JEFF PEDOWITZ
My guest today has done something most people only manage manage one of. He's founded five AI companies over thirty-five years with three exits, mostly in the brutally hard world of using AI to discover new medicines. He teaches the next generation of engineering managers how to actually use this technology, and he writes dystopian novels about AI agents and humanoids gone wrong, not for fun but as a way to stress test how these systems could fail before they do. So, we have a builder, a teacher, and a novelist sitting in one chair. We get into where AI medicine is real versus a bubble, why
JEFF PEDOWITZ
he says AI is no longer the story but the substrate and what imagining the dark future taught him about building the real one. Ed, welcome to Unscripted. Let's get into it.
ED ADDISON
Thanks Good to see you again.
JEFF PEDOWITZ
Good to see you, sir. You've been doing this longer than most of us even know what AI was about.
ED ADDISON
Well, I took my first AI course in 1985, believe it or not. Most people think AI didn't exist back then, but it did. And I've been in and out of AI ever since. Not much out of it, you know, I I I my ventures have all been at uh AI products, but really focused on specific products rather than the AI. The AI was the tool and it kept changing, of course, as you know.
JEFF PEDOWITZ
Of course. So, having been in it as long as you have, have there been some long-held beliefs that you had but that you recently changed?
ED ADDISON
Well, early on when I was at MIT in the 80s, I like most of the people there thought that the rule-based approach to both expert systems and natural language processing was going to win. Uh, and venture capitalists did too. They were investing in um expert system shells and things like that. But it quickly became a reality around 1987 or 1988 that this stuff just was too brittle to be used in uh for real users.
ED ADDISON
So, uh theoretical people can could develop their research around it, but um it just it's the coverage was just not what it needed to be.
ED ADDISON
uh so more emphasis on neural nets after that. And in the 90s was all about neural nets. But the neural nets in the 90s, they just weren't that accurate, you know. Maybe 98%, which sounds like it's pretty high, but it's really not when you're talking about speech or language. Um, and then when you got to big problems, the toy neural nets couldn't do it. They had to go to much deeper neural nets.
ED ADDISON
Somewhere around 2006, I think, uh deep learning came out and that changed everything because the computational power was there, uh and large data sets were available, you know, with the help of the internet. And so images, um speech, uh and uh you know, deep learning classification problems, all of them were were really hot back in 2006. I got into AI drug discovery in 2011, which was one of the early players, Cloud Pharmaceuticals.
ED ADDISON
cloud had a good run for 10 years. Uh when uh the pandemic hit, we had to break the company up into three pieces because of financial reasons. But uh uh some of that still exists. Um and uh there's some new initiatives going on there. Um but
ED ADDISON
in that course of that time you see 2014 generative AI came out first with generative adversarial networks. And then, uh, then 2017 you had the transformer architecture which only the insiders really knew what it really meant. It took till about 2022 to get a large language model out of it, but, uh, you know what we've learned in answer to your question, you know, natural language is best processed statistically.
ED ADDISON
And expert systems just don't work very well. So, um, generative AI is kind of saving the day on on that, you know, your expertise can be programmed, can be trained into a large language model.
JEFF PEDOWITZ
So, um, you've founded, I mean, I I have only founded one company still running over 20 years. You founded five. Uh, mostly in drug discovery. So...
ED ADDISON
One of them was in drug discovery and uh one of them was two of them were sort of surrounding technologies, bioinformatics and clinical trials enrollment. Uh but after after cloud, I became advisor to four or five companies in the drug AI drug discovery and still still doing that.
JEFF PEDOWITZ
Given your breadth and depth though and uh what what is it about drug discovery that's pulling you back in? What's the draw?
ED ADDISON
There's a lot of money being thrown at it. And, you know, you might ask a college graduate today, is it hard to get a job and they'll say yeah, but when you get to people with senior skills, um, it's the opposite. And I'm not looking for a job. I'm I'm doing a lot of consulting and uh teaching and and having fun, but I keep getting things thrown my way and when a good one comes along, I listen to it.
ED ADDISON
I sort of have the attitude that, you know, I'm having a good time, but gee, if one of these really good ones comes along, um, I just might uh ride that one. Because I I I... Is your wife not asking you to retire or slow down, kinda? Mmm, well, so, you know, when you reach come close to retirement age and haven't been in AI for 40 years and you see the field going like nuts, you don't want to miss out. So I'm not I'm not going anywhere for the next 10 years or so, you know, I'm I may shrink and swell the amount of work I do and vary it a little bit, but I I I I want to ride this out.
ED ADDISON
Um, I had enough years off over the years from selling companies and taking a break that it's just not that big of a deal to me right now. And I I tend to try to control my time. I have a lot of grandchildren, so I keep my weekends clear. And you know, we're all going to the beach next week, you know, putting five families in a beach house. So, you know, that's that's that's what's fun to me is family and travel. Uh but but work is uh um is not work to me. I'm I'm I'm just having a good time.
JEFF PEDOWITZ
Yeah, I mean, you have a lot of passions, right? You you've been a founder, an entrepreneur, you are a consultant, you're a professor, you just wrote a novel, you're a pilot, you swim. Um, how how do they all feed into each other?
ED ADDISON
You know?
ED ADDISON
Uh, I was just thinking about that the other day. Um, and I don't think, you know, our our method of measuring hours and saying that hours is proportional to work is just not true. Not in the knowledge work area. I've I've gotten great ideas while I'm swimming laps in a pool. And that when I wasn't working at all. Then there are other times when my I'm on mental overload and it serves no purpose to keep sitting at a desk and I just get up and go play my piano or something for a little while. But uh, and then that breaks that frees up my mind.
ED ADDISON
Um, I also find that the playing the piano in particular, especially I I like like jazz and I like improvisational piano. I don't want to just play sheet music that somebody else wrote exactly the way they wrote it. That's boring. But when you can improvise over it, it makes it fun. And it's just the kind of method, the abstract method of thinking that you have to go through. It's both left brain and right brained. Also applies to AI. You know, there's a left brain and right brain involved in AI. You know, everybody thinks it's, you know, logical, but there is there is a lot of art to it, I think.
JEFF PEDOWITZ
So, I mean, given your your focus, drug discovery, I mean, I don't think a lot of people realize how much effort and how hard it is, you know, to get the right drug because sometimes the best, most popular drugs often end up being accidental, right?
