Latest / The Adults In The AI Room / [Earley AI Podcast] Episode 73: Mark Anderson - Unlocking Hidden Patterns: A New Approach to Enterprise AI

Transcript

[Earley AI Podcast] Episode 73: Mark Anderson - Unlocking Hidden Patterns: A New Approach to Enterprise AI · The Adults In The AI Room

• 0:14
  SPEAKER_00: The early AI podcast is sponsored by early information science, political transformation journey with design and deployment of innovative technology solutions.
• 0:28
  SPEAKER_02: Hey, welcome to uh to the early AI podcast.
• 0:32
  SPEAKER_02: My name is Steph Early.
• 0:33
  SPEAKER_02: I'm your host.
• 0:33
  SPEAKER_02: Today we're going to be talking about a different approach to AI.
• 0:38
  SPEAKER_02: And while most of the world is focused on large language models, there are other groundbreaking technologies changing the way we understand and use data.
• 0:45
  SPEAKER_02: And our guest today is at the forefront of this movement.
• 0:48
  SPEAKER_02: Joining me today is Mark Anderson.
• 0:50
  SPEAKER_02: He's co-founder and CEO of Pattern Computer.
• 0:53
  SPEAKER_02: Mark has led the company for nearly a decade, building the pattern discovery engine into a platform that delivers insights traditional models cannot reach.
• 1:01
  SPEAKER_02: He's known for his work as a technologist, a strategist, an author, and he spent his career helping organizations see the future of technology and put it to work.
• 1:09
  SPEAKER_02: He's a recipient of the Alexandra J.
• 1:12
  SPEAKER_02: Noble Award for Scientific Discoveries for Fundamental Contributions in Computing and Medicine.
• 1:17
  SPEAKER_02: Mark, welcome to the show.
• 1:19
  SPEAKER_03: Thank you, Seth.
• 1:20
  SPEAKER_03: Good to be here.
• 1:22
  SPEAKER_02: So you were on uh a while ago, and it's great to have you back.
• 1:26
  SPEAKER_02: I want to uh kind of refresh people's uh memories and just to provide the story.
• 1:30
  SPEAKER_02: You're approaching your 10th anniversary.
• 1:32
  SPEAKER_02: Tell us about the journey.
• 1:33
  SPEAKER_02: How has Pattern evolved from the early startup days to where you are now?
• 1:38
  SPEAKER_02: Um, how and what does the pattern discovery engine enable that other technologies are uh cannot?
• 1:44
  SPEAKER_03: Right.
• 1:45
  SPEAKER_03: So we have a pretty cool foundational story.
• 1:49
  SPEAKER_03: Um, I'll do it in a few sentences.
• 1:51
  SPEAKER_03: Um as you know, I'm a science guy from Stanford, and so I had the scientific method tattooed in my brain when I was five years old.
• 1:59
  SPEAKER_03: And then um somewhere along that run, it occurred to me that the method wasn't perfect.
• 2:04
  SPEAKER_03: The idea of the inference step, that where did the hypothesis come from?
• 2:08
  SPEAKER_03: No one was teaching that.
• 2:10
  SPEAKER_03: And that was the key issue.
• 2:11
  SPEAKER_03: So this led to pattern recognition as the core technology that I wanted to develop, turn it from an art into a science.
• 2:18
  SPEAKER_03: And I did a chip design, a concept design, uh, that was well received, called the the next, you know, the most important processor that's not yet been invented.
• 2:27
  SPEAKER_03: And then IBM called and DARPA, and they said they had been working in secret on the same kind of chip.
• 2:33
  SPEAKER_03: So then they brought me over and I got to talk to them.
• 2:36
  SPEAKER_03: And I realized this is really serious.
• 2:38
  SPEAKER_03: And so I run, as you know, I run the FIRE or Future in Review conference, which the economist calls the best tech conference in the world.
• 2:45
  SPEAKER_03: And I had the power to put 10 world-famous CTOs into kind of a reality TV setting every year called the CTO Challenge and make up what they had to do.
• 2:55
  SPEAKER_03: And it was always something hard, you know, something different.
• 2:57
  SPEAKER_03: So I said, okay, this year, you all are gonna invent a pattern of discovery computer.
• 3:03
  SPEAKER_03: And uh a pattern computer, I'll tell you what it is, off-the-shelf hardware and any software you want to make up.
• 3:09
  SPEAKER_03: And so for a week they go away hiding, they don't go, no sleep for a week, they come out ragged and ready, and they do their delivery on a Friday morning.
• 3:17
  SPEAKER_03: And we have an audience which is CTOs, CEOs, and investors, pretty tough crowd, and they got a standing ovation, which had never happened before.
• 3:26
  SPEAKER_03: And then they got another standing ovation, two, and they leave for coffee, and I'm leaving with them.
• 3:31
  SPEAKER_03: And by the time we got to the coffee, they were funded, and I say they, we they they were funded and a staff before coffee.
• 3:40
  SPEAKER_03: So this was an indication to me that I had missed something.
• 3:43
  SPEAKER_03: I was telling them that it was a panel, but it wasn't a panel, it was a company.
• 3:47
  SPEAKER_03: And so they convinced me it was a company.
