Commentator

Matt Hays, MD

Transcript

Mike Paskavitz

Hi, this is Mike Paskavitz, Vice President of Candello, a division of CRICO and the National Medical Malpractice Data Collaborative. Welcome to our podcast.

I’m very excited about this topic. I think there isn’t a person in the world who isn’t on some level. And we’re going to talk a bit about intelligence. And by intelligence we mean just that. And there’s two versions of it. Now there’s two versions of it. One is artificial intelligence, which seems to be everywhere. And the next is human intelligence, which has existed since the dawn of man.

Obsolescence is among the deepest fears that a human can experience. ’Robot replaces man’ has been in people’s consciousness since the industrial revolution introduced mass automation.

Many things have come true. Many things have been have been mythologized. And yet here we are, suddenly adapting to artificial intelligence becoming ubiquitous in all aspects of our daily life. And it’s happening yesterday.

So what is a human to do?

Today I’m thrilled to invite and to have a gentleman who is eminently qualified to speak about both artificial and human intelligence. Dr. Matt Hays is the senior vice president of research and analytics for amplifier, which is a very advanced learning platform. I have familiarity with it. Matt and I worked together some years ago, and the thing about Matt is, you hear the term rocket scientist, well. Matt is actually a brain scientist. He is a PhD in cognitive learning.

Matt, welcome. And please tell us a little bit about yourself and your professional journey.

Matt Hays

Thanks, Mike. I think all the things you said were true. My background is in how people learn and remember and forget, and how you can get a computer to help them learn faster, remember longer, forget slower, and transfer the things that they’ve learned more broadly.

I’m a recovering academic. I thought I was going to be a professor, and then I stood in front of a class full of students and then had to talk to them afterwards. And it quickly put me on a path toward working in industry. So I’ve been an amplifier for 12 years now and continuing to enjoy it and miss having you around there, Mike.

Mike Paskavitz

Yeah, I mean, everyone there has been great; it was a great experience. And I’m glad we stay in touch as we do. So, I had a galvanizing experience recently around the sobriety of AI. And again, there’s all these fears and feelings that go with it. There’s also excitement and opportunity that goes with it.

And I heard a presentation at the MPLA, Medical Professional Liability Association, annual meeting in Philadelphia few weeks ago. And it was by a futurist named Sam Jordan. And she gave the keynote address. And I have to tell you that the picture she painted of what life could look like in the years and decades ahead, if AI were unleashed to its full potential, was overwhelming to a lot of people in the room. Now, these are very seasoned business people, insurance executives, clinical leaders, and for whatever reason, as much as we may have heard about AI in the past up to the most recent times, there was some really amazing eyebrow raising moments in her presentation.

So, Matt, nobody has more permission than yourself to share their opinion about Sam Jordan’s message. So I’d love to know your thoughts specifically on her use of the word intelligence, for example.

Matt Hays

That’s an interesting question. So let’s talk about human intelligence first, because I think when we’re talking about AI in the way that everybody’s talking about AI right now, I am not convinced, and I think a lot of the people who have a lot of experience in how large language models actually work, I don’t think that they are convinced that this is actually intelligence at all.

There is something here and it is of incredible value, but it is a tool that produces output in a valuable way rather than something specifically intelligent. I think broadly defined, intelligence is the ability to figure out what to do when you don’t already know what to do. So if you know when you already know what to do and can reason your way through when you don’t, you are exhibiting intelligence. And an LLM-based AI isn’t actually doing that. I don’t think it’s doing almost any task that we’re asking it to do.

Just to respond to what Sam said. I think she’s an excellent speaker. I like to think about who else is weighing in on this. Yann LeCun is a Turing Award winner and is the chief AI scientist at Meta, and he has a quote that is, ’on the highway toward human level AI, large language models is an off ramp.’ And he says that ’LLMs will never be much like us.’ There’s a couple more: Gary Marcus, who ran AI at Uber, and he’s been a long-time skeptic of LLMs and other systems that are built like them. I think that he would agree every dollar spent on LLM development is a dollar that could have gone toward research on something that will get us toward human level AI.

And so all of the money that we have spent on that, we are further behind than if ChatGPT had never launched and everybody hadn’t gone wild over this.

