2882: From Chess Grandmaster to ML Innovator: Tal Shaked's Journey

[00:00:00] Is machine learning just an extension of traditional software engineering?

[00:00:06] Or is it carving out a new frontier in the tech landscape?

[00:00:11] Well today here on Tech Talks Daily I'm going to be joined by Tal Shackad and he is a chess

[00:00:18] grandmaster turned tech visionally at a company called Maloka.

[00:00:22] And he's going to be bringing with him today a unique perspective on the evolution

[00:00:27] of machine learning and its versioning role in our digital world.

[00:00:32] And as someone who has mastered the strategic complexities of chess, he now applies his analytical

[00:00:38] prowess to the field of machine learning engineering.

[00:00:42] And he's going to challenge us to rethink how we integrate machine intelligence in a

[00:00:46] human centric world and discuss the nuance differences between artificial intelligence

[00:00:52] and machine intelligence.

[00:00:55] We're going to learn a lot today so buckle up and hold on tight as I beam your ears all

[00:00:59] the way to the US.

[00:01:00] Well, Tal is waiting to join us today.

[00:01:04] So a massive warm welcome to the show Tal.

[00:01:07] Can you tell everyone listening a little about who you are and what you do?

[00:01:12] So I was a professional chess player up until the age of 20.

[00:01:15] That was back almost 25 plus years ago.

[00:01:18] And after that I took my first programming class in 1999.

[00:01:22] And right away I was fascinated with the ability for computers to use storage and compute to

[00:01:27] solve all kinds of problems that were really difficult for humans.

[00:01:30] And I started with computer chess, later wrote programs that played games and quickly

[00:01:35] got fascinated with AI, studied AI more than 20 years ago in grad school and quickly

[00:01:40] discovered that really all the action innovation was happening in the machine learning space.

[00:01:45] And since then, after doing a bunch of internships and joining Google full time about 20 years

[00:01:52] ago, I've been thinking about how to use machine learning to work on many interesting

[00:01:56] business problems across web search, advertising, recommendations, cars, vehicles and much more.

[00:02:02] And that journey has taken me to be the chief machine learning officer at Maloko today.

[00:02:08] Maloko is a company with a vision to empower businesses of all sizes to grow through

[00:02:13] operational machine learning.

[00:02:14] Currently we focus on performance advertising.

[00:02:18] And part of my role as a leadership team is to really think about our overall strategy

[00:02:22] in machine learning and to deliver as much value as possible to our customers.

[00:02:27] What a great journey you've been on and I've got to dig into your origin story a little

[00:02:32] bit before we go any further.

[00:02:33] A professional chess player at 20 and then going into the world of tech and programming

[00:02:38] and AI, etc.

[00:02:41] Was there any link between those stories?

[00:02:43] Do you think there's a lot of parallels between those two worlds?

[00:02:47] I think, you know, talk about this probably later every human is wired in a unique way

[00:02:52] just as computers are wired in a unique way.

[00:02:54] And for whatever reason, you know, I started playing chess at the age of seven.

[00:02:58] I just picked up the game really quickly.

[00:03:01] And you know, coach said, hey, you know, to my parents, hey, your son, you know,

[00:03:05] picking up the game really quickly, maybe you should get some lessons.

[00:03:08] And after getting some lessons, I ended up, you know, being one of the strongest

[00:03:10] chess players from my age, you know, from eight onwards and just really enjoyed playing,

[00:03:17] studying and getting better.

[00:03:19] And in a weird way, you know, by my mind operated, probably a little bit like a machine

[00:03:23] learning system, right?

[00:03:24] Where if I make good moves, I win, get rewarded if I make bad moves, I lose.

[00:03:28] And how can I keep playing and learning from those mistakes to get better over time?

[00:03:33] And it says, you know, machine learning is kind of similar, right?

[00:03:36] We have lots of data points with algorithms for how the system learns and it continues.

[00:03:40] And it learns to make better decisions over time.

[00:03:41] And so I think maybe that's how my mind worked.

[00:03:44] I just was really attracted to machine learning after playing chess.

[00:03:48] It's an incredibly cool story.

[00:03:50] And one of the things that put you on my radar was when I was reading about you,

[00:03:53] that you said, I think you consider machine learning engineering to be a

[00:03:57] superset of software engineering.

[00:03:59] So can you elaborate on that?

[00:04:01] And also the kind of challenges that might present in terms of balancing

[00:04:05] infrastructure with quality because I would imagine it's quite a fine balancing

[00:04:10] Yeah. So first of all, ML engineering has many definitions.

