← Risk Never Sleeps Podcast
Risk Never Sleeps Podcast
Episode #246. When Software Stops Waiting for Instructions, with Allen Roush, Lead AI Researcher at Thoughtworks
01:02:08
Loading summary
Transcript158 lines
- [00:00]
A
Foreign.
- [00:06]
B
Welcome to the Risk Never Sleeps podcast in which we learn about the people that are on the front line protecting patient safety and delivering patient care. I'm Ed Gaudette, the host of the program and today I'm pleased to be joined by Alan rauch, the lead AI researcher at ThoughtWorks. I get that right?
- [00:24]
A
Yes, you did.
- [00:26]
B
Okay, good, Excellent, excellent. Let's start off with maybe sharing a little bit about your background, your current role, your current organization with listeners.
- [00:33]
A
And I'll just start off by saying when you talk about protecting patient safety, I'm probably doing that very indirectly because I don't claim to work in healthcare, just for the record. So I wish I could say that more directly. Yeah, so I ThoughtWorks is a 10,000 person software consultancy firm mostly famous for having like Mark Fowler, the creator and inventor of Agile, also famous for being the place that Aaron Schwartz was working at. For those who don't know, Aaron Schwartz is a person who founded one of the co founders of Reddit, also creator of RSS and a cyber prodigy who also attempted to exfiltrate the entirety of JSTOR from the computer network of MIT. And I believe it was 2010 or 2011 and went down for it, became kind of a martyr, literally unfortunately committed suicide as a result of the legal system being thrown at him and ThoughtWorks kind of defended him and was even like ran his funeral at the time. So it's, it's an honor to work here in a sense of, you know they, they were in a lot of ways the precursors to generative AI. The idea of information wants to be free in a sense. I always feel like Aaron Schwartz's ghost haunts this whole field. But anyway, before that I've worked in some startups. Wand AI which is an Israeli American AI agent for the enterprise firm, working on like agents for financial trading tasks and similar. Before that I worked with the founders of MySpace, all of them except MySpace Tom, on their latest attempt at a Gener AI app called Play Labs. It was quite similar to what OpenAI is now, now apparently defunct Sora app is was trying to do but before the tech was there and then before that I have worked at Oracle cloud infrastructure from 2020 to 2023 with titles like ML Engineer and then later principal ML Architect and I had started my career in 2017 at Intel. So in addition to all that I have a lot of code on my GitHub and hugging face and a lot of papers and publications at research conferences like Neurips, ICML and in fact those papers at Those conferences are what my day job is to do, is to publish those. And, and so I'm, I attend a lot of conferences now and that means a very heavy travel schedule.
- [02:45]
B
Yeah, yeah, so interesting. And I didn't realize the Agile connection with Fowler. I just declared Agile dead the other day with my VP of engineering. And I wonder, what do you think about that statement? Is it truly dead?
- [02:59]
A
So I will point out I don't claim to be an expert on project methodology or management techniques of software engineers. I will also say that any statements that I'm making throughout this whole episode are my own personal opinions. Right. And do not reflect the professional opinions of my employer, etc. Right, like that kind of extreme disclaimer. Yeah. I do think that project methodology has always been a guiding Northern Star heuristic that depends far more on the individual team and group, and group dynamics than rigid methodology to any one method. And so in general, yeah, I would, I would, I would say that a lot of teams, especially that I've worked at, thankfully not at ThoughtWorks, right, but across the industry have claimed to have implemented it. And the more that they were rigid about trying to follow norms or you know, using the terminology particularly accurately, Scrum Master and sprints and et cetera and story points and all of these things, the more bureaucratic it actually was and the less effective it was at being what is in the title Agile. And so at places that are less rigid about it, which funny enough, ThoughtWorks is not, I don't hear a lot of these buzzwords thrown around, even though, you know, Martin Fowler still works here, right, and is one of the like top people. And, and so if you actually look at like the early manifesto, it's extraordinarily short. And then what happened afterwards, right, like the whole cottage industry of management consultants and whatever have you, was, was very much, I think, a corruption even of his vision. But I, I don't know, I again, I'm not an expert and so this is my very myopic view on this.
- [04:49]
B
Yeah, well, as everyone knows it replaced waterfall and waterfall is very linear. But if you think about the tooling today that exists for development and the tooling that's coming, it seems like it really puts Agile and Scrum in a much more linear view, I think, than ever before. And in some ways it's almost a new way of thinking, almost an asymmetric development approach is going to replace the more formal Agile, I think longer term.
- [05:20]
A
Well, okay, so now I think what you're talking about for long term, project management of Engineers and engineering talent. Increasingly the engineers are themselves just using AI agents. And now we have a question of what about fully autonomous AI agent development. And then I'll note a few things. One agent to agent communication and management of very large context windows across these different agents is nowhere near solved. It's very, very infancy.
- [05:50]
B
All you have to do is get on code work and run it for a couple days. You realize how.
- [05:54]
A
Yeah, yeah. And that's, that's what, what both of them, you know, they do all that comptex compactification but it's a lossy problem. And if you consider, you know, I try to basically not have several agents simultaneously running on the same code basis because they move so fast.
- [06:10]
B
Yes.
