right and I feel a lot

of the AI buyers today haven't went

through that exercise yet of kind of

understanding what are the criterias

that you should be looking at an AI

vendor what we are betting on is that a

lot of the pilots that are happening

today are going to fail and then

companies that actually did have true

technologies will will remain and that's

the state of every bubble that happens

you just raised money and we're seeing

in VCS saying the same thing they are

not betting big on a lot of AI companies

because they assume that most of those

Pilots will fail and those who are

actually going to pass 2020 for and show

clear Roi on the business or clear Roi

on the outcomes of for the patients are

going to be the AI companies that will

eventually survive that that that

wave um do you see the same do you see a

lot of Pilots happening in in around

your your use cases or is it like we're

either going to get in or not going to

even try the

product I mean you know we've been you

know in in our case I there are

definitely still companies that want to

run run like earlier stage like test

cases pilots and things um I would say

that in some cases we're a little bit

we're we we're we're some part beyond

that and that we are actually using our

methods in some phase two and and and

and we're uh working on some protocols

for some bigger phase three studies now

um so to a degree we're a little bit um

we're a little bit but I would say like

there's the innovators where we're just

beyond that pilot stage like like we're

deploying things now with some companies

where where there are some Forward

Thinking people um and then there are

other companies who we hear all the time

and I you want to know what makes me

irritated is this phrase which is I've

heard now multiple times that we want to

be your second customer in this disease

area or this indication whatever it's

like oh you want to be the second like

that is just like what's wrong in my

opinion that's like what's wrong with

medicine is that is that attitude uh um

so uh but yeah I mean I I certainly

think that there are lots and lots of

companies who are

um in in in the broader space I think

that there are a lot of companies who

are working or applying computational

techniques or applying various machine

learning Technologies to problems in

medicine or Healthcare I don't think

that there's a lot of companies that are

really trying to push the boundaries for

uh for how these things really work um I

think you know to me one of the biggest

challenges for for uh the thinking about

the way AI is moving right now for for

people in in this industry is they're

used to a really slow moving technology

um so you think about drug development

and Discovery right drug Discovery

development takes like 10 or 15 years to

develop and bring a new drug to Market

and so uh maybe it's a little faster for

a medical device I don't know five years

probably still seems fast though uh so

you're talking about you know five to 15

years is the pace o over which things

change um but in AI things are CH

changing right now every like every like

five to 15 days right and so part of

what I think is a challenge for example

is like you work you might work with

somebody if you're really an AI research

company then if you've done a pilot with

me three months ago you haven't piloted

my technology now you piloted my three

months ago my prior version we we have

the same problem where actually the

customer buys the product or gets into a

pilot and two weeks later we just

launched the newer model and the

customer is like two weeks be behind but

it's not only two weeks it's it's some

R&D work that has been happening for for

few but we talked before the webinar

about uh areas of like um when we were

setting this up we said look like we uh

areas of differences between both of us

would actually become would make this

webinar more exciting so here's one area

of difference that that that we can go

through when for example we were talking

about the idea of like uh headache in

the knee or something like that the like

what I've heard from you is what I've

hear a lot from from clients is like I

bet if you actually train a model on

more clinical data that essentially that

problem wouldn't be solved but when

mandal started 2016 essentially we have

a team of 100 plus clinical annotators

and we have tried every model under the

under the Sun from birth back in

2019 or something to like the latest and

the best today and it essentially does

not work it the systems that that amount

of data that we have um per patient is

too big where the model really gets

confused and starts itating nonsense

essentially around a patient Journey but

here we are like I I view you as a

clinical like as an AI expert but we

have to go through that debate back and

forth the problem is if you and like if

if if you and I are going through that

or someone on my team is going through

that there's some ground truth that we

can stick to it becomes 100x harder if

I'm talking to someone who has never

built an AI model who has never trained

an AI model it becomes a a losing battle

so what ends up happening is they do

follow their own path they come back six

months later or a year later and say you

know what we just lost whatever $5

million on that initiative and didn't

work one one actually true case that

happened with us is we've been working

with a big diagnostic company and they

have taken multiple approaches including

Google like approaches and there have

been voices in the company saying hey

like Google is always going to win over

a small company like Mel the approach of

machine learning is as an only approach

is good enough they went through that

they found say x number of patients that

they thought is eligible for a certain

project they had and using our approach