And the last one I’ll share is I’m not sure how to pronounce his name. I think it’s Demis Hassabis. He’s the CEO of Google DeepMind. He has been knighted, which is pretty good. He won a Nobel Prize, which is decent, and his position is that LLMs need a bunch of non LLM things in order to be intelligent. Sort of in the same way that like the language chunk of your brain, which is over here, around and up above your ear on the sort of outside of your brain, if you just take that out, it’s not intelligent either. But the other parts of your brain that, that this connects to, are hard to represent artificially as well.

So the idea of the value of intelligence going to zero, I think I reject the proposition in that you still intelligence is probably more important than ever to be able to sort out what are the outputs that you’re getting from these systems, or how can you use those outputs as inputs to other systems? And so the value of true intelligence wielded appropriately, I think, is only going to go up as these tools proliferate.

Mike Paskavitz

Great. Thank you. And I think the term ’going to zero’ is what was sort of shaking about that.

So you’ve said, Matt, that you use you use AI routinely in your work. What is AI good at and what is it not so good at? And the second part of that question is AI is designed to learn and adapt, right? So as AI experiences more, it presumably learns more. It becomes better at being good.

What do you find AI to be good at and what is not as good at?

Matt Hays

There’s a myth about when and what it’s learning. So the model that you interact with over time will have some historical information about you that it uses this context for conversations, but the model that underlies it only improves when the rules around it have been changed, or when the training set has been augmented or replaced.

So it’s not that the model is better as a result of having a conversation with me or having a conversation with everybody at my organization over the course of the last three weeks. It’s only better when the training that the people producing the model…it’s only better when that training results in an incremental release of the model.

I mean that’s not that’s not the entire story. Like if you’re running your own model and you’ve got functions that do additional training periodically. But it’s not just that using AI and interacting with AI doesn’t necessarily make it better or learn. And I think that absence of learning is another thing that makes it good at some things and bad at some things, is a limitation on how you would define it as intelligent and restricts what it might ever be good at.

Mike Paskavitz

Yeah. No, I think our audience, the people that we speak with and work with, are sort of in the insurance community, but also the health care community, which sort of is driven by knowledge and intelligence and judgment and all of those things. So how AI works and supports their work and the use of it is really important across the, the, the industry that we that we work with and serve.

And I think everybody is trying to find their place with it.

Matt Hays

Okay, so here’s the thing AI is not good at. I don’t think AI is good at estimating its confidence, because I don’t think it is actually doing an estimation of its confidence. It is providing an output that looks to you like something that is evaluating its own confidence. But I don’t think there is a thing that could have confidence to be evaluated. I’m actually sure of that.

So once you recognize that AI is not something that is doing the task that you ask it to do, but is rather generating an output that looks to you like the result of having done that task.

It really helps to understand what an LLM based AI is going to be good at versus bad at. So AI excels when the task that you give it resembles what it has seen before, particularly when you have a way to verify the output.

But even non LLM based AI has been really good for a while at classification tasks. One of them is for pathology for like skin lesion identification classifying it as cancerous versus not. It’s fantastic because the training for it is you’ve already gotten the pathology reports for all of these. You’ve got a picture of the skin cancer and you’ve got the pathology report that goes to it. So you can give it a million pictures of cancerous and a million pictures of not cancerous.

And it can extract the features that determine whether something is going to be cancerous or not. It’s just so much more training than a human brain could get on this. It would be impossible to look at two million pictures of skin cancer in a reasonable amount of time, and then do something meaningful with it, with it being tagged as a one to a zero for cancer.

But the neural network model can feed that through and then see a new picture. And as long as it is related to the training distribution, as long as it is not some kind of cancer that it has not been trained on, it is I believe better, and I believe substantially better than the human eye. And the brain at diagnosing AI is great at classification tasks for things that are similar to its distribution, right?

Another thing that AI is, is very this isn’t something that it’s actually good at, but it is something really valuable that it provides. I don’t know how how ill you want to talk about Candello, Mike, but I’m sure that you have worked at an organization or two in your life where different groups have been siloed, where team A will be won’t be communicating with team B and will wind up doing the doing duplicate work, or worse, where team A has already overcome some obstacle that team B is still running into.

Those kind of silos don’t exist in an LMS model. AI does not care about those, the disciplinary boundaries that our world has sorted itself into. Like if two things are related, they are going to be proximal in the embedding space, this information space that emerges from its training. And so you wind up getting not quite original thought and maybe not even original synthesis, but something genuinely new can emerge.