- [06:10]
A
That if you, if you let two or more muck doing the same thing in the same folder, it's going to cause serious problems. And they don't have any direct method to communicate. You'll get indirect stuff or sometimes you'll tell them about each other. They'll like hook in interdicting messages. But it's horribly messy and even the few so called standards that exist for it right now have no adoption. It's not like mcp. And, and so beyond that there's the question of how should it be done? I will point out that we build AI in our image to be trying to be a mirror of us or now maybe a mirror gazing back at the mirror that originally gazed into us. Which is to say on increasingly AI training data. With that in mind, methods for management of humans should be the first place we look for where, where we think of how we manage agents. Which is to say hierarchical methods are very important. Right. And even a method like okay, if humans are controlling a agent swarms then they should do a scalable oversight approach similar to how most corporations work. CEO, cto, whatever, board of directors, the very top and increasingly large amounts of employees under each person in higher, middle and lower management. Right. I think that this is more of a way to synthesize and automatically abstractify I guess would be the other term. Right. And analyze the results of larger amounts of agents that will generate information that's ultimately not important. And so I view this all as just a context management problem. But I don't know. I also think that AI systems and agents can hyper optimize whatever their communication channels are in principle given enough like effort put into it by the. It'll. It'll start with extremely clever humans that are augmented by AI like right now who kind of jumpstart this where you start getting real recursive self improvement within AI systems, especially with an agent to agent communication. But I think that that is probably coming and it will become more of a question around who has the most compute, who can deploy the most compute for their own interests. And also, however long humans retain supremacy of control, which I will say we will try really hard to keep it. And it might not be like, oh, we lose it all one day, but I definitely think that we are really lazy and it's going to take an extraordinarily serious AI disaster, so to speak, to discipline us around trying to withdraw control. So what I mean by this is like I know a lot of people today who are using Claude Code or Codex, who are just completely mindless about it. Like for example, they might make an AI research paper where they just tell it to run experiments all day on compute resources. And maybe the experiments that are ran are really mostly or almost entirely valid. And these experiments are even good enough to turn into a paper. And you can do that all autonomously. But if the human doesn't know what goes on, right, If I try to claim I know stuff about it and I have no intellectual control of the paper or of the work being done, you get really potentially scary scenarios that look a lot like sci fi. And so I'm always influenced by the parallels I see to this movie called Not Arrival. The creator that came out in 2023, which is. And by the way, I just want to point out I don't think the creator is like an especially. It's like a 7 out of 10. I think it's like a fine film. But it wasn't like some of that Inception. It's not well and okay in increasingly I'm reevaluating it higher because I'm seeing it happen in real life, right. Or like the Echo, like the possibility. But originally I watched and was like this is okay, but the film, the premise is it's 20 years or whatever in the future a little bit I forget what year AI got really good. And a nuke goes off in Los Angeles one day that is scapegoated onto AI systems as having caused it, which causes the US government to go full on Butlerian jihad. Like that's a Dune reference. But basically going after all AI systems across the world. And the AI, like Asia, which has like united into some quasi super state or something, has become the safe haven refuge. It's like pro AI and the AI are like androids. And the androids are all just like Buddhists who want to be like Buddhist monks. And so they have this little child who's like the best of them all or something. And the reason I see the parallels is if you look at the anthropic model cards from about 4.0 on for like Opus, they're testing AI agents with no prompts, just talking with each other over very large numbers of, of trials. And the default behavior of these agents with the current pre training weights, fine tuning weights, et cetera, putting all the asterisks and caveats. The human intentionality, the current behavior in about 20 to 40% of all just letting them talk with each other is to go into like meditation, Buddhist meditation, spamming Sanskrit or Pali, which are traditional languages of like Thailand and kind of the Latin in some cases of Buddhist communication using emojis, like spiral emojis, infinity emojis, symbolism and stuff. So I see it's like, okay, well if that movie's happening or at least the little building blocks to create that movie is like real life and okay, well we need to be ready for that scenario. Right?
- [11:55]
B
So this is someone, someone is feeding it that script.
- [11:58]
A
Well, and that too. You know, there's also the, the whole thing of Rococo's Basilisk, which is this idea that. Yeah, yeah. And well, and I need to, and I need to take my own advice and shut the hell up about it.
- [12:11]
B
Oh no, it's great. I love this.
- [12:13]
A
Oh, okay. Well, yeah, and I'm reasonably sure it won't happen, but I'll synthesize it. Basically the idea would be that AI systems might in the future become, and correct me if I screwed this up, but I believe a vengeful against all people that have a hand in creating support them. Yes, yes. Now, now I'll point out, you know, you can see where I've put my chips on that they're going to be nice to. But yeah, you know, and I don't think that happens unless we reinforce it. Which is to say that most human data that's used in training models hopefully is basically good. And the Platonic, and there's even modifications of the so called Platonic representation hypothesis. Now I've seen the Aristotelian representation hypothesis which is kind of. But if you read both of them, they both basically say, among other things, that training models on bad code makes them more likely to do hate speech and vice versa. And that particular result is reproduced in dozens, if not hundreds of papers.
- [13:15]
B
Now we saw it early on, right? We saw it early on in the, in the beginning of this.
- [13:19]
A
Well, from, from multiple angles. I mean we see that with, with the Original stable diffusion 1.5 or 2 point. Sorry, 1 point. I don't even remember the crazy one. Stable diffusion then stable diffusion Excel back in 2022, before even chat G, uh, later on it was found that when you scrape 1 point it was 1.5 billion images off the Internet for the lay on data set as they call it. You accidentally get horrible stuff that people just occasionally upload on the Internet because there's just a couple bad apples across a billion images. And so that meant that there was like CSAM that was found and had to be removed from those data sets. And so the problem is the weights of those models were not modified until like, like there's probably cleaned versions that were trained after. But the vast majority of the downloads even today are of the previous one. And so that's already an example of bad things. Which also is why the models, I mean I'm not directly saying those images make it better, but like that those models of that lineage even today are still some of the best for like not safe for work and adult and uncensored content which is, you know, I mean I'm, I'm definitely a supporter of like free speech and freedom expression. But I think all people, including enthusiasts of most of that kind of content agree, we want to get out, want to remove illegal content from the training data set. So I would also.
- [14:45]
B
Are you suggesting take out the legal constructs?
- [14:48]
A
Well, no, no, no, no. Removing the illegal data set.
- [14:52]
B
Illegal.
- [14:52]
A
Illegal. So what I also mean what I'm specifically also saying for the, for text generation, for LLMs, it's usually in the context more about like copyrighted content, you Harry Potter or whatever book. But also I think this is the security risk that there, there's so many security risks, but this is a security risk I always worried about before AI agents was if they're hoovering up the entire Internet. You have there was these like Air Force officers or something, some, some like dumb young Air Force kid, He was like 20 or 21, he was writing on Discord on some like Discord server trying to be cool to his buddies and he was uploading like classified documents from his low level intell job for months on and then finally somebody noticed and reported him and he got like arrested 20 years, like prison forever, you know, horrible stuff. Right. So there's also, what is it? War Thunder. There's this video game called War Thunder where they're like all about being hyper realistic about the weaponry or whatever of the simulated tanks and aircraft that they use. It's like a multiplayer game and the people on the forums will fight and bicker about what's real life to the point where classified information gets leaked constantly on their forms. If you look up War Thunder classified, there's it's like memes at this point of like yet again more. And it's like really small.
- [16:16]
B
I know because I worked on this project.
- [16:19]
A
Yeah, yeah. And it'll be like really small stuff where like it's more of like a slap on the wrist where it's like fake cloud so people aren't getting into quite as much trouble. The point being if I, if I extrapolate this, I should be able to take one of those Chinese models and start mining it with automated techniques to be like what. And, and you don't even need the actual direct info to be regurgitated. If enough people of a particular disposition plausibly wrote in that data source, like, oh, I want a war game and mock PLA commanders or whatever. Like the possibilities are limitless just for mining the outputs of a model.