• 3:50
  SPEAKER_03: We actually hired everybody and we took the money and we started pattern.
• 3:54
  SPEAKER_03: So that's our foundational story.
• 3:56
  SPEAKER_03: And then they had to do a skip a few years forward in stealth mode, and it was like, okay, you built it.
• 4:02
  SPEAKER_03: Will it work?
• 4:03
  SPEAKER_03: And you know, so here's the first ever discovery engine where uh it's not based on something else, it's it's literally a proprietary mathematics.
• 4:13
  SPEAKER_03: Uh, we had help from Berkeley Labs in terms of what direction to go in, and we're done, and we're gonna track the sausage handle and see what comes out.
• 4:20
  SPEAKER_03: And Berkeley was in charge of a very important data set in the disease called triple negative breast cancer, which is the worst kind of breast cancer.
• 4:31
  SPEAKER_03: There's a lot of reasons why it's bad.
• 4:32
  SPEAKER_03: It's not, it kills Iowa.
• 4:34
  SPEAKER_03: And so we thought, let's try the hardest thing.
• 4:36
  SPEAKER_03: So we applied it, we applied the discovery engine to that to see where we would go.
• 4:41
  SPEAKER_03: And in fact, the result of that, even to this day, is that we created new drugs that could actually kill those cells.
• 4:48
  SPEAKER_02: Yeah.
• 4:49
  SPEAKER_02: So when you task this uh group of people with this activity or this problem, how did you frame that?
• 4:54
  SPEAKER_02: What did you tell them that they needed to do and how is this new thinking?
• 4:59
  SPEAKER_02: Uh, and and how did they come out with this this new approach?
• 5:02
  SPEAKER_02: So trying to get to the core of what was the original task that you you uh assigned that when they came back with this great presentation that got the standing ovation.
• 5:11
  SPEAKER_03: Right, right.
• 5:12
  SPEAKER_03: I said to them, I don't want you to deliver an accounting system that can do patterns.
• 5:17
  SPEAKER_03: I want you to think out of the box, give us something original.
• 5:20
  SPEAKER_03: There are no rules about this.
• 5:21
  SPEAKER_03: Uh, the complete architecture is up for grabs.
• 5:24
  SPEAKER_03: We are not gonna make chips.
• 5:26
  SPEAKER_03: So we'll anything else is fair game.
• 5:29
  SPEAKER_03: So we're gonna use existing chips, and at that time was gonna be the true north chip, which was a pattern recognition process.
• 5:34
  SPEAKER_03: And come back with a new architecture with whatever you have to do to have one job and one job only.
• 5:40
  SPEAKER_03: That job is to recognize patterns more effectively and tell us more about them.
• 5:45
  SPEAKER_03: I want to see more of them and more about them.
• 5:48
  SPEAKER_03: It'll be a true machine for finding patterns, purpose-built.
• 5:53
  SPEAKER_02: So when you're assigning this this uh task, you know, it seems very broad to say find patterns.
• 6:00
  SPEAKER_03: It's not really.
• 6:01
  SPEAKER_03: It's broad in application, but it's narrow, it's broad in application.
• 6:06
  SPEAKER_03: It's interesting that way.
• 6:08
  SPEAKER_03: There are a lot of things you can do with you know, self with ASICs and with various chip designs and architectures, including you know, IBM doing payroll.
• 6:17
  SPEAKER_03: But um even though they're all called general, they are really designed for particular types of tasks.
• 6:25
  SPEAKER_03: You know, one might be adding and subtracting, one might be rocket uh launch, you know, trajectories.
• 6:32
  SPEAKER_03: So there are a lot of things you could do, but having the generality of pattern discovery, which is truly general, but the specific job of just doing that, it was very focused.
• 6:42
  SPEAKER_03: And and in fact, it led to what we thought it would take us to, which is a new need for semantics about patterns.
• 6:51
  SPEAKER_03: So not only a new mathematics or a new architecture, a whole new way of talking about patterns and trying to deal with what you might call them old statistical problems that hadn't been really solved.
• 7:03
  SPEAKER_03: As an example, there are a lot of problems in clustering, there are a lot of problems in high-dimensional reduction, there are a lot of problems in how do you rank when you have 50,000 things, how do you rank the top 10 out of those?
• 7:13
  SPEAKER_03: How do you create a symbolic equation that is truly delivers the outcomes that are exactly what you said they would be?
• 7:19
  SPEAKER_03: So not only are you making predictions, but you're showing the actual symbolic equation, a human legible math equation you cut and paste out of an Excel sheet into another cell.
• 7:29
  SPEAKER_03: All those things were pretty much undone.
• 7:32
  SPEAKER_03: So um we had to have breakthroughs in explainability.
• 7:36
  SPEAKER_03: We had to have breakthroughs in in the mathematics of these things I just said, uh, all of which are integrated into one engine over time.
• 7:45
  SPEAKER_03: And then, you know, you push a button, it makes it sound easy, but you're doing you know 10 years of math to get that to work.
• 7:53
  SPEAKER_02: So so what what are the patterns that the system is looking for when you're uh when you're looking at that uh triple negative breast cancer drug research?