In other words, one particular approach to compromising security and another approach, if there is some overlap with how a system is implemented, the model doesn’t care that nobody who uses these tools ever uses these tools. Everything’s on the table and you get very you get very interesting insight in that way. But it sort of just again, it’s still precipitates out of the model the way like rain falls out of a cloud.

It isn’t thought out of a cloud, right. But it’s still useful.

Mike Paskavitz

Yeah. Gotcha. Okay.

Matt Hays

That’s the that’s stuff it’s very good at, particularly when you have a verification system that is either a human or a rule-based computer. So like a test suite where AI writes your code and then a test suite can deterministically check to see whether that code works. And if it doesn’t, say, “here’s the error I get, here’s the problem,” and feed that back into the generative AI system, and it will get a different output that comes through the test suite. And then once that is successful, that’s a that’s a fantastic use of it. And that becomes a loop that you can set off on your own. So you let it go and it does it, and you get a good output out of it eventually. That becomes really valuable.

Where it is, where it struggles, is once you’re trying to look for something that that isn’t in its training distribution. So if you showed it, if you showed the classification system, I was talking about a new kind of skin cancer. It might make a bad judgment, it might happen to make a good judgment. But there is no there has been no learning about the thing that it is trying to do. And so there is nothing that can come out of the model that’s going to be of reliable value, let’s say. It really struggles whenever the whenever the why behind a decision wasn’t written down. So like the judgment that a senior engineer or executive or doctor or lawyer makes, by understanding all the levers and how they map into the wider world:  unless that’s in the model, unless that is part of how it shapes its output, it won’t shape its output.

And so you’ll get things that are directionally incorrect or inconsistent with what your organization wants to do.

Mike Paskavitz

So this is what I want to talk about next, which is really what everyone in our audience is committed to an interest in, which is the human intelligence. The keys to it. Knowledge, judgment, confidence. And that’s the work that you have been doing and that your organization has been built upon.

So can I just start with a question to you in your role, you know what happens to the human brain that drives learning? What is it that what do we know about learning now that has come so far in such a short time?

Matt Hays

Wow.

That is a that massive question. Let’s see. One of the one of the things I’ll tell you a few things that we know about it. And we can talk about both intelligence and learning. And there’s a couple of ways to think about intelligence. So intelligence has been conceptualized in a few different ways. But there’s this idea of this dominant, very generalizable capacity. That is how you do when you’re trying to process information, particularly applying concepts, which, again, I’d love to get back to talking about that. AI gives the appearance of having applied a concept, but can’t because there’s nothing that would be doing the applying. So one of my favorite conceptualizations of intelligence is this idea that there is fluid intelligence. This is tells model. Fluid intelligence is your ability to reason through novel problems. And crystallized intelligence is all of the knowledge that you have accumulated.

So fluid intelligence sort of peaks when you’re 25 and gradually fades off. And then but the crystallized intelligence, the amount of stuff that you know and learn and that your fluid intelligence can apply that grows for the rest of your life, that grows as long as you are experiencing new things, getting surprised and getting curious and discovering. And so then human learning becomes the thing that adds more and more of that crystallized intelligence over time, adds more and more of what the processing has to work with.

And the thing I think is most interesting that we’ve learned about learning over the last 50 years or so, is that your internal sense of it is really bad, your experience of what processes are happening in your dome is not even just unrelated to what’s actually happening, but is actually, I would say, negatively correlated with it.

So, for example, there are ways to set up your learning where, if you were going to learn about mitosis and meiosis, let’s say, I mean, in every single biology textbook and in every the brain of every single person who’s listening to this right now, it’s like, well, you should obviously learn about one and then the other. And the only debate is which to learn before which to learn first. Right. Mitosis and meiosis or the other way around?

And really what you find is if you mix those two topics together so you were learning both of them at the same time, you learn what prophase is in one and in the other, and anaphase and telephase and whatever else my my brain is rattling up from eighth grade, if you’re learning both of those at the same time, it will be harder and it will feel like you are learning worse and you will learn better and remember longer. And the reason for that is if you learn one and then the other, whatever, you just learn them. But if you learn both, I have to learn what the first phase is of mitosis, and I have to learn what the first phase is of meiosis, and I have to create some higher order cognitive structure so that I don’t confuse the two.