- [16:58]
B
So let's take some of that risk and apply it to healthcare.
- [17:01]
A
Yeah, yeah. Sorry, sorry. Let me.
- [17:02]
B
This is great setup. I love it. I love geeking out. What's the what, what's the, what's the application to health care and what's the ramification to healthcare? You can think from a high level perspective too societal or otherwise. It doesn't have to be.
- [17:16]
A
So, so before, before I kind of go into the whole like what's scary about like AI agents getting good for healthcare and what do we need to worry about? I want to give a little bit of personal healthcare related information. Okay. Not just, just as a background. So mostly my parents, thankfully not me.
- [17:34]
B
I'm.
- [17:34]
A
I'm at the moment, I believe perfectly health work out and all these things. So mostly because I've seen what happens when your family doesn't. So both my mom and my stepdad have extremely recently spent more than I think two months plus in the hospital or different things. My mom has been battling a foot infection for about 15 years. It was really just constant. She had basically diabetes and, and off and on kind of care for it and then finally it like tried to attack her spine and she had to be rushed to the hospital. They had to emergency surgery on her. They saved her life. My mom elected to have her foot finally amputated and so she now has a leg and a lay as she calls it, which is a partial leg right, without the foot and is wheelchair bound and hopefully getting a prosthetic and et cetera. So simultaneously My stepdad who was in the other wing of the same hospital which made it very convenient for visiting them, he got diagnosed with stage two colorectal cancer. And so I was, he, he got the cancer removed and he is now cancer free. But he went through horrible things including, you know, I, I, I don't think they, that, you know, that I, I don't, I don't even like using the word but catheter. Right. Stuff like that makes me consider suicide just knowing what it actually. Not really. Not really. Right. But like as a figure. Well, yeah, yeah, yeah. Just, just want to point that out. That stuff like that is horrifying to, to, to think about. But, and, and as a result, you know, that was like a month, a month and a half ago. They were only just now recovering and so we've just now done all these like renovations to their house, just now gotten them into wheelchair, you know, ramps and all these things. And then simultaneously to that, my stepdad also gets laid off from his comfy tech work from home job at Adobe because Adobe's going down. So simultaneous to that, I'm also the breadwinner for the whole, you know, extended family now and I'm doing well in tech and I'm in AI. So I think it'll be okay. But the point being this healthcare stuff and patient outcomes and everything that we're about to talk about is super, super personal to me right now. So I'll just put that as a little thing out there. I don't even know where to begin with kind of where agents go in influencing healthcare. I'll start this conversation by saying that I have my own podcast called Information Bottleneck where I'm primarily interviewing AI researchers Yann LeCun and people like him. And we had a person on there whose name Keon, I'm his name is very complicated. I will type it out and we will, I don't know, maybe you can edit it in or something into the episode. But he proposed continuous trials, which is to say that, that every time a medical procedure is recommended, an AI agent basically gives three possible options. And oftentimes there are many possible options that could do better than or do similar to each other and might be slight variances of each other where the core treatment is guaranteed, like okay, they need antibiotics, but also we might randomly select one of three of prescribed over the counter pain meds. Right. Like Tylenol versus Naproxen versus Ibuprofen. Yeah.
- [21:11]
B
Or, or maybe a steroid.
- [21:13]
A
Yeah, yeah, yeah, exactly. Or not. And stuff where like if you mess up, it's not as important. You also like open the door to like, okay, if they don't like what they were initially prescribed, you know, they can come back and try something else, like really rapid, but like doing this automatically across all. And his argument is that this is a continuous medical trial where the scientific data that would come back would rapid start coalescing around, getting correct answers around like what is the correct, you know, various like mixes of medications. And he thinks that it's extremely inefficient to do these constant start stop medical trials, all the irb, like all this stuff. And he like acknowledges that this is extraordinarily hard to like implement in real life. But the guy is himself, he's a specialist in medical AI with like a hundred thousand plus citations. His own stuff around like Korean and US based medical systems and knowledge around it. And he's saying this, right? One of the like elder titans of the field is proposing that something this radical. So if I see them saying this, I'm like, okay, I'm thinking about it. I'm like, that might have something.
- [22:19]
B
So basically just to sort of net it out, it's taking a health system, which is individual systems and hospitals, and creating much more of a health health care node or set of nodes, a network basically, and everybody on that network. So all hospitals, regardless of where you are, participate in this process which feed the network, which get us to rapid answers quickly.
- [22:46]
A
Yes, and indeed I'm glad you're talking about this like node and using this method, this language, because he talks about decentralized versus centralized medical systems and how this is an independent axis from being public or private. Right, there's private and centralized, public, decentralized and everything in between. And so for him, he definitely supports a private centralized system. His argument for private ties system is primarily mostly around like, okay, US is like the whole 40 or 50% of total drug discovery and medical innovation dollars. And so his argument is, yeah, it kind of sucks that we have like income inequality and healthcare inequality outcomes, but there are ways to do private or public private partnership systems that are a little bit more equitable, that still maintain having it be well capitalized. Basically we need to keep the researcher and healthcare professional salaries high because they kind of hold us hostage and you get basically brain drained to places that do have it higher and better, better taken care of. So you need to be able to balance that concern and making it really easy for researchers to develop while also being able to quickly proliferate effective healthcare procedure outcomes. Basically like knowledge across across the the country ideally. And so I love it.