it was like 20% of that number and

essentially what happened was like well

if that's the case Mendel is wrong the

other company is right and we had to

push for manual error analysis of the

two sets of data and guess what ended up

happening is actually we were right but

this type of friction and like the proof

is on you type of type of uh type of a

headwinds is what I'm seeing today in

the market as uh the biggest challenge

for AI companies um including yours or

or or

mine yeah you know the other thing

though with sort of I would say the flip

side of some of these things that's

really good I I in my opinion about

working within kind of

space is that um the best marketing the

best marketing that any that any of us

AI companies can have and it's

absolutely been true for us it's just

publishing an actual research

paper just publish you publish research

paper and people at Pharma read it right

uh that's absolutely true um for us like

I've had people at the FDA uh ask us

questions about our papers on archive

not even peer-reviewed papers the FDA is

reading our archive papers so so I think

that that's one of the best parts about

working within a scientific field is

that if you engage not only not if you

actually really engage you be like here

are the algorithms that we're using

here's these things here's how they work

and here's examples of them working in

these areas that you get real scientific

engagement from the other side I like

1,000% agree with that and I think

actually talking about like how can an

AI buyer assess a vendor I would say if

you don't have any peer-reviewed

Publications that chances are you don't

have any sign on your team because it's

not only for marketing your team will

ask I want to publish about this I I

want to compare against the status quo I

think this is a really great point and

it's like uh easy easy to go through

that checklist if you don't find

peer-reviewed Publications com like

talking about the technology or

comparing the system to like status quo

that is for me is already a yellow flag

that that something is is not going well

um this year we've decided actually to

allocate around 70% of our marketing

budget towards publishing peer-reviewed

Publications rather than actually

spending on conferences and and and

things like that but I 100% agree with

that uh point so what's the most like U

provoking AI marketing message that kind

of made you feel like oh man that's come

on that's too much like something too

like like making a big deal of of an AI

capability that does not exist or

something that just foundationally

Incorrect and you felt like

that that is

outrageous

um that's actually a hard question I I

don't know if I have seen something

that's been like an outrageously

overhyped claim I don't know if I

actually don't really think that that is

that that is really what I have seen um

in fact I remember uh JP Morgan one of

the previous JP Morgans maybe like JP

Morgan early 2023 or something like that

I was on a panel with endpoints and I

said something about like by the end of

the year uh an AI system is going to

pass like the bar exam the medical the

the mcap the ls it's going to pass all

these things just one system that's true

right at the time people acted like it

was hype um so I actually think that at

least for my opinion AI within the

healthcare space is under hype I think

it is under hype because I think that

people focus so much on what has been

done they go what have you done and of

course that's very very important you

got to be able to get things into these

clinical trials you got to use them or

or whatever use Cas is um but again I

think that what people are missing is

that it's a rapidly advancing technology

um so I to to to me it's less around

like people's claims around what AI can

do or will be able to do I think that

that's largely underhyped to me the

stuff is that like when somebody's like

hey we use we use sophisticated Ai and

then you go what is it and it's like

it's it's like sklearn import PCA and

it's it's like well that's not that's

that's not really what I had in mind um

and so but that's actually often but

that's kind of kind of what I think is

challeng is that's really difficult to

tell sometimes uh what people really

really are doing um so you have to

really I'd say re I've seen papers like

this I'm not going to call out who

specifically but that I have like a

specific paper in mind for example where

the whole beginning of the paper talks

about some this is uh some deep neural

network based model or whatever I forget

it is and then halfway through the paper

it just switches to using a random Force

the title of the paper is like uh deep

learning for something something

something like this but the Deep

learning model loses to just the random

force in the in their paper and so the

whole rest the whole second half of the

paper just like and so we use this D of

forest for the rest of our analyses and

that's like a fairly well-known AI

startup in the Medical Healthcare space

that published that paper uh but again

they published at least they published

the paper so you can read the thing and

see what they really did

did so um talking about the you're

saying like well-known startup um coming

into the um we're seeing I'm seeing a

lot of the hype happening now on the m&a

side of things um so obviously every

company is trying to take an AI approach

that is going to boil down to like buy

or build and the buy is not only like

buy from a vendor it's actually bu the

vendor Al together and and use the AI

capability we're seeing a lot of that

happening um

but there is a fairly like uh also known

startup in in in in our field that has

been essentially using um apis from

another AI vendor which basically means

like you can just go buy your those apis