And that thing that my brain only makes if they are shuffled together, it’s called interleaving. I only make that structure if they’re interleaved. And that thing is really durable, and lasts a really long time. But nobody would ever think of doing it. Nobody would. There’s no world in which you’d be like, ’I should probably learn meiosis at the same time as I’m learning mitosis.’ Because it is confusing and it makes it feel harder. And people routinely interpret that experience of difficulty as evidence that they’re learning worse when it is actually evidence that they’re learning better.

Mike Paskavitz

Right? Right.

Matt Hays

There’s a bunch of these. And it’s actually it’s one of the reasons that AI is okay at instructing, but not great at instructing. Because for every paper that Lisa Son writes about interleaving, there are a million, I don’t know, maybe not a million, 10,000 biology textbooks that have mitosis and then meiosis in the next section. And so when all of those books and Lisa’s paper are fed into a large language model, and then you say, hey, teach me about mitosis and meiosis, it is doomed to do one and then the other. Even if you say ’apply the concept of interleaving,’ it will do it. But you have to invoke it. It won’t do it spontaneously.

Mike Paskavitz

Yep, yep. Understand. So Matt, one of the things that initially interested me in the work that amplifier was doing going back years ago, really stemmed from my understanding experience around medical error. So I’ve been in healthcare for a very long time in the patient safety risk management space. I think the evidence that there are knowledge-based root causes, if you will, behind a lot of medical errors, the body of evidence that that’s true even now in our own data, our malpractice data still is one of the top contributing factors to medical errors and events that ultimately lead to malpractice claims. And I think the same is true in serious reportable events, near misses, all the things that we study to understand what went wrong.

And the idea that I really clung to was this notion of confidently held misinformation, and just the idea that you are confident but wrong, but your confidence is what makes you act, resonates in the space of medicine and health care and malpractice. So tell me a bit about what you’ve learned about, through your experience with the health care organizations that you work with, confidently held misinformation, tell a little bit about what it is and what could be learned from it.

Matt Hays

Well, you nailed it. Confidently held misinformation is when you are sure you’re right, but you aren’t. And the risk as you identified is, when you are uncertain, you will look something up, or you will ask a peer or you’ll actually look at the smart card, or you’ll hit F1 and get the help. But if you’re sure you’re right, you’ll just act. And then when you’re when you’re in reality wrong, that action is either a mistake or it burdens your safety net and causes a near miss.

And the thing we’ve learned, it’s actually funny generally across domains, across doctors, across nurses, on average, just like in the low 20% of things that they would be expected to know, they are sure they know and they are wrong. So you’ve got about 20%, 23% confidently held misinformation generally across the board in medicine. And this isn’t this isn’t to impugn medicine. I have 20% commonly misinformation about the stuff I’m supposed to be an expert in. I’m sure it is in aviation. It’s in economics, it’s in accounting. People are generally not great at understanding their own learning, which is why you think that harder conditions are worse when sometimes they’re better. You also don’t have great insight into what you know and what you don’t know, and you don’t have great insight into the extent to which you should be confident in your knowledge versus not. You’re likely to be wrong in this belief. And so as a result, you end up being confident and you end up being wrong about 20% of the time.

Mike Paskavitz

In that context, talk a little bit about forgetting and remediation around misinformation.

Matt Hays

So forgetting. It’s interesting. People always seem to think that forgetting happens just because time goes by, but you really don’t. Well, actually, let’s go back. Right. So people think that forgetting is when a thing is no longer in their brain, when they don’t have that piece of information anymore, when what was storing it, the box that that piece of information was in, is now empty. And that is pretty much that’s pretty much only the case when you sustain an injury, when the when the tissue in the brain is damaged, that’s the only time you actually lose information out of the brain.

This is the crystallized intelligence we were talking about that only accumulates over the course of your life. What you lose is access to it. And so you can, after a certain amount of time, which is really like a certain amount of interfering, intervening material, you lose access to a piece of information or to a skill that you had. But this is why. This is why everyone says it’s like riding a bike, right? You don’t have access to that information, but the context of getting on the bike, your first wobble and you’re like, oh yeah, I know how to do this. It comes back as if it had never disappeared because it had never disappeared. You just lost the access to it. But you used to know it so well that the access comes back really rapidly. This is a model of this is a model of memory called the New Theory of Disuse from Robert and Elizabeth Bjork. It came out in 1992. So it’s not so new anymore, but it describes how rapidly you gain or lose access to a piece of information based on how well you already knew it and how well you know it at the time.