• 8:02
  SPEAKER_02: So it's what is what are the inputs and then what are the patterns that you're you're looking for?
• 8:07
  SPEAKER_03: Right.
• 8:07
  SPEAKER_03: And this comes to really an important point, which is the data.
• 8:10
  SPEAKER_03: So we learned very quickly that if we were going to get the most out of this system, we had to have not just good data, but data of a certain type.
• 8:19
  SPEAKER_03: Not fuzzy, not maybe, not handwritten, you know, not someone's personal feelings, but you know, really great data.
• 8:26
  SPEAKER_03: And we had that in this data set, it was called Metabrick.
• 8:29
  SPEAKER_03: 20 years open source, and what it was was of the genetic expressions of genes of women who had died of triple negative pests.
• 8:37
  SPEAKER_03: No question they died.
• 8:39
  SPEAKER_03: And so we had a really good, strong data set.
• 8:42
  SPEAKER_03: And even so, that we we did a review of the literature, and I think well say 80% of the papers on that data set, again open for 20 years, were with one gene.
• 8:52
  SPEAKER_03: So we're going back to the GWAS days here, you know, and then maybe two genes, maybe three, but not much.
• 8:58
  SPEAKER_03: Well, we were able to do um in like levels of six, and we found instead of a family of GWAS genes that were small, we found about 50 genes that were expressive.
• 9:07
  SPEAKER_03: So we found more genes, we found more patterns of expression of the and that allowed us to do a lot of work in making a drug design.
• 9:16
  SPEAKER_01: Okay.
• 9:17
  SPEAKER_02: So you're comparing uh the uh the uh genotype and the and the and the phenotype, you're looking at at how expressions uh genes are being expressed, and then you're looking for, and then how are you how are you associating that with drugs?
• 9:33
  SPEAKER_03: So we're actually creating drugs, but the way we're doing it is different from I think the normal Merck.
• 9:39
  SPEAKER_03: So we weren't taking a known drug and burying it a little bit, or we weren't taking a known target and trying to make a little pocket fit for that known target.
• 9:48
  SPEAKER_03: That's the normal thing.
• 9:50
  SPEAKER_03: But what we did, and we did this for a reason.
• 9:53
  SPEAKER_03: We wanted to have the lowest risk, highest speed to market, highest reward that we could do.
• 9:59
  SPEAKER_03: Save more lives more quickly instead of the usual 30-year run to it.
• 10:04
  SPEAKER_03: So what we did was we first did the dimensional reduction.
• 10:07
  SPEAKER_03: So we had a very large number of genes expressing, and we we reduced them to uh, I don't know, I guess I said 50 or 40.
• 10:15
  SPEAKER_03: We looked at them in different ways using the engine.
• 10:17
  SPEAKER_03: One way was um just do the reduction and show us who's contributing in what order.
• 10:23
  SPEAKER_03: Another way was let's filter this with another engine and only look at drugs that are already FDA approved or close to it.
• 10:29
  SPEAKER_03: Okay, fine.
• 10:30
  SPEAKER_03: Let's find drugs that are on PAT instead of off pat, which is a weird thing to do.
• 10:34
  SPEAKER_03: And then let's let's also make combinations.
• 10:36
  SPEAKER_03: So let's find, first of all, how are these that you're looking at now?
• 10:39
  SPEAKER_03: Imagine a constellation of stars, which ones are acting in families together to harm this patient.
• 10:46
  SPEAKER_03: All right, here's one family, here's another family.
• 10:48
  SPEAKER_03: Let's block two places instead of one, because we know from cancer clinical work that if you do that, you're more likely to save the patient and the cancer is less likely to escape.
• 10:58
  SPEAKER_03: So we did all that, and then we said, okay, though we're gonna make a drug or try to out of those two things, combine them.
• 11:05
  SPEAKER_03: Now we'll have to test them, see how they work together, what are their interactions, and against a disease that was unknown to be you know vulnerable to those drugs.
• 11:13
  SPEAKER_03: And so we did that, and we always did that work at some other place like Charles River or Berkeley Labs.
• 11:20
  SPEAKER_03: So we didn't test them ourselves.
• 11:22
  SPEAKER_03: And those two drugs that I've just mentioned to you, keep in mind no direct treatment at the time, even now, hardly.
• 11:29
  SPEAKER_03: And those two drugs killed, each of them killed 100% of the cancer cells.
• 11:34
  SPEAKER_03: In one case, zero healthy cells, which means truly targeted, and in one case, 15% of the healthy cells.
• 11:40
  SPEAKER_03: So we had found a way of doing something that hadn't been done.
• 11:44
  SPEAKER_02: And so, you know, there's a lot of attention on large language models.
• 11:49
  SPEAKER_02: You know, patent discovery is is very different than that.
• 11:52
  SPEAKER_02: Can you talk about how uh they work together or how they are complementary?
• 11:57
  SPEAKER_02: How do they work together?
• 11:58
  SPEAKER_02: I guess in what they are very different, right?
• 12:01
  SPEAKER_03: They are different, yeah.
• 12:02
  SPEAKER_03: And and I'm really glad that someone named them large language models because that's exactly what they are.
• 12:07
  SPEAKER_03: People try to say that they're more than that.