yourself they just have a cooler

interface than than what you you would

get from the original vendor and we just

got that just got to see that this

company acquired at like high high high

high dollar value and in my mind

man like for $5 million you could have

done exactly the same thing and you

could have just saved another like

hundred million or something um so I'm

seeing also some of that hype happening

on the m&a side where companies that

kind of got the right Banker or got the

right type of packaging um with some

papers that have very interesting titles

are and end up doing some interesting

transactions that wouldn't have happened

I guess had then their more rigor around

the AI uh criteria so what are the what

are the criteria like if you are buying

an AI company or if you are buying an AI

product in general what would you advise

people to think

of um well I honestly actually I might

go in a little bit of a different

direction on that on that Kem I even if

somebody wants to use like an

off-the-shelf wrapper they probably

still should buy an AI company to do it

um uh you know so so at least from my

own perspective I mean I can speak from

like you know I so I'm a PhD in

theoretical physics right um and I'm

really an AI researcher and I I

attempted to go and work at a big Pharma

company and I would describe my

experience there kind of as I I like my

colleagues and everything but I kind of

describe it as an organ rejection like I

like it was like there was a very clear

oil and water kind of situation um and I

I think that that's quite broadly that

that the the the culture of companies

who are really building these new AI

systems very different from the culture

of of a traditional company in in

healthcare medicine Pharma Life Sciences

um I don't think that they go together

um and so I I think that that that that

building these types of capabilities

internally uh having tried to do it

myself okay is probably impossible I

think it's probably impossible um and uh

that doesn't mean people are going to

try they're going to try uh but I but I

think it's impossible and so I actually

think that if people do want to have

these internal capabilities they they

should go to look for companies that

that uh for Acquisitions they should

leave those companies semi-independent I

mean so one of the examples of like what

and I guess I don't know too much I have

a few friends who are at it now but um

would be like preent design so so

Genentech uh in the drug Discovery Space

went out and acquired this this AI uh it

was a small a company it's actually

grown I think a lot since it's been part

of genetic but I think that type of

model in which if you're looking to

really do these things yourself that you

you do go out and acquire a company uh

for the culture and leave it independent

so that you don't mess its culture up I

actually think of as being important

actually yeah would you do the same if

so let's talk maybe this is the last

point and we can jump to questions let's

talk about Talent um in in my mind like

the biggest deterministic Factor around

an AI company is the talent uh I think

first Mar some VC said they assume

there's like lesser than 2,000 people in

in the world today who are capable of

actually doing a breakthrough in AI or

have the the right background for it but

anyways it's long way to say like talent

and we're seeing the talent war between

like Big T can even startups every every

capable AI scientist today is getting

multiple offers from from from multiple

places right so how how much weight

would you put on talent because to your

point even if company is just wrapping a

technology just acquired for for the

culture and I would agree with that but

I would say the rest of the statement

would be if they have ai talent because

if you're buying culture you want to buy

the AI culture but if you have a company

that mainly 100% are software Engineers

that has been wrangling apis those are

no different than the software Engineers

that you already have in in your company

and Ive seeing company like Mosaic AI

for example acquired by data bricks for

enormous amount of money but again it

boils down to the talent and even if the

company didn't have that much traction

so how much weight would you put on on

Talent or is your point that as long as

they have attempted some sort of

innovation they probably have the right

traits so still take the risk on that

but like how much weight would you put

on Talent essentially is my question uh

no I mean 100% although I wouldn't put

down I wouldn't necessarily put down the

software engineering talent I think

software Engineers are very important in

in in particularly in in ml um but like

you mentioned something earlier about

you know if your ml team is just a bunch

of NBAs

then that's that's not really it's not

really an ml team uh I do I agree with

that you know um I I think that you know

companies who but I I that's kind of

partly my that's that is my point though

effectively what I talk about that the

cultural difference which is that you

know at unlearn most of our employees

have PhD in physics or math or computer

science um and it's very very different

you know you go to to a life sciences

company and everyone also has phds but

there phds in biology or their MDS and

these are very different disciplines um

and and not only the different

disciplines because the different

backgrounds and stuff but I the there's

really there are cultural differences

between the between what people care

about who come go into those fields and

I think the challenge for a lot of the

companies uh who are trying to be sort

of this te new tech bio set of companies

it's actually how do you integrate these

cultures um like if you're a small

biotech and you're you know like say

you're recursion and you're trying to