But people mistake the ease with which something is accessible, they mistake that accessibility, for how well it is stored. And so that’s why, like if I was preparing to give a talk, I could get ready for that talk. It’s in a couple weeks. So I could do a practice today and then watch the recording and look at my notes tomorrow, and then practice again the next day so I could spread it out over the course of, if I’m going to do four runs through this talk and do a run, I look at my notes the next day, I do a run, the next day I’m better, but I’m not great. I look at my notes. The next day I do a run I do like, but by the end I’m feeling pretty good about it. Or I could just practice my talk four times in the same day. I could run through it, look at my notes, run through it, look at my notes. By the third time, it’s clean, and by the fourth time I feel like I’m just going to fly through it.

And then the talks the next day in the timeline where I did all my studying the day before, I will have forgotten everything that I learned on the second, third and fourth tries. That’s not true. I’ve forgotten, like I’ll only be marginally better than if I just run through it once. But if I add that space in between,

I will have felt worse. The information isn’t as accessible in my brain, but every time I do more learning on it, it is entrenching it more. It is crystallizing it more and so I forget it less overnight. So when I go give that talk the next day, I do better. So you again wind up in this situation where how you experience your memory and the actual quality of your memory and its applicability in your daily life or are inverted, are anti correlated.

Mike Paskavitz

Right.

Matt Hays

Bob Bjork calls these desirable difficulties. These conditions of learning like spreading your learning out over time or interleaving them together or delaying the feedback like I gave an example. He calls them desirable difficulties because they are of course more difficult. Your performance increases more slowly. You feel worse. You’re not as fluent, the information is not as accessible. But it is a desirable difficulty because that makes you remember it better in the end.

People don’t choose these things spontaneously, because who would ever think to shuffle mitosis and meiosis together? Or how much better do I feel if I practice the talk four times in a row like that? We are completely, seduced by the feeling of fluency in our brains to think that that is the thing that reflects the learning that we’re actually doing when it, when it is the more accessible.

In fact, I would say the easier your learning feels. It is probably the case that that means the less you’re getting out of it.

Mike Paskavitz

Right? So, Matt, in the context of the world we live in, right, which is risk management, patient safety, whatnot, and the insurance that essentially underwrites that risk, if confidently held misinformation is a risk, what is the ability of that to be remediated and how does that happen?

Matt Hays

So the ability to remediate, you have to do it with intention. So confidently held misinformation is confidently held because it comes to mind very readily. We’ll do a couple of hard questions, but we’ll start with an easy one. Who invented the assembly line for automobile production? Everybody who’s listening to this knows the answer is Henry Ford. It’s not Henry Ford. And I’m not lying. This isn’t just a demonstration. It’s not Henry Ford. But the ease with which that piece of information came to your mind makes you 100% confident that that has to be it. When I tell you that the answer is actually Ransom

Olds, as an old symbol, began producing the old mobile curved dash in, I think, 1901 on an assembly line. Henry Ford improved the assembly line to make it the moving assembly line, but the model T didn’t roll off the moving assembly line until 1913. So he got beat to the industrial innovation of the century by a dozen years.

So what we have just done is, anybody who heard that just now is never going to get that wrong again, because you asked the question first, which is valuable. You wait to provide the feedback, which is valuable. But there’s also this thing that is that is particularly true about confidently held misinformation, which is there’s this hyper correction effect where if you are sure that you’re right and then you’re wrong, being confronted with that wrongness makes your brain pay extra attention to it.

But the thing is, if you never get that correction, if you’re just making decisions or classifying skin tags or whatever you’re doing and you’re not getting feedback or you’re not getting explicit training about this threshold for categorizing someone as overdue or unlikely to pay or whatever. If you’re never getting that corrected, it is just going to get further entrenched.

So it’s such a solvable problem and so hard to actually solve. In many cases because people don’t even look at confidence. They give people a multiple choice test. But somebody fills in B and you don’t know if they’re like, I don’t know, maybe B or they’re like, oh, Henry Ford, come on, get out of here. Give  me a tough one.’

So it’s fixable. It is just difficult to know that it needs to be fixed because the person in whose brain the wrong answer bubbles to the top instantly will believe that that is the thing least in need of training. The thing that they would study if you let them be in charge of their own learning, that’s one thing where they’re like, oh, I don’t need to read that, that chapter or that section. I already know that. So you have to have some externally guided training provide that correction for them.

Mike Paskavitz

So this is obviously fascinating to me.