• 12:09
  SPEAKER_03: Um, so a large language model in my mind is a context extension machine.
• 12:15
  SPEAKER_03: And the context used to be predicting uh tokens one or three or four ahead, you know, now it's way beyond that.
• 12:23
  SPEAKER_03: It's still doing the same work, but it's on a much higher scale, larger scale.
• 12:28
  SPEAKER_03: So I see this as a language contextual engine, which uh allows us to chat with it basically, or talk to us.
• 12:36
  SPEAKER_03: That is really impressive, I think.
• 12:39
  SPEAKER_03: It's a big deal.
• 12:40
  SPEAKER_03: But it isn't the same thing as saying, can you make a major discovery in physics or chemistry or so?
• 12:47
  SPEAKER_03: It's not the job of that system to do that.
• 12:50
  SPEAKER_03: We try to force it, but it doesn't force well.
• 12:52
  SPEAKER_03: So then the question would be if you want to work together with these things.
• 12:56
  SPEAKER_03: Pretty obvious answer would be if you have the a pattern discovery engine that does do that, now you've got the best of both worlds.
• 13:03
  SPEAKER_03: You could actually talk to a system and say, I'm looking for a new drug of cancer triple negative breast cancer, how would I do that?
• 13:11
  SPEAKER_03: It would look at things that are like papers.
• 13:14
  SPEAKER_03: But the papers, by the way, about half of them are not repeatable, which is a huge scandal, which we won't talk about today.
• 13:19
  SPEAKER_03: But that's a that's a fail in the idea of just reading stuff, you know, and that's there's a limit there.
• 13:26
  SPEAKER_03: But if you said, okay, it understood what I wanted, it reviewed what had been done so far, and then it jumps over to the pattern discovery engine to do the work of making the discovery of the drugs.
• 13:37
  SPEAKER_03: Now we're really in business.
• 13:38
  SPEAKER_03: This is pretty interesting.
• 13:40
  SPEAKER_03: So I think the two things are highly complementary.
• 13:42
  SPEAKER_03: You've improved communications with human beings in ways that they understand that the we I think the system knows us better than we know ourselves in terms of like chat GPT files.
• 13:53
  SPEAKER_03: Sure.
• 13:54
  SPEAKER_03: It's a reflection.
• 13:55
  SPEAKER_03: It is a mirror.
• 13:56
  SPEAKER_03: Yeah.
• 13:57
  SPEAKER_03: So that's the that's the front end.
• 13:59
  SPEAKER_03: And the back end, if you're a scientist or a data scientist in business, is is the discovery.
• 14:05
  SPEAKER_02: And so how do you see that current technical landscape of AI and machine learning and vis-a-vis?
• 14:13
  SPEAKER_02: In a way, they are predicting patterns or they're predicting sequences, but they're not necessarily identifying those patterns.
• 14:20
  SPEAKER_02: But how do you see that technical landscape evolving?
• 14:23
  SPEAKER_02: And what is your take on the on the current technology?
• 14:27
  SPEAKER_03: Yeah, I have about the same attitude, I think, or opinion about that, that I think uh Elon has, although he can't do so.
• 14:34
  SPEAKER_03: I'm gonna say a few provocative things here.
• 14:36
  SPEAKER_03: Um, I like Elon, but I don't think he can do it yet.
• 14:39
  SPEAKER_03: He named his company XAI.
• 14:40
  SPEAKER_03: He can't do XAI.
• 14:41
  SPEAKER_03: Um, I think Eric Schmidt XAI being explainable.
• 14:44
  SPEAKER_03: Explainable AI.
• 14:45
  SPEAKER_03: He likes the letter X, so maybe that's why he did that.
• 14:48
  SPEAKER_03: Um, Eric Schmidt, very smart, you know, used to run Google.
• 14:51
  SPEAKER_03: He has been quite clear, it has to be interpretable, has to be no black box, you know, it's got to be consistent, transparent.
• 14:58
  SPEAKER_03: So these are all very important attributes that these guys and I are looking for in what I would call the next generation of platforms.
• 15:06
  SPEAKER_03: And so those are not available through LLMs today, but that doesn't mean they're not useful.
• 15:11
  SPEAKER_03: LLMs have their own use.
• 15:12
  SPEAKER_03: But that's the next step.
• 15:14
  SPEAKER_03: So if you're gonna put a timeline in order going forward, say five years, through now till four, you probably say, okay, we got the platform for LLMs.
• 15:23
  SPEAKER_03: Bang, that's done.
• 15:24
  SPEAKER_03: We're probably never gonna escape from what we can see the idea of hallucinations.
• 15:30
  SPEAKER_03: That's apparently scientifically not gonna happen.
• 15:33
  SPEAKER_03: So, but we do have the middle of this, which is the applications layer being built like crazy on top of LLMs.
• 15:40
  SPEAKER_03: There'll be a lot of money made and a lot of success there, still with hallucinations.
• 15:45
  SPEAKER_03: So gotta be careful.
• 15:46
  SPEAKER_03: And then what's next?
• 15:48
  SPEAKER_03: And then now we come back to what I was saying about Eric and Elon and myself.