like really simultaneously be a leader

in Ai and ML and in drug Discovery I

think it's hard probably uh to to to be

able to actually really truly have these

two separate cultures and get them

together so no I agree with you that

1,000% you want to be looking for

companies that are higher in math

physics computer scientists uh but I do

I I would say that engineering Talent is

as important as research Talent enormous

yeah an enormous amount of of of of the

progress in ml particularly the last

like five years has really been

engineering driven more than research

driven there is Yan Lun also said this

uh let's engineer the heck out of it uh

which is essentially if you look at open

eye it's like a lot of engineering that

happened on a technology like

Transformers that has already been there

before that but it's like if you a

clinical analogy it's like you cannot

have a hospital with nurses only or

Physicians only and I think it's the

same thing you cannot build an AI

company with Engineers only or with phds

in in math and physics only I think it's

a lot of collaboration between both and

it's a spectrum of how much the density

has to be between those um I know

endpoints want us to jump into questions

and we have a lot of questions here um I

think the most interesting question was

for you to say which medical startup

company has had that funny up the paper

not goingon to say that

yeah it's a published paper so people

can go look it up themselves yeah uh I

think one question for you that I

thought was interesting and maybe I can

take one for me is like for you was like

can AI cut down the time it takes to

carry out the clinical trials in

particular phase three oh yeah I mean a

thousand per this is this is exactly

what we're what our goal is to work on

um so the basic idea about what we do if

you think about so the applying this

kind of idea of a digital twin in in a

clinical trial so a clinical trial you

want to know what would happen to a

patient if they got a

drug in comparison to what would happen

if they don't and so what we're

effectively doing is we're using these

models to simulate one of those outcomes

for a patient um and if that model was

perfect you could actually run a

clinical trial that doesn't have a

control group so you could just say I'm

going to take each patient I'm going to

create their digital twin I'm going to

simulate what would happen if they were

if they were assigned to control group

and and just by doing that comparison

you would be able to to get an estimate

to the treatment effect

um like you were saying earlier about

hallucinations and other things today's

AI systems are not perfect uh so the the

types of things that we're working in

there's still randomized studies where

you do have some patients uh being

randomized to a real control group um

but it's fewer than an an typical study

um so you know we might work to say you

know half the size of a control group as

you would need in a typical study and

that cuts off you know potentially a

couple of hundred patients for a large

for a large trial makes a trial more

attractive this is why people are

participating in trials in the first

place is typically to get access to a

new experimental therapy so it's really

a win-win and that we're able to do

things that that speed up the trials for

sponsors but do it in a way that keeps

sort of uh the the science rigorous so

you get the same kind of rigorous

results out of these RCS as other ones

uh and aligns with What patients want

out of Clin iCal trials in the first

place I think one other question um for

both of us actually was um what do you

think um I know we're going to disagree

on this one so let's go through it what

do you think uh regarding the criticism

of deep learning by some known people

like Gary Marcus um that current AI are

in worrisome hands regarding AGI and how

would that impact medical AI in general

um and related more to both of your

companies how do you think another new

AI Theory could boost your companies you

want to take this

first sure uh I'm trying to decide how

mean to

be

um

so look I I'm sorry but like fun the if

you wanna if you want to to go out there

and think about like two today's AI

systems have have some limitations they

do they do that's why people like me

exists that's what my job is is to think

about how to overcome how to invent new

kinds of algorithms that that will

overcome these limitations and um do I

think that those new types of things

will with be within the realm of what we

call Deep learning I do I think that

there will be different architectures

and different training techniques and

different loss functions and things like

this than we use today but but I think

that they will fit with under the same

umbrella um I think that uh you know

again if you look at what we do at

unlearn we actually use a totally new

architecture we use a completely new

kind of architecture that no one else in

the world uses actually for our machine

learning approaches it's a deep neural

network but it's a really weird kind of

deep neural network super strange uh so

again we're inventing new kinds of new

kinds of machine learning like totally

new kinds of things I I think that um

the the criticism where people just say

like hey it won't work it's I I don't

understand I don't understand what what

Gary is doing frankly I like go out

there and try to build a neural network

then that that he has never ever

published anything in which he actually

trains a neural network that works well

or trains some other machine learning

technology that works well on something

go do it stop just telling everybody

they're gonna be wrong and go go show us

how to do it if you think you know so

well I don't understand I don't

understand so anyway uh it ended it

started nice and it got

[Laughter]