ED ADDISON
Well, let's take 2020. What what really happened over the last 10 years in AI drug discovery is that AI has replaced a lot of the lab components like high-throughput screening. But at the end of the day, you still have to put those molecules in a lab. You can make it more efficient with uh AI for the analysis or maybe the regulatory filing. Still have to go through a clinical trial, maybe you can make the management of a clinical trial more efficient with AI. But that speedup isn't worth as much as much as you think. Um, ninety percent of the cost is in the clinic.
ED ADDISON
Um, and so, you know, getting getting there better, faster, sooner is nice if you get to shave a little time off off or leave a little time on patent for collecting, you know, non-generic fees for for drugs. But, uh, the real value of AI drug discovery is yet to be proven, but it's it's it's widely believed and that's that AI will reduce the failure rate.
ED ADDISON
and because it comes up with novel molecules that you couldn't find the other way. And we won't be able to prove that until we get
ED ADDISON
several drugs out of the clinic and into approval. None have gone through approval yet. There are several in the clinic, both at uh Recursion and I think a couple at uh Silico Medicine. Some of the other ones claim they're entering phase one. Uh but these companies are valued by the drug and where they are in the clinic. They're not valued by their algorithms.
ED ADDISON
Algorithms are a dime a dozen if you ask me. You can always come up with a new algorithm. What's more valuable than an algorithm is the data. Big Pharma owns the data, so they they're in the catbird seat there.
ED ADDISON
beyond the data, what's really valuable is the drug. And the young companies are being valued by the number of drugs that are in the clinic and not by their AI algorithm. Although they get their first attention from the AI algorithm, maybe their first venture capital comes from it. Maybe it's supposed to be a flywheel, but uh again, the numbers are too small to prove anything yet. But the belief is still there and and what drives investment at insane valuations is FOMO, fear of missing out.
ED ADDISON
I would not put my money into companies like, say, Zara, X-A-I-R-A, that's that had a billion dollars of seed funding. The hurdle's too high to get out on that. The probability that they'll overcome that is low and the multiple that you can get is low on that. I would never put, um, angel money or personal money into a company that had an inflated valuation when it was just getting started. You know, angels like to put money in companies that are valued at five to ten million dollars, not that are valued at a billion dollars on day one. Right. FOMO's driving that.
ED ADDISON
I think that's going to have to adjust. And and today these companies are valued by by assets.
JEFF PEDOWITZ
Curious, I mean you you talk about the blur, the line blurring between an AI company and a biotech company. We do here in the States live in a highly regulated, you know, environment. But in other countries, you know, it's easier to get drugs to market. So I guess two sides of this. Are you, do you A, see more opportunity for AI startups to make, um, inroads at some of these, uh, companies that have less regulation, countries that have less regulation? And B, is there then a concern that maybe drugs can get to market too fast to become accessible if they haven't been...
ED ADDISON
I'm not sure that that practice is being widely pursued. And the reason that's a that's a practice that any biotech company could do. They don't have to be an AI drug discovery company to do that because it doesn't even come into consideration until you're in the clinic. And once you're in the clinic, most of the AI is already done except for the clinical management or the data analytics for clinical trials.
JEFF PEDOWITZ
Okay.
ED ADDISON
So, so I don't I don't see that as specific to AI by any means and and it's a strategy that some companies have pursued, maybe some successfully, but it's not a mainstream strategy. Venture capitalists don't really like it.
JEFF PEDOWITZ
Well, I guess just thinking of the dynamics, for example, with um the GLP products, right? That, you know, this last year have started to become more, I mean, while people have been taking them now for several years, they were expensive. If you couldn't get them prescribed, so people were...
ED ADDISON
Tell me about it. No, I tried it.
JEFF PEDOWITZ
Yeah, but it's right there going to Canada or going to other countries to get it because they couldn't get it here or it wasn't accessible. Now, you know, you've got like grow.com, these other services, and the demand is such that the price is coming down. So it's becoming more accessible. Just wondering in a
JEFF PEDOWITZ
capitalistic society, you know, are there, why wouldn't, I guess a company try and pursue more avenues besides just go straight through
ED ADDISON
for GLP
JEFF PEDOWITZ
or anything really.
ED ADDISON
the big pharma companies are in that market and and own that market. It's kind of tough for uh a small company to enter it, but unless they're they've got a really novel drug and they're going to flip it to another big company who can compete with the players that are in there.
JEFF PEDOWITZ
GLP was more of just as an example, not for a company specifically to get into the GLP market, right? My because you are an investor in these these uh drug discovery companies, um newer companies that are trying to use better techniques and bring drugs to market. I I guess I wanted to explore a little bit more both the ethical and and uh capitalistic grounds, you know, in terms of how a company could jump the shark so to speak and and get to market.
ED ADDISON
there are a couple of ways. I think precision medicine is wide open because big pharma is still pursuing the blockbuster and you know, frankly, these AI drug discovery companies are too, because they want high valuations and they need to support their overvalued stock price. You know, so uh, I don't know how long that's going to last, but uh, precision medicine is something that it's hard for big pharma to pursue because it means you have to have multiple drugs per indication and uh, and and then that if let's say you have five drugs for an indication instead of one, then the development cost is five x.
ED ADDISON
Now, where AI can potentially reduce that is using real-world evidence and stratify the patient world, and combinations of drug repurposing of existing drugs, and new drugs that apply only to subsets of the patient population, um, I think are are is where the market should go. But there's financial disincentives for that, and if AI can truly reduce the failure rate, then it can enable that. So, I think that is, it's not going to happen overnight because it's more than just AI.
ED ADDISON
And uh another more long-term area for AI in drug discovery is to take a step back and say, hey, the the prevailing paradigm in in drug development is targeted therapeutics. In other words, I identify a target, I design a molecule to attack that target, to inhibit it or excite it. Well, that's an on-off switch.
ED ADDISON
And if you have a systems mindset and and my background is in complex systems. Uh if you have a systems mindset, you can take a step back and look at that and say, hey, wait a minute,
ED ADDISON
Shouldn't drugs be modulated biology rather than on-off switch? Shouldn't shouldn't they be control functions? Like solving an optimal control problem. Well, the answer is they probably should be, but that's even more expensive to develop because there's more molecules that have to be part of that, you know, time-release capsule or whatever it is that you're going to use to time to time the control function. But I think AI can help solve that in combination with systems biology. That's a much longer term problem than precision medicine.