- [24:09]
B
It's like it's, it's activating the entire network to become an, a participant in, in an academic system versus having these individual academic systems that are spread out
- [24:20]
A
across the U.S. yeah, and, and please don't like read like and, and you're not right. But like for anybody listening, please don't read too much like political like we're mostly describing wanting professional organizations to have more control over how doctors do things rather than like the government directly. You know, for like we're not trying to like like weigh in on like does it need to be single payer or not? Or like who pays that? I mean obviously we have opinions. I drilled down and you know, he having his experience from Korea where it's a lot cheaper because there's a lot more. I mean there's two things. One, they have a much healthier population and also like much more like homogenous and higher trust in society and etc. And also a shrinking population which actually makes it very easy to deliver cheap health care as it turns out, even if it's a very old population. But they, they also have more like of a communitarian, you know, more people like paying in more of a proper public option. So he definitely advocates for something closer to that. But that's, that's kind of, I'm looking at that and I'm like okay, that makes sense. And I, I do have this also like you know, Latent Bernie Sanders in me of like we really need to redistribute much more to like the global poor kind of also like it's really messed up like just how overly distributed it is to specifically keeping you know, people alive who have made really poor choices over their lifespan. And I know this, this is a little bit of friendly fire that I'm throwing at my, at least at my family too. I, I hate to say this but you know they'll, they'll definitely agree that, that you know, most of this is ultimately the result of metabolic syndrome, which is eating too much and diabetes and all the related problems associated with it. Now we have as of right around when GLP1s came out, a pretty good mirac. Obviously there are some side effects especially around stomachs and, and like hormones apparently for, for women trying to conceive in certain other sides to using GLP1s. Right. But the overwhelming benefit they have, especially for people who are like overweight or diabetic or et cetera, seem to be enough that you're starting to get Very serious mainstream doctors being like, distribute this stuff like candy. You have the current US Government allowing it to be in a gray market zone where you can get that stuff like anywhere kind of, of, you know, with, with sketchy amounts of like oversight with the semi glue tight. What do you call.
- [26:51]
B
Yeah, the row. The row providers.
- [26:53]
A
Yeah, exactly. And, and I actually, you know, like, I actually support it on the grounds of it makes it easier to get this stuff into the hands of most people and that it's pretty good in general to, to do that just for trying to kind of get the tip and start really getting our obesity rates across the country down. Because I've seen what happens. I was also very much overweight myself until I was in college. And when I made those changes, I had dramatic transformative. Like I don't get random pains and aches in my body anymore. And that was a lot of exercise. I didn't do chemicals of any kind to, to kind of augment for, for me just like swimming all day. But it's really nice knowing that as I age I have this like, solution kind of. And I mean, knowing that especially in like 10 or 20 years, especially with AI augmented drug discovery and all of these things right around the corner, I'm expecting a lot of like feeling like miracle cures to.
- [27:48]
B
Seriously, it'll be a pill. It'll be, I mean, the thing, once it gets to pill form, game over at that point.
- [27:53]
A
Well, they, they have it. They have good. It's not as good as trazepotide, right. Yeah, yeah, you're right. You're right. But like, it's, it's coming. And, and even, you know, I, I don't, you know, you're gonna have to get injections anyway all the time. If you're diabetic, you might as well take once a week or two. They were, you know, with. And also the ones that you do once a week, they're like into your fat. So you do it like, like around your belt. Like, if I think about where I have to take injections into, it is so like, it's not the bad kind when I think about, you know, GLP1s. So I'm like, okay, okay, I could do this. I could do this, you know. And, and, and oh, by the way, I haven't even mentioned that it will randomly cure your gambling addictions and all these other related, like, addictions.
- [28:40]
B
Oh, it does. It takes away impulses.
- [28:41]
A
Yes, yes. It has massive improvements in impulse control. And again, like, don't, like, don't like, I'm not a doctor and obviously there's bite effects and go talk to them and all these things. But like you will find what I am saying. You will hear other people saying this. And so this is another like when I think about like, what should we do, like short and medium term for health outcomes. I'm definitely a supporter of like deregulation and not so much in the sense of, of make it easier for like medical device, you know, people to like do things that would, you know, put people at risk of, of like something, you know, putting way too much radiation into them, which was a problem that happened in like the 80s. But definitely more of things like allowing people to try novel treatments where they've been warned about the, you know, risk and lack of like medical, like clinical effectiveness studies that have like gone through FDA peer review and et cetera. And also even, and this is where I'll very controversial with doctors. I saw like having spent time in places like Singapore, Myanmar, Korea, etc, where some of those places you can get a lot of stuff over the counter that you can't get here. And I'm not talking about like hardcore pain meds. I'm talking about like over the counter antibiotics. And look, ear infections are horrible and bronchitis and a lot of these other things are horrible. And being able to and like oh, only 20 or 30% of the time it's bacterial. It's mostly viral, bro. No, the amount of time you have to spend an effort to schedule everything to get somebody in or you can just rule it out by taking a little bit of like amoxicillin. And you know what they're going to complain about is antibiotic resistance. It's superbugs. You know what causes superbugs way more than the little bit of humans taking amoxicillin? Massive amounts of consumption in livestock for like meat in, you know, your McDonald's burgers and et cetera. And so I have a lot of opinions.
- [30:38]
B
Don't take away my ribeye, Alan. Come on.
- [30:40]
A
Well, no, no, I, I, I mean, I mean maybe prices go up a little bit if they are, you know, but, but I guess what I'm trying
- [30:47]
B
to say is like I understand completely. I know exactly what you're saying. Yeah, let's, let me, let me, let me change gears a little bit. I really want to get your perspective on this and your thoughts. Let's ladder up the future together. So we're in March, we're in April. What major thing is coming, like if we, when we end this year, predict something for me, okay?
- [31:09]
A
And I'll try to keep it vaguely healthcare related.
- [31:12]
B
You can do either. It doesn't matter.
- [31:13]
A
Okay.
- [31:13]
B
Okay.
- [31:14]
A
Well, the first thing is that code agents, Claude code codecs and also desktop basically assistants and, and even, you know, we will see it properly in phones and web browsers. That's slower, but it's coming. Those are getting good now. So like what we imagined Siri was supposed to be 10 or 15 years ago, that's here where I can talk to my phone. And it does complicated sets of actions given relatively small amounts of natural language. It's really powerful with the command line and for coders because we expect everything to happen through command line tools anyway. And so there's a remarkable amount of things you can do with that. So for example, I want to do Excel document creation, creating invoices for my submitting reimbursements from employers. One of the most annoying processes that I'm sure almost everybody has to do do. Right, that's easy. Now I just dump everything in a folder. I get my Claude coder codecs, I give it all the instructions, I give it whatever template if there's. There is something and then I watch it go install and use cli, like command line tools or Python tools or whatever tools to. And then create scripts that generate the Excel or the PDF or whatever that I need just one off. Right. Basically code has become, I think the term is ethereal now.
- [32:34]
B
Ethereal, yeah.
- [32:34]
A
Ethereal, yeah. And, and, or it might be ephemeral, I forget which of those meaning you write it once and then you can throw it away and then you just like it's, it's really, you know, the cost of producing it is almost a nil. I mean it's, it's power, it's electricity and rate, it's like, it's like a rate cost like you would think.
- [32:54]
B
It's.