• 15:51
  SPEAKER_03: So now we want to have transparency, interpretability, consistency, safety, no hallucinations.
• 15:58
  SPEAKER_03: Great.
• 15:59
  SPEAKER_03: Where's that?
• 16:00
  SPEAKER_03: And that is what people are trying to do now.
• 16:01
  SPEAKER_03: And it's if you looked at the patterns that you said, 10 years we've been doing this, so we have a lot of proof cases.
• 16:06
  SPEAKER_03: Mira Marathi's trying to do this, Ilya Suskeber, the people who are flooding out of OpenAI and out of their safety groups, they're all saying we're doing different math now.
• 16:15
  SPEAKER_03: We want to have it be safe, super intelligent.
• 16:18
  SPEAKER_03: And so I think that the you know the target is shared by some very smart people, including open AI refugees, and and everyone's trying to go do the next step of this next generation AI.
• 16:30
  SPEAKER_03: And that that's what it'll be like.
• 16:32
  SPEAKER_03: And those all those attributes are are embedded in the pattern discovery engine.
• 16:38
  SPEAKER_03: Say that again, all of those attributes are embedded, they're part of the pattern discovery engine.
• 16:43
  SPEAKER_02: How can you can you uh dig in a little bit more deeply on that?
• 16:46
  SPEAKER_03: Okay, so we have a family of products called true XAI, and under that family, we have one XAI which we use with our own products, you know, so our own engines.
• 16:56
  SPEAKER_03: So we have what Berkeley Labs and the DOE have called the most advanced explainable AI anywhere, and we use it through our own mathematics.
• 17:04
  SPEAKER_03: So, of course, we understand how it's doing, why it's doing this.
• 17:08
  SPEAKER_03: Uh, we have another product which we have used against uh it's pretty interesting, existing neural networks.
• 17:16
  SPEAKER_03: That's a hard thing to do, by the way.
• 17:17
  SPEAKER_03: So I don't know if anyone else has done it yet, but we have done that, and so we can look inside a neural network, it's a normal run forward neural network, and and come back and tell you how it did, what it did, and why it did what it did.
• 17:32
  SPEAKER_03: And then we've got other things like that.
• 17:34
  SPEAKER_03: We've got a product which isn't out yet, but if it were out, it might be able to do the same thing with LLMs, and we would probably call that GAI XAI.
• 17:44
  SPEAKER_03: So that's underway.
• 17:45
  SPEAKER_03: We're pretty close to being done with that.
• 17:47
  SPEAKER_03: So we have this whole family of explainable AI products, which are really important.
• 17:53
  SPEAKER_03: If we wrote a paper with 35 co-authors a while ago saying, what will be the impact of explainable AI when it finally comes?
• 18:00
  SPEAKER_03: And the answer was it'll change the scientific method, it'll flip it.
• 18:04
  SPEAKER_03: So instead of having a hypothesis and then you run, you know, you wind and go again, it's the opposite.
• 18:09
  SPEAKER_03: You're not allowed to have any hypothesis, you can't bring your bias to the to the game.
• 18:13
  SPEAKER_03: And instead of that, you have good data, you run the data, and you generate the hypothesis.
• 18:19
  SPEAKER_03: That's that's the right way to solve problems.
• 18:22
  SPEAKER_03: Yeah.
• 18:23
  SPEAKER_02: So so I mean, I I think I it's I have a hard time getting my mind around that because you think of, well, what is my what am I what am I trying to accomplish, or what is my theory, or what is my hypothesis, and then I'll look to try to refute that or support that with yeah, it's different.
• 18:39
  SPEAKER_03: Those are two, those are separate things now.
• 18:41
  SPEAKER_03: So the hypothesis part goes away, but the you need what we call response variables.
• 18:46
  SPEAKER_03: So the part you said, what do I want?
• 18:48
  SPEAKER_03: You have to know that.
• 18:50
  SPEAKER_03: That's the key issue now, not the hypothesis.
• 18:53
  SPEAKER_03: So, as an example, you have to say, I want to know what gene expressions are killing these women.
• 18:58
  SPEAKER_03: All right, that's your R value, we call it an R value.
• 19:01
  SPEAKER_03: But you can't bring that, you can't say, and my professor at Stanford taught me that it might be this.
• 19:06
  SPEAKER_03: No, no, you don't get to do that.
• 19:08
  SPEAKER_02: I see.
• 19:08
  SPEAKER_02: So you're asking the question without presupposing some kind of theory of the answer.
• 19:13
  SPEAKER_02: You're not trying to, you're not trying to articulate a potential answer, you're you're just asking the question.
• 19:19
  SPEAKER_03: And that's why you can make major discoveries, much larger than you personally would have made otherwise.
• 19:25
  SPEAKER_03: Because there will be things you didn't know about.
• 19:27
  SPEAKER_02: And then when you talk about explainability around neural networks, what what is the output of that?
• 19:32
  SPEAKER_02: And when you when you know explainability is kind of an ambiguous concept, right?
• 19:36
  SPEAKER_02: Yeah.
• 19:39
  SPEAKER_02: What is what does it really mean with a neural network?
• 19:42
  SPEAKER_03: So to be very specific, first of all, and I kind of was a minute ago, everyone claims to have this now.