ED ADDISON
But it's one that would completely change the paradigm in biotech. That's probably ten fifteen years out because, um, we have to reduce the failure rate enough to be able to consider that.
ED ADDISON
that's that's that's a problem I've been eyeing that problem for years. And it's way too soon to enter. And I might be way too old when it's not too soon to enter, but I'd love to enter that market.
JEFF PEDOWITZ
I don't know, with your with your energy level and everything, you you could be doing this another 50 years. I don't know, you'd be surprised.
ED ADDISON
No I have several friends who are entrepreneurs in their mid-80s and still going.
JEFF PEDOWITZ
Love that.
ED ADDISON
and I knew one that was an entrepreneur until he died in his 90s. And so, I don't know, once an entrepreneur, always an entrepreneur. Sometimes our work changes form, you know, sometimes I'll go back to academic work, sometimes I'll do consulting work, but I've always got my ear to the ground for the next big opportunity, whether that's a startup company or to work for an existing company and help them launch a new market.
ED ADDISON
And I think AI drug discovery is fascinating. The stakes are high right now because because of the inflated valuations and the and the, you know, it's overcapitalized in some respects.
JEFF PEDOWITZ
So do you think we're in a bit of a bubble then?
ED ADDISON
There might be a financial bubble, it's not a scientific bubble. There's a long way to go with the science, uh but you know, you saw what happened back in the dot-com world. There was a dot-com bust, but the internet never stopped. It was just a pullback.
ED ADDISON
valuations went down by 80% and then they slowly tracked back up. I think we could see some of that in AI drug discovery, but it might be more muted because it's a longer term thing.
JEFF PEDOWITZ
I think we'll see that more broadly in in the AI market, let's just say. I mean,
ED ADDISON
Yeah, but we're early yet, you know, and generative AI is so early, um that um I don't think that AI is a bubble. I think the financial uh evaluating evaluation of AI is a bubble.
JEFF PEDOWITZ
Yes.
ED ADDISON
and it's and it's one that's and what's going to happen is somewhere along the way and it might be different in different markets, there'll be pullbacks in valuation with public companies or even private valuations. uh private companies is a real problem.
JEFF PEDOWITZ
but yeah, I mean we're already seeing this even with the public companies where, especially SaaS companies, they're coming in, they're doing great, their sales levels are good, but they're getting crushed because they don't have some sense of an AI business model.
ED ADDISON
No, I mean, generative AI, uh, agentic AI is crushing SaaS, especially for small and mid-sized businesses.
JEFF PEDOWITZ
Yeah, there's no reason to buy a traditional subscription license when you can buy code or kind of come up with something at a fraction of the cost that meets your needs. So, uh, there's definitely some concerns there. Um, you know, when you argue about the winners uh not being the teams with the most models but the ones closest to biology and experimentation. So, why do you think some of the AI drug startups are getting that backwards?
ED ADDISON
Um, I'm not sure you could say they're they're getting it backwards. It's backwards right now. When it is when it is eventually proven that AI can design drugs right the first time or right early on and really reduce the uh the failure rate, um, it won't be backwards. But we aren't that close to that yet.
ED ADDISON
We have succeeded, I think, in replacing wet lab functions with AI and reducing some costs in finding novel drugs, but we have not gotten them out of the clinic and tested. We can't do pharmacovigilance on on the approved drugs because there aren't many. And when there are, we can do that and then when we get enough data points, say ten or fifteen drugs, we might be able to measure with some at least reasonable degree of accuracy, uh, the failure rate change due to AI.
ED ADDISON
but until I can measure that, that's the single most important variable is the failure rate. It's not the speed, it's not the lower cost, and it's not really the greater novelty, although the greater novelty is related to it. Uh, is the single most important factor is failure rate because that acts like a lever arm on the cost of drug. You know that.
ED ADDISON
2.6 billion dollar cost to develop a drug is calculated by total R&D spend divided by total approvals, which means all failures are in that number. So if you reduce the failure rate, you dramatically take the 2.6 billion down. If you look at a successful drug and look at the costs of just that drug getting to market, it's nowhere near 2.6 billion. The rest of the money is all the failures that you took trying other things along the way.
ED ADDISON
you know, and pharma does maybe, you know, ten times as many discovery projects as they take them to the clinic. And a lot of them fail and some of them are good ones that they put on the shelf because they can only afford to take so many to the clinic.
ED ADDISON
You know, so those numbers are also a factor. Uh, we got to get the cost of clinical trials.
JEFF PEDOWITZ
how do you, I mean, from being in the business, it's understandable, but then, you know, it's the consumer when when the prices are going through the roof, how do you translate that? I mean, it would be awesome if we could use AI to help lower the cost of prescriptions. I mean, you don't want to take
ED ADDISON
The way you do that is look at a different problem. You take all generic drugs and repurpose them and find pill packages of generics that apply to specific patient classes. And if you can do that, then there's cheap drugs out there for those people and maybe there's a a use patent for the person who finds it, but that's just a premium on a generic drug, so um there's probably not enough of that going on. And there has been a lot of looks at repurposing, but it it hasn't been from a polypharmacy point of view.
JEFF PEDOWITZ
Should it be government sponsored or should it be just a capitalistic entrepreneurial endeavor or maybe both as a partnership?
ED ADDISON
You know, government sponsorship would help accelerate it.
ED ADDISON
because the big farm is not that interested in repurposing because they don't make as much money. And the small companies are not capitalized well enough to do it. And they're gonna they're currently chasing getting, you know, the next blockbuster in the clinic so they can support that billion dollar valuation that they shouldn't have.
ED ADDISON
So it's tricky from an investment point of view. All these companies are over over overvalued now and you're you're placing a bet on how long they're going to remain overvalued before the pullback or before adjusting to a different multiple.
ED ADDISON
And that affects the early investor a lot. It doesn't affect the later investor. They they become the conservative investor who just gets their 10% a year.
JEFF PEDOWITZ
Right.
ED ADDISON
Okay.
JEFF PEDOWITZ
I mean there's but I think that's the whole nature of investing, right? There's, you know, the earlier you are, there's more inherent risk, but there's also a higher return.
ED ADDISON
So there is, but but you know, one risk is the risk of success of a company. But now we also have the risk of of valuation, not just the risk of success. You could have a very successful company whose exit is below what you paid for it.