Matt Hays

This is fascinating. I feel like to everybody it’s such a well yeah. It’s such a cool technology.

Mike Paskavitz

It’s a cool technology and an incredible time we’re in. So we’ve talked a bit about artificial intelligence and what it is, what it is. And we talked about human intelligence and how it’s driven and different dimensions of it. What would you say to our listeners out there, Matt, about how to approach AI in your professional life, your personal life, and also how to think about your own, your own knowledge development? Because I think what AI has done is it’s raise the bar for humanity, I think. And so I’m just curious what advice you would offer to people out there trying to figure out both themselves and what they’re capable of, and then also AI in its place in their lives?

Matt Hays

If you can use AI to take away the boring, repetitive manual parts of your work, you will like your work more because it will rely more on the things you can do that a machine can’t replace. This is my favorite thing about AI. You know how annoying conditional formatting rules are in Excel? Yeah. Have you ever? Well, do you even try and do it anymore because you’re like, oh, well, apply these rules, but it’s only applied to some subset of cells. Oh my gosh.

Mike Paskavitz

For me, Excel is like a game of pong on my TV from the 1980s. It’s just so yes, I have I have a challenge with that.

Matt Hays

Conditional formatting is one of the banes of my existence, and it is a delight now because I’ve got the Claude plug in for Excel. I’m like, ’highlight all of these progressively further from gray to either green or red. If they go above or below the median value of this, go.’ And then I just go look at something else, I go do something that isn’t going to be this waste of my time. And I come back and it’s so good at that sort of thing, and it’s done. I hate doing that sort of stuff. Or like, I want this PowerPoint animation to work in this sequence. And I could describe it out loud, but I couldn’t implement it on here without spending three hours. But it can. And if I can implement it out loud, I can tell it to do it. Then it does it.

So if you can, one of the things I would say is change your job so that all of the little things that are annoying, or that would take you a while to figure out, you can have Claude build you a little tool to do. Like it’ll build me a little JavaScript HTML tool for doing some modeling for scenarios that I want to game out, or the plugins for PowerPoint or Excel. I love using it for that sort of stuff.

And so there are a lot of jobs that are going to change. I don’t think it’s going to be the case that, like, 90% of us are going to be out of work in two years because of this. I don’t think it’s going to be the case that 15% of us are out of work because of this in two years.

There are going to be some people who are some executives who are so excited about this that they’re going to fire more people than they should. They will quickly realize that they shouldn’t have done that and will begin hiring some of them back, which you can. The anyone who’s listening to this can right now look and find dozens of articles about that happening at tech companies where they’re like, no, we shouldn’t have done that. We actually need these people.

But any time you feel like the first part of a customer support call when your internet’s not working, right, and they’re just going through the script and you’re like, yeah, I already tried all this. This is actually a unique situation. Once you’re in the unique situation and you need someone actually thinking if that once you’re that part of your job, give AI all the rest, all of the stuff that’s really scripted and well understood, but is hard to write as if thens, but is easy to articulate in English. And you will enjoy the job that you’re doing and you will be able to. You’ll enjoy it more. You might not necessarily enjoy it, but you’ll enjoy it more. It will burn you out less, and you will be more valuable to your organization because you will have gotten more work done.

So to the extent that you are making decisions that you understand and that you understand why you’re even being asked to make a decision, or that you could think remotely strategically about what should be done or that something being done should be done differently, and then get AI to make you a little demonstration of how. Because if you could already do it, you’d have done it. But now you can get AI to do it, and you can come to somebody and say, here’s a proposal for this that I got AI to write me the script for. And then I get it to make the visualization for. And then I got it to make the PowerPoint for. Now all of a sudden, instead of being worried that my job is getting eaten by AI, I’m doing a different job that is the job I wanted to be doing and all the crap parts of my job I have excised, I love that.

Mike Paskavitz

Yep. Yeah, I think that’s it’s kind of a nirvana for somebody thinking about their, their, their work. So, Matt,  we have covered a lot of ground. You and I could spend hours talking about this and we will, I’m sure, in the future. But thank you so much for your time and your expertise. This is again a journey we’re all on. I’m a big fan of the work you’re doing and have done, and I really appreciate the time and insights you’ve given on this topic. So thanks to Dr. Matt Hays, Senior Vice President of research and analytics for amplifier, this is Mike Paskavitz. We’ll see you next time.