• 19:49
  SPEAKER_03: But here's a test.
• 19:50
  SPEAKER_03: Can you give me a complete symbolic equation with a that's human readable, legible, which uh can be I'm sorry, what what kind of equation?
• 20:02
  SPEAKER_03: Symbolic equation, which we do really well, and you know, give it to me as an equation that I can read and understand, and that I can cut and paste into an Excel sheet and create the the exact outcome.
• 20:15
  SPEAKER_03: Deterministic, yeah.
• 20:17
  SPEAKER_01: Yep.
• 20:18
  SPEAKER_03: That's the test.
• 20:20
  SPEAKER_03: Pretty good test.
• 20:21
  SPEAKER_03: So um one of the things that we did to prove the case was we wanted to show the Europeans that we could um help them enforce GDPR and the AI Act, which we believe is a good idea.
• 20:32
  SPEAKER_03: And so we did a run against a lender I won't name, an American lender, and we had um uh a lot of data.
• 20:40
  SPEAKER_03: They made it public, and we ran a neural network against that data, and then we went into the neural network with our product and came out and explained it.
• 20:50
  SPEAKER_03: And what we found, this probably will make you smile, the top four determinants of who they lent to were zip code numbers.
• 20:59
  SPEAKER_03: I'm sorry, we're what zip code numbers.
• 21:02
  SPEAKER_03: Okay, now as a banker, I can tell you I'm not a banker.
• 21:05
  SPEAKER_03: The redlining is illegal, right?
• 21:07
  SPEAKER_03: Sure.
• 21:08
  SPEAKER_03: Yeah, you you shouldn't probably lend based on zip call.
• 21:11
  SPEAKER_03: But uh, they were in fact doing that.
• 21:13
  SPEAKER_03: So very useful uh application, uh particularly I think for Europe, if they wanted to find, you know, what were banks doing and and were they doing the right thing?
• 21:23
  SPEAKER_03: It would be useful to the banks if they cared about the enforcement, right?
• 21:27
  SPEAKER_02: Right.
• 21:28
  SPEAKER_02: So so when you think about some of the challenges to adoption, every time you have a breakthrough in any technology, there are always hurdles.
• 21:35
  SPEAKER_02: What have been the biggest challenges in bringing uh patterns approach to market?
• 21:39
  SPEAKER_02: And what opportunities do you see for the enterprise looking beyond, say, means right.
• 21:45
  SPEAKER_03: Biggest problem have been that people assume things.
• 21:48
  SPEAKER_03: You know, they assume if you say AI, you mean LLM.
• 21:53
  SPEAKER_03: That's just a given today.
• 21:55
  SPEAKER_03: And that has been a real challenge for us.
• 21:57
  SPEAKER_03: It's matched with the challenge we just you.
• 22:00
  SPEAKER_03: I just went through where like you it's hard to get your arms around a new way of thinking if you're really not using the old scientific method.
• 22:08
  SPEAKER_03: It's hard to get your you know your brain into a new place.
• 22:12
  SPEAKER_03: And and it takes it's just as hard for a data scientist at an enterprise as it is for a scientist or a normal human being.
• 22:20
  SPEAKER_03: So, you know, that's something which is a huge challenge.
• 22:23
  SPEAKER_03: You have to say things three times, you have to explain them in detail because they're assuming you're just like the last hundred companies they talk to, and we aren't.
• 22:32
  SPEAKER_02: So, what are some of the enterprise applications that that you see or that you've gotten traction with?
• 22:37
  SPEAKER_02: I mean, you're doing a lot in science and uh and medicine, and the data sets are very large and complex.
• 22:43
  SPEAKER_02: Do you see similar enterprise applications or what have you?
• 22:46
  SPEAKER_02: Uh we do.
• 22:47
  SPEAKER_03: We do.
• 22:48
  SPEAKER_03: Matter of fact, I think enterprise is pretty well documented by MIT and others, not doing a full-scale implementation of LLMs, using it for certain jobs, but not for all jobs.
• 22:58
  SPEAKER_03: And you know, if if imagine if you're in the enterprise and you knew you could make new discoveries in whatever you were doing.
• 23:06
  SPEAKER_03: So I'll give you an example.
• 23:08
  SPEAKER_03: We've been very successful at manufacturing quality control for a tier one aerospace company.
• 23:14
  SPEAKER_03: We solved a huge problem for them.
• 23:18
  SPEAKER_03: Mining in Australia and how to operate the machinery to be more productive.
• 23:23
  SPEAKER_03: How to run airplanes in America, domestic airline operations, improvement so they won't be laid more often.
• 23:29
  SPEAKER_03: This thing is agnostic.
• 23:30
  SPEAKER_03: And if you're in the enterprise and you have many divisions, and each would benefit not from improved language, contextual expansion, but through major new discoveries.
• 23:41
  SPEAKER_03: How many of your divisions would qualify?
• 23:44
  SPEAKER_03: And then to back that up, we just have launched, unofficially, we actually have launched this for alpha users, the pattern DE trademark online website, not available publicly yet, but we are running uh influencers through it now.
• 23:58
  SPEAKER_03: And I'll give you one example of the joy of this, I guess.