ED ADDISON
And then you you don't make money, but maybe the founders do, depending on what kind of stock deal they had.
JEFF PEDOWITZ
Right.
ED ADDISON
whether they get wiped out by the preferreds or not.
JEFF PEDOWITZ
So let's uh spend a couple minutes on your book, uh Probability Doom.
ED ADDISON
Probability of doom Oh, yes. That's that was what I call the accidental novel. It was never planned.
JEFF PEDOWITZ
How did you accidentally write a novel?
ED ADDISON
So you remember about a year ago, uh, at a Coldplay concert there was a jumbotron incident where an AI executive got caught kissing his girlfriend on the screen and his his wife found out about it.
ED ADDISON
I went to Kings Dominion with my daughters and their kids last summer, and there's a lot of we have a lot of grandsons who play video games. And
ED ADDISON
We have a five-hour ride back home. My daughter's driving, I'm sitting in the front seat, the boys are in the back seat. My daughter says to me, "You know, you ought to write a little novel about that Jumbotron incident. That's pretty fascinating." And I said, I pushed back and I said, "Nope. Not that story, but here's the story that you ought to look at." Because I had this problem on my mind: what happens when corporations start deploying humanoids and agents all over the place, and still humans? You got a multi-agent company that's a hybrid.
ED ADDISON
and having been a complex systems person, I knew about the principle of emergence, which says that you could be doing everything right and emergent behavior that you couldn't predict happens.
ED ADDISON
Um, so I said, we could write a we we could sort of borrow from that, but do it in this context. And and she said, how do you mean? And I said, well, think of uh a fictitious company in Las Vegas, and we called it Universal Business Dynamics. And then we dubbed them as the largest reseller of humanoids to corporations worldwide in 2031.
ED ADDISON
And I said, what happens if their chief robotics engineer gets fired, gets disgruntled and decides to rewrite or or script in, um, script in prompts into the guardrail software supporting, I mean because a humanoid, a good humanoid is going to have be a multi-agent system. It's going to have to have, uh, generative AI and and guardrails inside of it to be able to do the flexible kind of things that it
ED ADDISON
So, so the story that we concocted in the car, just and this was still not a novel, it was just fun talk, was that
ED ADDISON
The head robotics engineer, PhD from MIT, had a girlfriend who had an affair with the CEO. When he found out about it, he decided he was pissed and he decided since he had programmed the CEO's humanoid to bring him a glass of wine every day at five o'clock that he's going to reprogram that. And he does. Instead, the humanoid comes in at five o'clock one day. Instead of giving him his wine, he flips the bird. And uh, and the CEO being one of these egotistical, you know, slender British guy types, um, think of Donald Sutherland in Disclosure back in 1995.
JEFF PEDOWITZ
Okay.
ED ADDISON
Um, he uh calls his fixer in, wants to find out who did this. And when he does, he fires him. So this guy not only loses his job, he loses his girlfriend. Goes out and gets drunk because he's feeling sorry for himself and while he's still drunk, he reprograms the robots, which which were disseminated through the internet to all the robots worldwide that they were supporting.
ED ADDISON
programs and and puts in uh prompts in there to have them go recognize guns, go find them, fetch them, and learn how to shoot them.
ED ADDISON
Well, this was what he did while he was drunk, not not having as much ill intention as it seems. The story goes that he was the first victim of a humanoid shooting. So now the person who wrote the virus is gone.
ED ADDISON
And all of a sudden humanoids worldwide are shooting people. How do you solve that problem? That's the novel.
ED ADDISON
That's the basis of the story and my grandsons just lit up and said, you know, because they play video games and they saw Terminator and all this stuff. So we we outlined a story while we were in this five-hour car ride. And I said, you know what? I'm I'm going to write this. I'm going to get AI to help me write this. Because we know the characters, we know the story, we invented the story, but I don't want to do the so we we made ChatGPT and Grammarly and Claude partners in this book. But we controlled it and in the course of five months it actually turned into a novel. And
ED ADDISON
At the end of the day, it wasn't meant to be a dystopian future given everybody doom gloom about AI. It was intended more to, because of my background, to inform insiders in AI what to look out for when writing guardrails. It's not just, um, you know, blocking an intruder's prompt, but it's what if an insider decides to reprogram it or what if a virus is put into these these things. Um, and even if not, what if emergent behavior becomes bad? You know, the unintended consequences type stuff.
ED ADDISON
So, there were examples of all that throughout this thing because I intentionally put them there. But it was still a woven novel and there was like three different scenes going on that were interwoven throughout until it concluded at the end when some uh uh motley gang of hackers figured out how to use uh use uh Skynet or something to to to reach all the robots and uh and teach them to to stop their bad behavior. Uh and that was the end.
ED ADDISON
But there was more to the end. you'd have to read it to to see the end because there's some details
JEFF PEDOWITZ
we can all go get it on Amazon, yeah.
ED ADDISON
Yeah, it's it's an inexpensive, you know, less than 20 bucks paperback on Amazon. I don't get much of it, but I get something. Uh, but it was not, I'm not, I'm not trying to make money off a book. I use it as a talking piece. And uh, and it's also a great, I I use it in the classroom too as I say, you know, why do you need guardrails and multi-agent systems? You know, read this story and you'll see why. And then all of a sudden, you know, the students light up and they think, I would have never thought of that. You know, so it's pretty cool. And it wasn't, we never intended this to be a real book, but it is. It's not a big seller yet, but it is selling.
JEFF PEDOWITZ
I I bet it will be.
ED ADDISON
I have two fellow members of ASF AI who want to help me turn it into an AI generated feature-length film. Um, I won't name them because I don't have their permission to, but uh, and I don't know if we're really going to do it, but uh, it it seems like taking an AI generated novel and turning it into an AI generated feature-length film of AI characters only, that might be an interesting first. I doubt it would be a big seller, but it might be very novel and interesting as a case study.
JEFF PEDOWITZ
Why not? Um, I'd love to get in on that project.
ED ADDISON
Okay, well, you know, with your expertise, maybe you could.
JEFF PEDOWITZ
I thought so. Uh, we are out of time for today, Ed, but as always, it's amazing. Uh, every single minute is just chock full of good stuff. It's good seeing you, my friend.
ED ADDISON
All right, well great, thanks for the opportunity and I I look forward to uh uh viewing unscripted, not just this one but your other episodes.
JEFF PEDOWITZ
Thank you, sir.