um but I I actually have um probably a

different point of view on that um I see

like sometimes it's that to your point

there is some extreme in in in how

shooting I feel like in trying to become

objective we're becoming extreme in

shooting down some other technology uh

but to be clear also it's like if you go

to Home Depot and you buy a hammer and

you say trust me I can solve every

problem in the world with that hammer

nobody's going to believe you and I

think the AI Community is doing the same

we're picking machine learning and

neural networks and we're saying every

single problem Under the Sun is going to

be solved with this and I feel like this

is also a failing approach folks like

Yan Lun again I'll keep coming back to

that is running an AI department at

Facebook and they are shipping models

they have shipped llama and they're also

talking about the limitations of of the

AI Technologies in certain areas

especially around reasoning um I've

personally attended

um uh presentation by a big AI company

that I'm also not going to mention their

name one of the most famous AI companies

that just were in the media because

their CEO left and they brought the CEO

back but I'm not going to mention their

name and they went through like a 20

minutes presentation around applications

of this technology in healthcare and

then someone comes in and asks but what

happens when hallucination goes on uh

does it ever hallucinate and the person

says yes it it does and says okay how

can I tell he said we actually don't

have a way yet but we warn our users to

not take the results as is but the

reality of the matter is he spent a 20

minutes presentation not talking about

it allog together he only brought it up

when he was asked so it is not that they

are telling their users that there is a

lot of downsides and there is a lot of

limitations for today's technology I

would say the second thing is once you

come into the area of machine learning

and neural networks essentially the

training data dictates a lot like

technology aside training data

dictates a lot and the fact that the

company owns what data the model is

going to be trained on and don't have to

always disclose what is this data can

create insane amount of biases that

literally boils down to the CEO and the

executive team deciding should we should

we do this more or not so in my in my

use case I can train the data on a

certain drug like I can basically show

the model patients responding really

well to lung cancer if they take a drug

that some Pharma company paid me more

money for and I can bias my training you

know like I can bias the training data

easy to do something like that and I can

always say sorry it's it's the model or

the model hallucinated um but there is a

lot of ethical considerations on what

data the models are using there's a lot

of ethical considerations on how we

present it that being said I also think

it's a beautiful piece of technology and

don't get me wrong 80% of our 70% of our

stack is boils down to machine learning

models it's a beautiful piece of work

but on it it on its own is again is like

a hammer that's going to solve every

every problem Under the Sun I know we

have two more minutes left but there was

an interesting question for you which is

what is it uh can and I'm excited about

digital twins so I'll I'll I'll give you

the last two minutes could you help us

understand in simple words what are the

biggest benefits of using digital twins

in clinical

trials oh um

sure I mean the big biggest benefit is

that you can run a clinical trial in

which fewer people have to be randomized

to

control so that allows you to run the

trial faster because you just you don't

if you have fewer patients to enroll

then your time to complete enrollment is

faster uh it's usually months faster um

it aligns better with the expectations

of patients so you get more people more

a larger fraction of people uh getting

access to the experimental treatment

which is usually why they participate at

all um and we can do this in a way in

which we make sure that the clinical

trial produces accurate and unbiased

estimates of how effective the treatment

is that's how we've been able to go in

and work work with the Regulatory

Agencies to to to get these methods

qualified for phase two and three

studies um how all of that works uh I

the simple terms in two minutes is maybe

a little to too hard uh but um uh but

we'd be happy if anyone wants to to know

more about about those things just reach

out to to us and uh and we'd be happy to

talk about it yeah I mean we have a

couple of minutes left so I don't know

if uh at least I'll offer that and I

know that uh Charles will do the same

but I would say like like some closing

notes Here is like if you look at

Charles's background it's celebrate or

die my background is Muhammad Ali

hitting down someone there is a lot of

uh people really trying to scratch the

surface here on what AI can do in

healthcare and we're we're essentially

not doing this only to build companies

and and make money I think there's both

of us we talked about this like uh both

of us come in from a lot of like like

there is no there is no motivation for

you to like go build Crazy Technologies

unless you truly believe in the outcome

of that so please reach out to us if you

have any questions around AI in general

if you're trying to to buy a AI

technology and you want some advice on

what considerations should should you

consider I think both of us would be

more than happy to like um be helpful as

much as we can um we hope those 60

Minutes were the start of the

conversation and not the end of of the

conversation and uh we wanted to bring

it to you as organic as it can be and as

casual as it can be also so thanks

Charles for for that and um for

endpoints also for helping us put this

together yes thank you thanks for

invting me fortunately that is all the

time that we do have for today I know it

was a great discussion lots of great

questions coming in so thank you

everybody for tuning in thank you Kareem

thank you Charles for sharing your time

and expertise and thank you to mendle

for sponsoring the discussion today and

points webinars if you'd like to rewatch

today's webinar or share it with

colleagues a link for on demand viewing

will be emailed to you tomorrow I'm

Carrie bball for endpoints news thanks

for joining and we hope to see you at

the next Take Care thank

you