• 24:01
  SPEAKER_03: So a company which has been working on a really important question.
• 24:06
  SPEAKER_03: They think that this is an ag tech company.
• 24:09
  SPEAKER_03: They think that they kind of know that children get asthma based on the microbiome of the soil where they live worldwide, right?
• 24:17
  SPEAKER_03: Yes, it's a theory called Old Friends.
• 24:20
  SPEAKER_03: No one has been able to prove it well or define the actual parts of the microbiome that caused this by their absence.
• 24:27
  SPEAKER_03: And so they had an 18-month project with Oak Ridge running on the top supercomputers in America, Capitan and Pioneer, I think.
• 24:36
  SPEAKER_03: And then they had a great relationship with Rob Knights Lab, UCSD, good guy, friend of mine, top guy in microbiomes in the world.
• 24:44
  SPEAKER_03: Couldn't quite bring it home.
• 24:45
  SPEAKER_03: They came to our Pattern DE site after 18 months of that work and they solved it, drum roll, in 56 seconds.
• 24:53
  SPEAKER_03: And now they're publishing a paper on that.
• 24:55
  SPEAKER_03: And they know exactly which taxa of those of that microbiome are harming those children.
• 25:01
  SPEAKER_02: So what was the I mean, it sounds it it's it's hard to understand how something something that couldn't be solved in 18 months can be solved in 56 seconds.
• 25:10
  SPEAKER_02: But again, what what are what's what was the approach that they were using that made this intractable for them?
• 25:17
  SPEAKER_03: I would put it simply and just say root force doesn't always win the game, no matter how many runs you do or how many cycles you use.
• 25:25
  SPEAKER_03: There are mathematics matters, the tools matter, and if you don't have tools that can handle the stuff we talked about at the beginning of this podcast, then you just won't be able to achieve those goals no matter how hard you try or how much you want.
• 25:41
  SPEAKER_03: It's just not available.
• 25:42
  SPEAKER_03: So it's less a matter of cycles or time in the run than it is of having the right tools.
• 25:48
  SPEAKER_02: So if um you know an executive, a CTO, CIO's, you know, um innovation officer were to want to explore this approach, what what what are the mechanisms that they can use to do that?
• 26:02
  SPEAKER_02: So so how would somebody kind of investigate this case?
• 26:07
  SPEAKER_02: In other words, yeah, there must be other organizations that are thinking similarly uh about this.
• 26:13
  SPEAKER_02: It's always when people when people don't have any competition, it's kind of like, you know, is there a market there, right?
• 26:19
  SPEAKER_02: So tell me a little bit about the landscape of other researchers and investigators, you know, in this space.
• 26:25
  SPEAKER_03: Yeah.
• 26:26
  SPEAKER_03: Um there's a landscape in the world of research, there's a landscape in the world of problem solving, and and they're kind of different.
• 26:33
  SPEAKER_03: So the academic world, which is you know what that is, and then the business world has this need.
• 26:39
  SPEAKER_03: The hardest part there, I think, is just convincing them that they can actually make major discoveries after years and years and years of being frustrated that they didn't get that out of AI.
• 26:48
  SPEAKER_03: I think there's really a big gap in management between the CEO, the CXO level, and the data scientist group.
• 26:54
  SPEAKER_03: Like, what'd you do for me this year?
• 26:56
  SPEAKER_03: You know, well, it's never quite as much as they hoped.
• 27:00
  SPEAKER_03: Where do I drill this oil?
• 27:02
  SPEAKER_03: How am I going to make more money on this mining?
• 27:04
  SPEAKER_03: You get a 3% or a 5%, but you don't get 100%.
• 27:08
  SPEAKER_03: It's not a major discovery, it's more like an improvement.
• 27:11
  SPEAKER_03: And I think that's what's missing.
• 27:12
  SPEAKER_03: And you have to realize it's available to you now.
• 27:15
  SPEAKER_03: And the way we've done this is this Pattern D website, you can come there and see a known data set, which everyone has already known about.
• 27:22
  SPEAKER_03: It's open source.
• 27:23
  SPEAKER_03: You can run it in a few seconds and see what happens and compare that to how your tools work, yeah, pretty quick.
• 27:29
  SPEAKER_03: And then you can bring your own data in through Python or through the front end and try that out and find out oh, look, this is my own data and it worked.
• 27:39
  SPEAKER_01: What would a good data set be?
• 27:41
  SPEAKER_01: What a would a good target data set be to try?
• 27:44
  SPEAKER_03: Yeah, the simplest answer is anything you put into a tabular data prepare preparations, CSV files or things like that.
• 27:52
  SPEAKER_03: And doesn't matter what the data is, it can be anything.
• 27:55
  SPEAKER_03: It doesn't matter, but it's helpful if it's in a tabular data form format.
• 27:58
  SPEAKER_03: We can do other things, but the the pattern DE site wouldn't probably do them out of the box.
• 28:04
  SPEAKER_03: You know, I think what you do is call us up and we would work on something else.
• 28:08
  SPEAKER_03: So we've done things like major improvements in X-ray, kids having uh pneumonia from bacteria versus viral versus some other cause, and we were able to see what others couldn't see on X-rays.