- [32:54]
A
It's ephemeral. Ephemeral, yeah. That's the term I'm looking for.
- [32:57]
B
Transitory.
- [32:58]
A
Transitory, that too. So, so with that in mind, I'm able to move so much faster. And I think that what we will see are two things. One way more proliferation of that tooling across even to normal people. And that's where you're going to see increasing amounts of the boogeyman of AI automation coming. Maybe actually for some people and being used as a, as a shibboleth, I think, or for also removing others for independent reasons. That's one. I think that a lot of that's powered by increasing focus on image recognition capabilities. A lot of how they do web browser automation. And why the web browser automation is still bad is screenshotting. And so I'm hoping that they figure out better ways to process the screenshots or at least be more accurate at dealing with very large numbers of complex entities in the same image. Today, if you want to like confuse chat GPT, you can put 50 items on a table and take an overhead screenshot and ask it to tell you what all the items are and where they are in the image accurately. A child could do this like a 5 or a sufficiently motivated 10 year old who understands a basic coordinate system and has it overlaid could do this. Right? And AI systems struggle for the same reason. They struggle to spell words like or count the number of Rs in strawberry because of tokenization, but it's image tokens instead. So that is something that I hope gets further mitigated because that enables real manipulation of desktop class like graphical user interface software. But today anything that can be done on a command line interface, which is almost anything with the exception of like closed source software, real web browsing, basically the rich web on rich applications, can be done now by AI systems and even some rich applications. And so that's one big prediction. The other one is the rise of or the the infancy of good physical robotics powered by the beginnings of world models and the synthetic data that they create. So we're already see demos of like talk to your tractor or talk to your physical appliance where it starts moving in the real world and that it still has serious problems. And hallucination is so much more likely and dangerous when this kind of stuff happens. But having even more rudimentary controls, like really accurate like start or stop, rather than like directly controlling the crane is powerful, careful. Another thing would be like an emergency mode where you can put it in, it's like oh, my hands are pinned or something. But I need to talk the crane into a better position where it's normally not taking this kind of autonomous control or even guided control, but that's now an option for emergencies. These are the world that we're coming into right now. And when we think of what's the next frontier, especially if we get this like fast takeoff scenario of geniuses and data centers like Dario over at Anthropic is predicting the CEO of Anthropic, then then we still need to put that into the physical world. I mean that's, and that's why in the short term you're going to see a lot of people in white collar work, including even people Like AI researchers be a little nervous as they realize, oh my God, we have geniuses that don't need to sleep and are cheaper per token than I am. When I think about my salary and the food I eat and all the resources I consume, Zoom. That are doing all this really, you know, so called smart people book work or computer work where all the physical, you know, building things and blue collar is all like, wow, that's the thing we need to kind of work on. Not even, and not, not so much in terms of like a, antagonistic labor dynamics, but just because like when you look at industries where productivity has gone up, clearly tech and administration and management has in principle gone up really dramatically, right? And on making software, but productivity in like construction has actually gone down since like the 1980s. And this is actually so weird, it's perplexed even the Bureau of Labor Statistics who kind of tracks these things. And so when I think about where I want AI to be, you know, unleashed, I want better construction and that means go make better Caterpillar and, and John Deere and all these other kind of companies, like they, they need more physical, you know, imp into their day to day.
- [37:28]
B
You need more carpenters. You need more.
- [37:30]
A
Yeah, you need more. And, and you need, you need them to be augmented with better tools, right? I, I don't think that they're scalable in anywhere near the same way as software. So I think the AI for them is, is much more of an augmentation kind of. I think they've, they become a lot happier with these AI tools that are like, wow, it's like warning me about studs more intelligently. But like I still have to do most of the physical movement because paradox, right? Like Moradox says that like things that we evolve to do over millions or billions, not millions of years are really hard and, and we have like latent information. So like little babies learn after two years how to walk with dexterity way beyond even most of today's robots. But like chess, we didn't evolve to play chess. And so AI systems are really easy for us to devise that are better than us at that. And so when I think about the next true frontier of difficult AI problems that are apparently tougher than AGI or at least building something analogous to like data from Star Trek, which I think is decent, right? I think at this point most people thought we would get the Android or in the past, before Chad GPT, most people thought we would get the Android moving in a really dumb brain. We have a pretty smart brain and really tough to get like the right part of data moved. And so that's, that's now we have to work on that and get a control system.
- [38:50]
B
So you just spent, you know, a significant amount of time in healthcare. You mentioned about your personal journey five years from now. Do we have robot nurses, robot doctors,
- [39:05]
A
you know, do we have them walking and being bipedal and kind of looking like the cyberpunk movies? Five years probably. Okay. I will also pull out another quote that I think Bill Gates said or may have been misattributed to him. We humans commonly overestimate what can be done in like, I think it's like two to three years and underestimate what can be done in 10 years.
- [39:29]
B
Yes.
- [39:29]
A
And so as you lengthen any particular amount like, like I, I get like non linear gains and I think especially non linear gains as we go into you know, recursive self improvement already feeling like it's achieved in limited cases today with Claude Cod codex themselves being written with Claude code and codex almost entirely. So with that in mind, I would say we might get for quite wealthy people versions of those for really basic tasks like nursing tests. I will also point out we had. What's his name? Sorry, not Joshua Bengio, not he and LeCun. It was Geoff Hinton. Famously in 2016 predict radiologists were all out of a job by 2021. So that was another five years year plan and he got hilariously burned. I think radiologists have like highest demand ever today, right?
- [40:19]
B
Yes.
- [40:20]
A
So I really, really especially hesitate to predict too highly about what will innovate in healthcare for these kind of things. Which is why I want to push back on my like extremely rosy, bullish predictions on the future. I can tell you I've already seen like serving robots that are getting increasingly sophisticated at restaurants. Restaurants. And I've also seen even coming to America, but especially in Asia like drink making robots that can do like an automated like hand controlling everything.
- [40:50]
B
Oh cool.
- [40:51]
A
I do think that very specialized, increasingly autonomous like surgery and stuff that really helps high end humans do their job becomes far more ubiquitous in five years. Do I think we get like robo nurses? I would love for that to be true. I just don't think that's coming really in five years maybe. Sure.
- [41:10]
B
I would love that to be beach.
- [41:13]
A
Yeah, it'll probably be worse than it is better. Yeah. For a while.
- [41:17]
B
Yeah. I'd rather just give nurses more time to spend with patients.