ED ADDISON
Thank
My guest today has done something most people only manage manage one of. He's founded five AI companies over thirty-five years with three exits, mostly in the brutally hard world of using AI to discover new medicines. He teaches the next generation of engineering managers how to actually use this technology, and he writes dystopian novels about AI agents and humanoids gone wrong, not for fun but as a way to stress test how these systems could fail before they do. So, we have a builder, a teacher, and a novelist sitting in one chair. We get into where AI medicine is real versus a bubble, why
JEFF PEDOWITZ
he says AI is no longer the story but the substrate and what imagining the dark future taught him about building the real one. Ed, welcome to Unscripted. Let's get into it.
ED ADDISON
Thanks Good to see you again.
JEFF PEDOWITZ
Good to see you, sir. You've been doing this longer than most of us even know what AI was about.
ED ADDISON
Well, I took my first AI course in 1985, believe it or not. Most people think AI didn't exist back then, but it did. And I've been in and out of AI ever since. Not much out of it, you know, I I I my ventures have all been at uh AI products, but really focused on specific products rather than the AI. The AI was the tool and it kept changing, of course, as you know.
JEFF PEDOWITZ
Of course. So, having been in it as long as you have, have there been some long-held beliefs that you had but that you recently changed?
ED ADDISON
Well, early on when I was at MIT in the 80s, I like most of the people there thought that the rule-based approach to both expert systems and natural language processing was going to win. Uh, and venture capitalists did too. They were investing in um expert system shells and things like that. But it quickly became a reality around 1987 or 1988 that this stuff just was too brittle to be used in uh for real users.
ED ADDISON
So, uh theoretical people can could develop their research around it, but um it just it's the coverage was just not what it needed to be.
ED ADDISON
uh so more emphasis on neural nets after that. And in the 90s was all about neural nets. But the neural nets in the 90s, they just weren't that accurate, you know. Maybe 98%, which sounds like it's pretty high, but it's really not when you're talking about speech or language. Um, and then when you got to big problems, the toy neural nets couldn't do it. They had to go to much deeper neural nets.
ED ADDISON
Somewhere around 2006, I think, uh deep learning came out and that changed everything because the computational power was there, uh and large data sets were available, you know, with the help of the internet. And so images, um speech, uh and uh you know, deep learning classification problems, all of them were were really hot back in 2006. I got into AI drug discovery in 2011, which was one of the early players, Cloud Pharmaceuticals.
ED ADDISON
cloud had a good run for 10 years. Uh when uh the pandemic hit, we had to break the company up into three pieces because of financial reasons. But uh uh some of that still exists. Um and uh there's some new initiatives going on there. Um but
ED ADDISON
in that course of that time you see 2014 generative AI came out first with generative adversarial networks. And then, uh, then 2017 you had the transformer architecture which only the insiders really knew what it really meant. It took till about 2022 to get a large language model out of it, but, uh, you know what we've learned in answer to your question, you know, natural language is best processed statistically.
ED ADDISON
And expert systems just don't work very well. So, um, generative AI is kind of saving the day on on that, you know, your expertise can be programmed, can be trained into a large language model.
JEFF PEDOWITZ
So, um, you've founded, I mean, I I have only founded one company still running over 20 years. You founded five. Uh, mostly in drug discovery. So...
ED ADDISON
One of them was in drug discovery and uh one of them was two of them were sort of surrounding technologies, bioinformatics and clinical trials enrollment. Uh but after after cloud, I became advisor to four or five companies in the drug AI drug discovery and still still doing that.
JEFF PEDOWITZ
Given your breadth and depth though and uh what what is it about drug discovery that's pulling you back in? What's the draw?
ED ADDISON
There's a lot of money being thrown at it. And, you know, you might ask a college graduate today, is it hard to get a job and they'll say yeah, but when you get to people with senior skills, um, it's the opposite. And I'm not looking for a job. I'm I'm doing a lot of consulting and uh teaching and and having fun, but I keep getting things thrown my way and when a good one comes along, I listen to it.
ED ADDISON
I sort of have the attitude that, you know, I'm having a good time, but gee, if one of these really good ones comes along, um, I just might uh ride that one. Because I I I... Is your wife not asking you to retire or slow down, kinda? Mmm, well, so, you know, when you reach come close to retirement age and haven't been in AI for 40 years and you see the field going like nuts, you don't want to miss out. So I'm not I'm not going anywhere for the next 10 years or so, you know, I'm I may shrink and swell the amount of work I do and vary it a little bit, but I I I I want to ride this out.
ED ADDISON
Um, I had enough years off over the years from selling companies and taking a break that it's just not that big of a deal to me right now. And I I tend to try to control my time. I have a lot of grandchildren, so I keep my weekends clear. And you know, we're all going to the beach next week, you know, putting five families in a beach house. So, you know, that's that's that's what's fun to me is family and travel. Uh but but work is uh um is not work to me. I'm I'm I'm just having a good time.
JEFF PEDOWITZ
Yeah, I mean, you have a lot of passions, right? You you've been a founder, an entrepreneur, you are a consultant, you're a professor, you just wrote a novel, you're a pilot, you swim. Um, how how do they all feed into each other?
ED ADDISON
You know?
ED ADDISON
Uh, I was just thinking about that the other day. Um, and I don't think, you know, our our method of measuring hours and saying that hours is proportional to work is just not true. Not in the knowledge work area. I've I've gotten great ideas while I'm swimming laps in a pool. And that when I wasn't working at all. Then there are other times when my I'm on mental overload and it serves no purpose to keep sitting at a desk and I just get up and go play my piano or something for a little while. But uh, and then that breaks that frees up my mind.
ED ADDISON
Um, I also find that the playing the piano in particular, especially I I like like jazz and I like improvisational piano. I don't want to just play sheet music that somebody else wrote exactly the way they wrote it. That's boring. But when you can improvise over it, it makes it fun. And it's just the kind of method, the abstract method of thinking that you have to go through. It's both left brain and right brained. Also applies to AI. You know, there's a left brain and right brain involved in AI. You know, everybody thinks it's, you know, logical, but there is there is a lot of art to it, I think.
JEFF PEDOWITZ
So, I mean, given your your focus, drug discovery, I mean, I don't think a lot of people realize how much effort and how hard it is, you know, to get the right drug because sometimes the best, most popular drugs often end up being accidental, right?