• 28:20
  SPEAKER_03: So we can do computer vision, but we've been able to do things which are military related where we detect bioattacks from unknown sources.
• 28:28
  SPEAKER_03: We're the best apparently in class with that under a DOD grant, national lab.
• 28:32
  SPEAKER_03: So that kind of work, just come to us and ask, and we'll work on that with the discovery engine outside of that website.
• 28:39
  SPEAKER_03: But you can go to the website with any tabular data if it's clean, and you'll get a result.
• 28:43
  SPEAKER_02: That's very interesting.
• 28:44
  SPEAKER_02: So looking ahead, what do you what do you see is shaping this technology?
• 28:49
  SPEAKER_02: How do you see this technology shaping uh the next five or ten years?
• 28:53
  SPEAKER_02: What will pattern discovery look like?
• 28:56
  SPEAKER_02: How will it play along with more uh familiar tools and how should organizations uh prepare to to leverage it?
• 29:03
  SPEAKER_03: Well, that's the most exciting question and the most exciting answer we've had today, right?
• 29:07
  SPEAKER_03: So, what is the world like when this becomes common?
• 29:12
  SPEAKER_03: And the tools are available in general, and anyone who's got a major problem, whether it's in research, physics, or in business, they make a major discovery about solving that problem instead of an incremental discovery.
• 29:25
  SPEAKER_03: That's just a different world.
• 29:28
  SPEAKER_03: People are very frustrated in the big sciences, they're very frustrated in enterprise that they haven't been able to go faster at making major discoveries.
• 29:37
  SPEAKER_03: And in a world where they could go faster, where that scientific method is flipped, but where they have access to a discovery engine, it's just a much more exciting, productive place to live.
• 29:48
  SPEAKER_03: You know, I it it just kind of almost gives me chills.
• 29:52
  SPEAKER_03: And you think if you're a scientist and and you can break out of all the peer review stuff, just do a run and find something out about physics that was unknown before.
• 30:02
  SPEAKER_03: Imagine that world.
• 30:03
  SPEAKER_03: And what if you could do something in in um you know, delivery of food services around the world of Africa, you know, and it was there's somewhere to do it you hadn't even thought of based on a combination of parameters you just couldn't keep track of.
• 30:16
  SPEAKER_03: You're doing weather in every country and you're doing politics in every country, and you're doing all this stuff at once, and you reduce it to its most simple form, and you find out something you didn't know before.
• 30:25
  SPEAKER_03: That'd be great.
• 30:26
  SPEAKER_02: Yeah, I could I can imagine.
• 30:27
  SPEAKER_02: So you wrote a chapter on the future of AI recently in the Ignite uh 2034 book.
• 30:32
  SPEAKER_02: How do you see the future of AI?
• 30:35
  SPEAKER_03: Well, because of all the things we're talking about, there are some pretty unusual connections that happen, I believe, my opinion.
• 30:43
  SPEAKER_03: And one of them is that you have to use the right data to get to these major discoveries.
• 30:49
  SPEAKER_03: And if you use the word ethical, which is usually means I'm not bringing in my bias about race, creed, color, and so forth, but I'm gonna extend that a little bit and say that you could use a mathematical definition of ethics, right?
• 31:04
  SPEAKER_03: Where it's real data, it's clean, it's not being fudged, it's not incomplete.
• 31:09
  SPEAKER_03: If you apply that idea of ethicality to your data, that itself will lead to major discoveries, guaranteed.
• 31:19
  SPEAKER_03: So suddenly we're in a world which almost mixes a spiritual aspect, ethical and moral aspect of life to science and discovery.
• 31:29
  SPEAKER_03: And I believe that's exactly where we're headed.
• 31:32
  SPEAKER_03: And so, you know, the human response to that will be pretty interesting.
• 31:35
  SPEAKER_02: Yeah, yeah, that's great.
• 31:37
  SPEAKER_02: So, Mark, thank you so much for joining us and sharing your perspective on um what it means to go beyond the hype and really unlock the patterns, the power patterns.
• 31:47
  SPEAKER_02: So, and to thank thank our listeners uh for joining us this week.
• 31:50
  SPEAKER_02: And we'll look forward to seeing you at the next AI podcast.
• 31:54
  SPEAKER_02: Follow us for more conversations, and uh, we'll have your uh LinkedIn information and your website in the show notes.
• 32:00
  SPEAKER_02: Uh again, thank you, Mark, for for joining us today.
• 32:03
  SPEAKER_03: Thank you, Seth.
• 32:03
  SPEAKER_03: I had a good time.
• 32:04
  SPEAKER_02: Absolutely.
• 32:05
  SPEAKER_02: We'll talk to you soon.
• 32:06
  SPEAKER_02: Thank you.
• 32:06
  SPEAKER_01: Thank you.
• 32:10
  SPEAKER_00: Thank you for joining us on another deep dive into AI innovation.
• 32:14
  SPEAKER_00: Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world.
• 32:20
  SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science and Vector.
• 32:26
  SPEAKER_00: To learn more, visit early.com.
• 32:28
  SPEAKER_00: That's E A R L E Y dot com.
• 32:31
  SPEAKER_00: Thank you for tuning in.