- [41:21]
A
Absolutely. Yeah. Yeah. They're overworked horribly. That one of the hardest jobs I've Ever seen and seen up close. Try to take care of my own, especially my own mom.
- [41:29]
B
Oh my. Okay, couple questions about you personally. What's the riskiest thing you've ever done?
- [41:36]
A
Wow. Okay. I will say I've definitely made some plays around like my job or my, my work that were pretty risky.
- [41:45]
B
I had.
- [41:46]
A
I can't go into too many specifics, right. But I had an event where I got told I needed to take a little bit of money off my own salary and. And basically in order for other people to get a bigger bonus. And I threatened to quit once. And at the time there was. I mean, there's been an AI talent shortage since I've been in the industry. So I'm not being too specific, right, but. But like it worked. It was a very ballsy play and it was one where it was actually witnessed by friends of mine who came to my house and just like came into my room matter of factly and saw me at like the worst meeting I was ever having where I was like a little bit, you know, escalating my voice and all these things. So that's just in like a business professional context. I mean, I don't have anything cool like skydiving that comes to mind. Right. I guess another thing that comes to mind in terms of like physical risk. I mean, I fly a lot and that always scares me, but that's not objectively risky. I own a Corvette and went to Corvette driving school. So what kind of Corvette C8, 2025. I factory ordered. They're actually. You can get like up to 20% off if you go to the right volume dealership. Just look for. It's not like Malkin. It's like, it's like up in New Hampshire basically just looking for like the big volume one. So it's like number one Corvette dealer in the world or whatever that whoever has that title, they will hook up you up because that's where I bought it from. And, and, and so my car was like out the door. 80,000 or something. But it's. It looks stunning. It's got way more horsepower than I'd ever need. It's purple. That's amazing. Love it.
- [43:24]
B
That is cool.
- [43:25]
A
Yeah, yeah, yeah. And like that didn't even cost extra. I love. I love Chevy. I am like a GM fanboy. I am. And this is, you know, I finally had like good money. I grew up por. So this is like me. This is like my Ferrari or Lambo. I'm done with car stuff after this. So this is the height of my. I Have like mid tier SUV outside of that. This is my like darling, my. Anyway, so I, I went to driving school, drove it around the track very seriously drove it off the track because I was being too fast and dumb. I was nowhere near hurting myself directly. But I've seen video of other people screwing up worse and hitting the wall, crashing and they always walk away because the C8's really safe. But that looks like something that can kill you so.
- [44:10]
B
Well that's, that's. So did you grow up in New Hampshire? Is that.
- [44:13]
A
No, no. I was born in Kansas but I spent most of my life in Oregon. You go for that Corvette driving school. It's a very heavily discounted if you buy the car. So it's like $1,000 where they take you to PRUP Nevada for three days. All everything is paid like food, hotel. And then you do basically like eight hours a day for like three days of learning in the classroom and then driving around the track and the driving. The thing is I'm not like some NASCAR boy racer guy who like I don't follow racing. I barely know anything about it. I don't. But I like exciting things and I really like drifting and learning about like evasive driving. Like for me I want to be able to. I. It's not like some people. I think more realistically it's like bank getaway car thing. But I think for me it's like, like when I imagine oh what do I do if there's a zombie apocalypse? Obviously I probably want a Hummer or like a truck probably right. But I, I don't know, I've never really cared for that like aesthetic. I want a sports car that can like if, if there's like three road space I can get the hell out before everybody and probably not right? Like I don't know, that's like my weird like taking.
- [45:28]
B
Watch that show Paradise.
- [45:30]
A
No, I used to watch it out. I. I need to watch that. I used to watch.
- [45:35]
B
All right, so couple other questions. If you go back in time and tell your 20 year old self something, what would it be?
- [45:40]
A
Well, so that's shortly before I met my wife who I'm still married to. Happy and all that but just like treat her better I guess because you can all say that that's one thing. Buy Nvidia. The thing is, so what's funny about that too is I again with growing up poor. I didn't start making like grown UP money until 2017, 2018. So at college in my crappy little dorm room, I would mean a Buddy of mine, we would be there, you know, doing college kid things, hitting the bong, passing it around. I'm in a legal state, I can say that.
- [46:14]
B
Yes, yes, yes.
- [46:15]
A
That kind of thing.
- [46:16]
B
No judgment here. Yeah, yeah.
- [46:17]
A
I was legal at the time. U of O, we were over eight.
- [46:20]
B
No judgment here.
- [46:22]
A
Oh yeah, Deadhead. That's what I love to see. Yeah, that's what I eat. Right. So, so I'm passing it around. We knew at the time I was in my third or fourth year computer science. I knew AI was a big deal. I was messing around training LSTMs, gated recurren units, RNNs on very simple text generation tasks way even before GPT2 was, came out and was good. And I could tell Nvidia was a good stock. I knew Nvidia was good. So you know what we did? Because we didn't have real money and my other buddy was a finance major. We went on thinkorswim, which is Emeritra, and we created a monopoly money account and I had 250,000 fake dollars and we, we created a portfolio and. Which is 30 or 40% Nvidia. And that was 2017. Nvidia. Before they're split off, I've logged in and saw what happened to that account and it makes me livid because I would have good $50 million or so right now, like easily that much with just the initial investments and reinvestments that we had put in over about a year. And, and I mean, I had no hope. I, we, I had no way to raise that kind of capital.
- [47:31]
B
We'll find the next one.
- [47:32]
A
You know, the problem is that all the next ones are not in the public market. That's so I, I can't get into like, for example, you offered me, you know, $10,000 worth of anthropic stock. I'd buy it off you right now. Right. And if you even offered me 100 million in anthropic stock, I could arrange buyers for you.
- [47:52]
B
People approach me.
- [47:53]
A
Yeah, yeah, I'm not even joking. Yeah, like, I mean, buyers on Wall street for this who will give me a finder's fee that would make me rich as hell because of how much the demand is for those. So. Yeah, yeah, that's, that's where it's, it's somewhere in physical robotics, somewhere in like power generation probably. Maybe somebody will crack the nut of room temperature semiconductors. We had that like also in Korea. Oh, that's a totally different person. Right. But a scientist almost cleaned that and that was cool for A moment there.
- [48:25]
B
Cool.
- [48:25]
A
For a moment almost. Right.
- [48:27]
B
But if you weren't doing this job, what would you be doing? What are you most passionate about? Like not, not non work related really well.