ED ADDISON
Well, let's take 2020. What what really happened over the last 10 years in AI drug discovery is that AI has replaced a lot of the lab components like high-throughput screening. But at the end of the day, you still have to put those molecules in a lab. You can make it more efficient with uh AI for the analysis or maybe the regulatory filing. Still have to go through a clinical trial, maybe you can make the management of a clinical trial more efficient with AI. But that speedup isn't worth as much as much as you think. Um, ninety percent of the cost is in the clinic.
ED ADDISON
Um, and so, you know, getting getting there better, faster, sooner is nice if you get to shave a little time off off or leave a little time on patent for collecting, you know, non-generic fees for for drugs. But, uh, the real value of AI drug discovery is yet to be proven, but it's it's it's widely believed and that's that AI will reduce the failure rate.
ED ADDISON
and because it comes up with novel molecules that you couldn't find the other way. And we won't be able to prove that until we get
ED ADDISON
several drugs out of the clinic and into approval. None have gone through approval yet. There are several in the clinic, both at uh Recursion and I think a couple at uh Silico Medicine. Some of the other ones claim they're entering phase one. Uh but these companies are valued by the drug and where they are in the clinic. They're not valued by their algorithms.
ED ADDISON
Algorithms are a dime a dozen if you ask me. You can always come up with a new algorithm. What's more valuable than an algorithm is the data. Big Pharma owns the data, so they they're in the catbird seat there.
ED ADDISON
beyond the data, what's really valuable is the drug. And the young companies are being valued by the number of drugs that are in the clinic and not by their AI algorithm. Although they get their first attention from the AI algorithm, maybe their first venture capital comes from it. Maybe it's supposed to be a flywheel, but uh again, the numbers are too small to prove anything yet. But the belief is still there and and what drives investment at insane valuations is FOMO, fear of missing out.
ED ADDISON
I would not put my money into companies like, say, Zara, X-A-I-R-A, that's that had a billion dollars of seed funding. The hurdle's too high to get out on that. The probability that they'll overcome that is low and the multiple that you can get is low on that. I would never put, um, angel money or personal money into a company that had an inflated valuation when it was just getting started. You know, angels like to put money in companies that are valued at five to ten million dollars, not that are valued at a billion dollars on day one. Right. FOMO's driving that.
ED ADDISON
I think that's going to have to adjust. And and today these companies are valued by by assets.
JEFF PEDOWITZ
Curious, I mean you you talk about the blur, the line blurring between an AI company and a biotech company. We do here in the States live in a highly regulated, you know, environment. But in other countries, you know, it's easier to get drugs to market. So I guess two sides of this. Are you, do you A, see more opportunity for AI startups to make, um, inroads at some of these, uh, companies that have less regulation, countries that have less regulation? And B, is there then a concern that maybe drugs can get to market too fast to become accessible if they haven't been...
ED ADDISON
I'm not sure that that practice is being widely pursued. And the reason that's a that's a practice that any biotech company could do. They don't have to be an AI drug discovery company to do that because it doesn't even come into consideration until you're in the clinic. And once you're in the clinic, most of the AI is already done except for the clinical management or the data analytics for clinical trials.
JEFF PEDOWITZ
Okay.
ED ADDISON
So, so I don't I don't see that as specific to AI by any means and and it's a strategy that some companies have pursued, maybe some successfully, but it's not a mainstream strategy. Venture capitalists don't really like it.
JEFF PEDOWITZ
Well, I guess just thinking of the dynamics, for example, with um the GLP products, right? That, you know, this last year have started to become more, I mean, while people have been taking them now for several years, they were expensive. If you couldn't get them prescribed, so people were...
ED ADDISON
Tell me about it. No, I tried it.
JEFF PEDOWITZ
Yeah, but it's right there going to Canada or going to other countries to get it because they couldn't get it here or it wasn't accessible. Now, you know, you've got like grow.com, these other services, and the demand is such that the price is coming down. So it's becoming more accessible. Just wondering in a
JEFF PEDOWITZ
capitalistic society, you know, are there, why wouldn't, I guess a company try and pursue more avenues besides just go straight through
ED ADDISON
for GLP
JEFF PEDOWITZ
or anything really.
ED ADDISON
the big pharma companies are in that market and and own that market. It's kind of tough for uh a small company to enter it, but unless they're they've got a really novel drug and they're going to flip it to another big company who can compete with the players that are in there.
JEFF PEDOWITZ
GLP was more of just as an example, not for a company specifically to get into the GLP market, right? My because you are an investor in these these uh drug discovery companies, um newer companies that are trying to use better techniques and bring drugs to market. I I guess I wanted to explore a little bit more both the ethical and and uh capitalistic grounds, you know, in terms of how a company could jump the shark so to speak and and get to market.
ED ADDISON
there are a couple of ways. I think precision medicine is wide open because big pharma is still pursuing the blockbuster and you know, frankly, these AI drug discovery companies are too, because they want high valuations and they need to support their overvalued stock price. You know, so uh, I don't know how long that's going to last, but uh, precision medicine is something that it's hard for big pharma to pursue because it means you have to have multiple drugs per indication and uh, and and then that if let's say you have five drugs for an indication instead of one, then the development cost is five x.
ED ADDISON
Now, where AI can potentially reduce that is using real-world evidence and stratify the patient world, and combinations of drug repurposing of existing drugs, and new drugs that apply only to subsets of the patient population, um, I think are are is where the market should go. But there's financial disincentives for that, and if AI can truly reduce the failure rate, then it can enable that. So, I think that is, it's not going to happen overnight because it's more than just AI.
ED ADDISON
And uh another more long-term area for AI in drug discovery is to take a step back and say, hey, the the prevailing paradigm in in drug development is targeted therapeutics. In other words, I identify a target, I design a molecule to attack that target, to inhibit it or excite it. Well, that's an on-off switch.
ED ADDISON
And if you have a systems mindset and and my background is in complex systems. Uh if you have a systems mindset, you can take a step back and look at that and say, hey, wait a minute,
ED ADDISON
Shouldn't drugs be modulated biology rather than on-off switch? Shouldn't shouldn't they be control functions? Like solving an optimal control problem. Well, the answer is they probably should be, but that's even more expensive to develop because there's more molecules that have to be part of that, you know, time-release capsule or whatever it is that you're going to use to time to time the control function. But I think AI can help solve that in combination with systems biology. That's a much longer term problem than precision medicine.