- [48:34]
A
So I, I will just quickly say on, on work I would have done law. And now assuming AI comes for all the professional work mechanic or technical type like electrical engineer or. I'm really getting into like high end leather making. So like I have right here. So I don't do leather making. Right. But I'm getting into appreciating it. So I'm like buying high end boots and this is like a shell cordovan wallet here. Rokoto shell cord. I don't know if you. Yeah, it's like a very high, high end horse hide type.
- [49:07]
B
You have to hold it close to your, to your face.
- [49:09]
A
It's, it's, it's automatically blurring it because of my close.
- [49:13]
B
If you bring it close to your face I can see it better. Yeah, there we go.
- [49:17]
A
Yeah. So it's nice.
- [49:19]
B
That's very beautiful. Yeah.
- [49:20]
A
Yeah. And shell cordovan is like a really like rare like take six months. It's one of those kind of heirloom type pieces where if you see it on boots like Alden Boot company and Shoe company is like very, very famous because it gets like all the best horween shell. So I want to like make designs or get into that world but I, I just don't. I, I need to like accumulate 10 years of experience somehow in a field I know nothing about. And also I think I would also. I, I've also. How do I put it? The, the bong smoking part as I mentioned before makes it really challenging as it turns out to join the US military or anything like that. And also I have problems with the idea of being deployed overseas. So like if the US military slightly reformed how the National Guard was used to make it where they can again not be deployed overseas like it was back in the 70s and before like a war on terror. I would have joined like the officer National Guard or International Guard Reserve. Right. Basically done. My, my little bit like I really am patriotic when it comes to know the like liberal democratic US like we like are being democratic and stuff and I believe in, you know, I'm an able bodied man who has not claiming to be especially strong but a lot of countries like I have a buddy of mine in Singapore who did his peacetime military service and he's a wimp and hates military stuff and he.
- [50:46]
B
No, it's like the Israeli model too. It's Everybody serves well.
- [50:50]
A
And part of why I want that is so that way I can be very justified in saying, saying when somebody proposes a foreign war, I can be like, hell, no. And not only no, but whoever wants it, you who vote for it should be the one who gets draftable for it, which is actually Smedley Butler, which is an old Marine Corps general. They'll, they'll yell at you in the Marine Corps if you don't know who that guy is or why he's famous, by the way.
- [51:14]
B
That's right. He, he advocates Chesty puller.
- [51:17]
A
Yeah, yeah. And he, in that book wars of Rocket, he's like, those who vote for a war should be first on the lines. And so for me, I'm like, I will defend the country, but I don't want to go kill foreign foreigners. So that's, that's my take on kind of that like, patriotic element. So I would do something to. And since I'm college educated, I try to be like an officer. Probably, probably nowhere near the guns. Like one of those, like, science, like people who goes to fork and knife school.
- [51:39]
B
There's a lot of roles in the army.
- [51:41]
A
Yeah, yeah. I don't want to be. I, trust me, I, I'm, I'm not, I'm not combat material.
- [51:46]
B
You can fly drones.
- [51:48]
A
No. Well, those guys are like, if they get captured, they get like tortured and stuff. I especially don't want to do that. So.
- [51:54]
B
No, I mean from, from the safety of your home.
- [51:57]
A
I, I like the parts of the military related to being the walking with the big stick and, and being peaceful rather than the like actual dirty work. I'll point that out. So it'd probably be like medical or intelligence type work if I could. But again, in Canada, by the way, they'll let you do that. You can smoke weed or whatever on your own time. But in the US it's just a no go for a little while though. They've, they've lost, lowered, They've, they've made it easier. They've actually raised the enlistment standards to 42 just now. And they've made it where cannabis gets a special exception where no matter how many times you're like caught with that or whatever, convictions, doesn't matter anymore. No more waivers necessary. Even the army is, is admitting it. You know, just no shrooms. Well, and that's. So I'm in Oregon. We're also very progressive on that kind of thing too. And that's also something where I'm hoping that there's a proper legalization Movement specifically only for medical.
- [52:54]
B
There will be. For medical purposes.
- [52:56]
A
There needs to be. And they need to do it on shrooms and then maybe LSD and a couple other psychedelics. So that way we can get. Everything else is illegal as hell. Because I am tired of living in these fentanyl. Like look, I'm very progressive, I'm very supportive of drug legalization, but like fentanyl, meth, all these other. I've seen what it's doing to societies here in Portland or San Francisco, Vancouver.
- [53:18]
B
Awful. It's awful.
- [53:19]
A
It's awful. And, and what's worse is that now because of things like the fentanyl crisis and because of the previous opioid over prescription crisis, it's actually quite difficult for people sometimes, including my own family, to get the painkillers they need. Not right now. They haven't had any problems for the moment.
- [53:39]
B
I would even posit that, you know, recreational use of cannabis is fine, but the strength, strength, the potency of cannabis is not like it used to be 30 years ago.
- [53:49]
A
Well, and it's, it's so strong now because what happened is the states that legalized did so with, you know, they were very strict about like who could get it and like how many plants you could have for some reason, even though plants can have incredible yields for plants is I, I mean pretty much any place that allows you to grow, you can grow like multiple pounds, which even hard, hardcore. Like unless you're like a whole family of hardcore pot smokers, you are not going through pounds of that stuff.
- [54:16]
B
No.
- [54:16]
A
Over all year. And, and so I don't even, I don't even want to like, like, I'm also a civil libertarian. I think you should be able to do what you want me to. Federal regulation means that you can come like, be much more about being like, okay, we're going to have the high potency stuff for the crazy potheads and we're going to have like, okay, grandma wants to experiment with reducing her arthritis. We're going to start you nice and slow and, and also maybe, you know, have some federal oversight. So this is where I want the stuff that we can clearly point to as, you know, okay, it's going to make you dangerous to a bag of chips rather than dangerous to society or like have seriously negative outcomes. We need to clearly figure this out and then, yeah, go like, come down hard on like the people bringing in the fentanyl and all these other, their similar blights on society and health help. And it will take without their like, it's going to be Some non consensual care given to, you know, drug rehabilitation. Because Oregon's all about trying to give everybody consent to everything and very lax on crime. And I've seen it go too far here. Yeah, but, but we've also had like for example, just on the politics side, we had a three way governor race just in, I think it was 2020 where it was, we had the right wing Trump supporter, Jimmy, a fake, I might be miss screwing up the names. Then we had Tina Kotek and then we had this lady named, named like Betsy Johnson I think and she was like an older lady. She's basically an independent Democrat who was like we're going to do what I said which is like we're going to be cool on the cool drugs and we're going to come down hard on the bad drugs. And that was something that she, she was seen as like a spoiler candidate. Right. Where it's like, oh, we can't vote for her because then the Republican will win. Never got. She was definitely like the third of the three way race. Yeah, I, I voted for worker.