ED ADDISON
But it's one that would completely change the paradigm in biotech. That's probably ten fifteen years out because, um, we have to reduce the failure rate enough to be able to consider that.
ED ADDISON
that's that's that's a problem I've been eyeing that problem for years. And it's way too soon to enter. And I might be way too old when it's not too soon to enter, but I'd love to enter that market.
JEFF PEDOWITZ
I don't know, with your with your energy level and everything, you you could be doing this another 50 years. I don't know, you'd be surprised.
ED ADDISON
No I have several friends who are entrepreneurs in their mid-80s and still going.
JEFF PEDOWITZ
Love that.
ED ADDISON
and I knew one that was an entrepreneur until he died in his 90s. And so, I don't know, once an entrepreneur, always an entrepreneur. Sometimes our work changes form, you know, sometimes I'll go back to academic work, sometimes I'll do consulting work, but I've always got my ear to the ground for the next big opportunity, whether that's a startup company or to work for an existing company and help them launch a new market.
ED ADDISON
And I think AI drug discovery is fascinating. The stakes are high right now because because of the inflated valuations and the and the, you know, it's overcapitalized in some respects.
JEFF PEDOWITZ
So do you think we're in a bit of a bubble then?
ED ADDISON
There might be a financial bubble, it's not a scientific bubble. There's a long way to go with the science, uh but you know, you saw what happened back in the dot-com world. There was a dot-com bust, but the internet never stopped. It was just a pullback.
ED ADDISON
valuations went down by 80% and then they slowly tracked back up. I think we could see some of that in AI drug discovery, but it might be more muted because it's a longer term thing.
JEFF PEDOWITZ
I think we'll see that more broadly in in the AI market, let's just say. I mean,
ED ADDISON
Yeah, but we're early yet, you know, and generative AI is so early, um that um I don't think that AI is a bubble. I think the financial uh evaluating evaluation of AI is a bubble.
JEFF PEDOWITZ
Yes.
ED ADDISON
and it's and it's one that's and what's going to happen is somewhere along the way and it might be different in different markets, there'll be pullbacks in valuation with public companies or even private valuations. uh private companies is a real problem.
JEFF PEDOWITZ
but yeah, I mean we're already seeing this even with the public companies where, especially SaaS companies, they're coming in, they're doing great, their sales levels are good, but they're getting crushed because they don't have some sense of an AI business model.
ED ADDISON
No, I mean, generative AI, uh, agentic AI is crushing SaaS, especially for small and mid-sized businesses.
JEFF PEDOWITZ
Yeah, there's no reason to buy a traditional subscription license when you can buy code or kind of come up with something at a fraction of the cost that meets your needs. So, uh, there's definitely some concerns there. Um, you know, when you argue about the winners uh not being the teams with the most models but the ones closest to biology and experimentation. So, why do you think some of the AI drug startups are getting that backwards?
ED ADDISON
Um, I'm not sure you could say they're they're getting it backwards. It's backwards right now. When it is when it is eventually proven that AI can design drugs right the first time or right early on and really reduce the uh the failure rate, um, it won't be backwards. But we aren't that close to that yet.
ED ADDISON
We have succeeded, I think, in replacing wet lab functions with AI and reducing some costs in finding novel drugs, but we have not gotten them out of the clinic and tested. We can't do pharmacovigilance on on the approved drugs because there aren't many. And when there are, we can do that and then when we get enough data points, say ten or fifteen drugs, we might be able to measure with some at least reasonable degree of accuracy, uh, the failure rate change due to AI.
ED ADDISON
but until I can measure that, that's the single most important variable is the failure rate. It's not the speed, it's not the lower cost, and it's not really the greater novelty, although the greater novelty is related to it. Uh, is the single most important factor is failure rate because that acts like a lever arm on the cost of drug. You know that.
ED ADDISON
2.6 billion dollar cost to develop a drug is calculated by total R&D spend divided by total approvals, which means all failures are in that number. So if you reduce the failure rate, you dramatically take the 2.6 billion down. If you look at a successful drug and look at the costs of just that drug getting to market, it's nowhere near 2.6 billion. The rest of the money is all the failures that you took trying other things along the way.
ED ADDISON
you know, and pharma does maybe, you know, ten times as many discovery projects as they take them to the clinic. And a lot of them fail and some of them are good ones that they put on the shelf because they can only afford to take so many to the clinic.
ED ADDISON
You know, so those numbers are also a factor. Uh, we got to get the cost of clinical trials.
JEFF PEDOWITZ
how do you, I mean, from being in the business, it's understandable, but then, you know, it's the consumer when when the prices are going through the roof, how do you translate that? I mean, it would be awesome if we could use AI to help lower the cost of prescriptions. I mean, you don't want to take
ED ADDISON
The way you do that is look at a different problem. You take all generic drugs and repurpose them and find pill packages of generics that apply to specific patient classes. And if you can do that, then there's cheap drugs out there for those people and maybe there's a a use patent for the person who finds it, but that's just a premium on a generic drug, so um there's probably not enough of that going on. And there has been a lot of looks at repurposing, but it it hasn't been from a polypharmacy point of view.
JEFF PEDOWITZ
Should it be government sponsored or should it be just a capitalistic entrepreneurial endeavor or maybe both as a partnership?
ED ADDISON
You know, government sponsorship would help accelerate it.
ED ADDISON
because the big farm is not that interested in repurposing because they don't make as much money. And the small companies are not capitalized well enough to do it. And they're gonna they're currently chasing getting, you know, the next blockbuster in the clinic so they can support that billion dollar valuation that they shouldn't have.
ED ADDISON
So it's tricky from an investment point of view. All these companies are over over overvalued now and you're you're placing a bet on how long they're going to remain overvalued before the pullback or before adjusting to a different multiple.
ED ADDISON
And that affects the early investor a lot. It doesn't affect the later investor. They they become the conservative investor who just gets their 10% a year.
JEFF PEDOWITZ
Right.
ED ADDISON
Okay.
JEFF PEDOWITZ
I mean there's but I think that's the whole nature of investing, right? There's, you know, the earlier you are, there's more inherent risk, but there's also a higher return.
ED ADDISON
So there is, but but you know, one risk is the risk of success of a company. But now we also have the risk of of valuation, not just the risk of success. You could have a very successful company whose exit is below what you paid for it.
ED ADDISON
And then you you don't make money, but maybe the founders do, depending on what kind of stock deal they had.
JEFF PEDOWITZ
Right.
ED ADDISON
whether they get wiped out by the preferreds or not.
JEFF PEDOWITZ
So let's uh spend a couple minutes on your book, uh Probability Doom.
ED ADDISON
Probability of doom Oh, yes. That's that was what I call the accidental novel. It was never planned.
JEFF PEDOWITZ
How did you accidentally write a novel?
ED ADDISON
So you remember about a year ago, uh, at a Coldplay concert there was a jumbotron incident where an AI executive got caught kissing his girlfriend on the screen and his his wife found out about it.
ED ADDISON
I went to Kings Dominion with my daughters and their kids last summer, and there's a lot of we have a lot of grandsons who play video games. And
ED ADDISON
We have a five-hour ride back home. My daughter's driving, I'm sitting in the front seat, the boys are in the back seat. My daughter says to me, "You know, you ought to write a little novel about that Jumbotron incident. That's pretty fascinating." And I said, I pushed back and I said, "Nope. Not that story, but here's the story that you ought to look at." Because I had this problem on my mind: what happens when corporations start deploying humanoids and agents all over the place, and still humans? You got a multi-agent company that's a hybrid.
ED ADDISON
and having been a complex systems person, I knew about the principle of emergence, which says that you could be doing everything right and emergent behavior that you couldn't predict happens.
ED ADDISON
Um, so I said, we could write a we we could sort of borrow from that, but do it in this context. And and she said, how do you mean? And I said, well, think of uh a fictitious company in Las Vegas, and we called it Universal Business Dynamics. And then we dubbed them as the largest reseller of humanoids to corporations worldwide in 2031.
ED ADDISON
And I said, what happens if their chief robotics engineer gets fired, gets disgruntled and decides to rewrite or or script in, um, script in prompts into the guardrail software supporting, I mean because a humanoid, a good humanoid is going to have be a multi-agent system. It's going to have to have, uh, generative AI and and guardrails inside of it to be able to do the flexible kind of things that it
ED ADDISON
So, so the story that we concocted in the car, just and this was still not a novel, it was just fun talk, was that
ED ADDISON
The head robotics engineer, PhD from MIT, had a girlfriend who had an affair with the CEO. When he found out about it, he decided he was pissed and he decided since he had programmed the CEO's humanoid to bring him a glass of wine every day at five o'clock that he's going to reprogram that. And he does. Instead, the humanoid comes in at five o'clock one day. Instead of giving him his wine, he flips the bird. And uh, and the CEO being one of these egotistical, you know, slender British guy types, um, think of Donald Sutherland in Disclosure back in 1995.
JEFF PEDOWITZ
Okay.
ED ADDISON
Um, he uh calls his fixer in, wants to find out who did this. And when he does, he fires him. So this guy not only loses his job, he loses his girlfriend. Goes out and gets drunk because he's feeling sorry for himself and while he's still drunk, he reprograms the robots, which which were disseminated through the internet to all the robots worldwide that they were supporting.
ED ADDISON
programs and and puts in uh prompts in there to have them go recognize guns, go find them, fetch them, and learn how to shoot them.
ED ADDISON
Well, this was what he did while he was drunk, not not having as much ill intention as it seems. The story goes that he was the first victim of a humanoid shooting. So now the person who wrote the virus is gone.
ED ADDISON
And all of a sudden humanoids worldwide are shooting people. How do you solve that problem? That's the novel.
ED ADDISON
That's the basis of the story and my grandsons just lit up and said, you know, because they play video games and they saw Terminator and all this stuff. So we we outlined a story while we were in this five-hour car ride. And I said, you know what? I'm I'm going to write this. I'm going to get AI to help me write this. Because we know the characters, we know the story, we invented the story, but I don't want to do the so we we made ChatGPT and Grammarly and Claude partners in this book. But we controlled it and in the course of five months it actually turned into a novel. And
ED ADDISON
At the end of the day, it wasn't meant to be a dystopian future given everybody doom gloom about AI. It was intended more to, because of my background, to inform insiders in AI what to look out for when writing guardrails. It's not just, um, you know, blocking an intruder's prompt, but it's what if an insider decides to reprogram it or what if a virus is put into these these things. Um, and even if not, what if emergent behavior becomes bad? You know, the unintended consequences type stuff.
ED ADDISON
So, there were examples of all that throughout this thing because I intentionally put them there. But it was still a woven novel and there was like three different scenes going on that were interwoven throughout until it concluded at the end when some uh uh motley gang of hackers figured out how to use uh use uh Skynet or something to to to reach all the robots and uh and teach them to to stop their bad behavior. Uh and that was the end.
ED ADDISON
But there was more to the end. you'd have to read it to to see the end because there's some details
JEFF PEDOWITZ
we can all go get it on Amazon, yeah.
ED ADDISON
Yeah, it's it's an inexpensive, you know, less than 20 bucks paperback on Amazon. I don't get much of it, but I get something. Uh, but it was not, I'm not, I'm not trying to make money off a book. I use it as a talking piece. And uh, and it's also a great, I I use it in the classroom too as I say, you know, why do you need guardrails and multi-agent systems? You know, read this story and you'll see why. And then all of a sudden, you know, the students light up and they think, I would have never thought of that. You know, so it's pretty cool. And it wasn't, we never intended this to be a real book, but it is. It's not a big seller yet, but it is selling.
JEFF PEDOWITZ
I I bet it will be.
ED ADDISON
I have two fellow members of ASF AI who want to help me turn it into an AI generated feature-length film. Um, I won't name them because I don't have their permission to, but uh, and I don't know if we're really going to do it, but uh, it it seems like taking an AI generated novel and turning it into an AI generated feature-length film of AI characters only, that might be an interesting first. I doubt it would be a big seller, but it might be very novel and interesting as a case study.
JEFF PEDOWITZ
Why not? Um, I'd love to get in on that project.
ED ADDISON
Okay, well, you know, with your expertise, maybe you could.
JEFF PEDOWITZ
I thought so. Uh, we are out of time for today, Ed, but as always, it's amazing. Uh, every single minute is just chock full of good stuff. It's good seeing you, my friend.
ED ADDISON
All right, well great, thanks for the opportunity and I I look forward to uh uh viewing unscripted, not just this one but your other episodes.
JEFF PEDOWITZ
Thank you, sir.
ED ADDISON
Thank