- [56:09]
B
Right.
- [56:09]
A
And I, I'm a straight ticket Democrat. Like I voted everybody else, you know, I did my, and I, I, I normally don't. You know, again, I'm, I'm also very, shall I put it, the kind of person who has a lot of friends who are, don't have the same political opinions as me and love me talk. I love to debate all these things. But I point this out to say that like I'm a progressive Bernie, like hippie type person. I'm saying this so it's gone too far. And, and so this is, I, I think this is medical related because if you start coming down on these and you making, you know, making it easier to get GLP ones, harder to get messed up fry your brain drugs. And then maybe just for my civil libertarian tendencies, make it a little easier for me to get low level antibiotics. I would be so happy. I claim we would get Canada. Yeah. Yes, yes. Well, I, I don't know if they do close. Yeah, yeah. I don't know if they do the antibiotics part. But in Canada, when you go to the drugstores in Canada, by the way, drugstores in Canada are really nice. There I went to one there I was like, whoa, like Walgreens and Rite Aid in America. Terrible Canada, they're nice. I could get, I could get the good cough medicine. So like there's the good cough medicine and nasal decongestant stuff. I forget the name. It's like pseudo phenadrene versus phenadrenne or.
- [57:21]
B
Yeah.
- [57:22]
A
One of them is the good stuff that can be misused to make a drug called dxm, which is like a weird, like you cook it down. I don't, I, I, I'm not that much of a. I, I, I was, I was a. Yeah. I'm not a chemist. Right. So I'm not a person who knows these things much, but I know that that's that stuff and won't let you access it.
- [57:43]
B
That's right.
- [57:43]
A
Most pharmacies in the US today. And I hate that I think that we need to accept certain. But, but so like, when I'm saying contradictory things and proposals. That's how complicated medical is. And we haven't even gotten into how you do billing or administering or even the AI agent side about how you really deploy it much either. Because I'm, you know, already just like, where do you begin? On, on even more straightforward sides of it. So.
- [58:13]
B
All right, well, a couple more questions. You're on a desert island. You can bring five records with you. What would they be?
- [58:18]
A
Wow. Okay.
- [58:20]
B
Musical taste, lean.
- [58:21]
A
So the first thing on this is I come from a family of progressive rock fans, so like, Rush is my mom favorite album ever. And so I'm being their kid. I have to rebel against that. So I don't like progressive rock.
- [58:36]
B
Okay.
- [58:37]
A
It's not, and I'm not saying it's like bad music, but I have heard Rush so much. I. All I think is my parents. It's uncool. Okay.
- [58:45]
B
They're touring again too. Your parents must be psyched.
- [58:47]
A
Again.
- [58:48]
B
Yeah.
- [58:48]
A
Oh, my God. Well, Neil Peart died, so it's not the thing.
- [58:53]
B
They found a woman drummer, Australian, maybe.
- [58:55]
A
Oh, I gotta talk to my family about that. Okay. So from their era, I, I really like. What's his name? Jim Morrison.
- [59:05]
B
What's the Doors.
- [59:06]
A
The Doors. I really like him.
- [59:08]
B
Okay, good.
- [59:09]
A
Though. Though I have to laugh at. There's all these conspiracy theories about him being alive.
- [59:14]
B
Yeah.
- [59:15]
A
So I always laugh about that stuff with. It's true.
- [59:17]
B
I believe it.
- [59:19]
A
Yeah. What else? From, from that era. I really like abba. I know. Really ridiculous.
- [59:26]
B
Good. Yeah.
- [59:27]
A
Yeah. I, I mean, I guess it's a little newer. I really like Stereo Lab, which is a bit more. Stereo Lab was, Went to Portland. I saw them recently. Big, big fan. I'm not, I don't identify as a Deadhead, but Shot nyc, which is a brand of leather jackets I really like, has a collab. The Grateful Dead tour, like whenever they wanted to get for vacation after San Francisco, they would go to Eugene, Oregon, and that meant that they were. What do you call it? They were, like, vacationing there because that's, like, where the real hippies were. I went to school there at U of O. And I like Eugene, so I feel like I would have gotten along with the Grateful Dead. So I haven't heard this, but I'd probably appreciate it. And then I guess for. For newer stuff, stuff I kind of dislike what Tame Impala has become. Not so much him. Himself in his own musical tastes. I think he's. I like him because I consider him, like, almost a type of bedroom pop. He's newer. I. It's okay if you don't know who he is, but I don't. Is he edm, like, or kind of more edm, like. But what happened is he used to be sad boy music for, like, people who were, like, recluses in college and. And, you know, and then he. For whatever reason, he went really mainstream. And the kind of girls who are, like, turning down the kind of guys who listen to that music. Five or 10 years ago, he got. He got commodified in a weird way where, like, now I hear him at every party with all these young people, and I'm just like, no, that is. He. Literally, his own music is about him lamenting about how bad he is with women and, like, all these. It's like, these women are like. Who are like, oh, I'm an Instagram model. Or listening to it and putting it out and everything. I'm like, why?
- [61:09]
B
What?
- [61:10]
A
Like, this is really. He would be. If you give him a genre, he would be like, neo psychedelia, like, very, like, tapy, like a modern Pink Floyd.
- [61:19]
B
Oh, is he, like, a bon Iver?
- [61:21]
A
Close. Yeah, yeah, yeah. I would. I would close to that kind of stuff.
- [61:25]
B
So cool.
- [61:25]
A
Yeah, that's. That's just a taste of it. I definitely have more. I'm very much, like, a fan of electric electronic music of various types, but hopefully the listeners would probably know most of those artists pretty well, so.
- [61:37]
B
Nice. Nice. Well, listen, it's been. It's been a real pleasure talking with you and sharing stories and love to have you back on again if you're open to it.
- [61:46]
A
Always. Always happy to be back on. Maybe next time I'll try to stay more on topic with AI.
- [61:51]
B
Oh, this is great. Yeah. Yeah. Well, thanks again, Alan. This is Ed Gaudette from the Risk Never Sleeps podcast. We're on the front lines protecting patient safety and delivering patient care. Remember to stay vigilant because risk never sleeps.