[00:04:14] So for me, I consider it to be the discipline of building ML powered software

[00:04:19] and products. And from that perspective, it's, you know, by definition,

[00:04:23] a superset of software engineering. To contrast this, I think of software

[00:04:26] engineering as typically referring to building traditional software

[00:04:29] products that are deterministic and predictable.

[00:04:32] You know, for example, building a data warehouse or payment system,

[00:04:36] just as clearly defined requirements and expectations and either

[00:04:39] it works or it doesn't. Right?

[00:04:41] Now, when you look at ML engineering, which builds upon this,

[00:04:44] you're building products for web search recommendation systems.

[00:04:49] And as we've seen, these systems are much more

[00:04:53] non deterministic and a bit unpredictable.

[00:04:55] And they kind of change as the user engages with them.

[00:04:58] And that's, you know, in my opinion, a really different kind of discipline.

[00:05:01] There's a lot of data science, dynamical systems, experimentation.

[00:05:06] In addition to all the software you need to actually build out these systems.

[00:05:09] And so the difference is, you know, in traditional software, you know,

[00:05:12] it's against deterministic, predictable.

[00:05:15] You just build a system so it works a certain way.

[00:05:18] While in these ML engineering problems like web search or chatbots,

[00:05:22] there's a quality component to it. Is the system creative enough?

[00:05:25] Is it hallucinating? Is hallucinating good or bad?

[00:05:28] And how does that, you know, depend on different users?

[00:05:31] So that's the quality perspective.

[00:05:32] I think that's what really makes these products difficult to design and build well.

[00:05:37] And I was also really about your, well, how you view the role of machine learning

[00:05:41] as transformative rather than merely additive in the,

[00:05:45] especially in the context of software development.

[00:05:47] So can you expand on that and maybe some of the implications

[00:05:51] that they might have for the future of the industry?

[00:05:53] Because the pace of change is moving at breakneck speed at the moment.

[00:05:57] It feels like everything is evolving before our very eyes.

[00:06:00] Yeah, so, you know, I joined Google 20 years ago

[00:06:03] and I was super excited to use machine learning to solve any problem I could find.

[00:06:07] But in many of these cases, those products and problems

[00:06:09] were kind of already well defined.

[00:06:11] And they're really just what I call additive ways at machine learning.

[00:06:14] Like, take, for example, a web search.

[00:06:15] You know, 20 years ago, web search was given a query,

[00:06:18] show a bunch of blue links and try to rank those blue links as good as possible.

[00:06:23] And so the additive way to add machine learning was to say,

[00:06:26] hey, instead of having a human right of function to rank those blue links,

[00:06:29] can we use a machine learning system to decide

[00:06:32] which of those blue links is the most relevant for a query?

[00:06:35] So that's an example, kind of an additive way to add machine learning.

[00:06:38] Now imagine that you have much more powerful machine learning.

[00:06:41] How can you actually rethink the product experience

[00:06:44] so it's more of a transformative change?

[00:06:46] And that's what we're seeing today with chatbots.

[00:06:48] Now the search experience isn't really just a query and a bunch of blue links

[00:06:52] or do we have these new capabilities that allows us to have a conversation

[00:06:56] and you can just extract the information from the web

[00:06:59] and present it to the user in a much more natural way.

[00:07:01] So that's one example of sort of additive versus transformative.

[00:07:04] Another one is an advertising, right?

[00:07:06] Your advertising has traditionally been more of a brand based advertising

[00:07:10] and, you know, show ad to users and hopefully that leads

[00:07:13] to better outcomes for your products.

[00:07:15] And then, you know, we started having ads at Google, right?

[00:07:18] And we might try to optimize for clicks

[00:07:21] and then later we try to optimize conversions

[00:07:23] and then we later optimize for lifetime value for a user.

[00:07:26] But, you know, there were still a lot of manual inputs

[00:07:28] like keywords and targeting criteria, what audiences you want to reach.

[00:07:33] But what's changing with machine learning is instead of saying,

[00:07:36] give me all this data and I'll match ads to users a little bit better.

[00:07:39] Just tell me where your objective is.

[00:07:40] You know, are you trying to find users that are just valuable for you

[00:07:43] and then let the machine learning system from the ground up

[00:07:45] figure out how to find those users that perform really well.

[00:07:49] So that completely changes the product experience

[00:07:52] for how advertisers provide input and how the system optimizes.

[00:07:56] With the evolution of technology, particularly, Al,

[00:08:00] how do you anticipate the landscape of software engineering

[00:08:04] will change over the next decade?

[00:08:05] And as I asked that, I know it's an impossible question

[00:08:08] because of just how much change we've seen in the last 18 months alone.

[00:08:11] But are there any trends or hints that you're seeing

[00:08:14] on how this will evolve in the years ahead?

[00:08:17] I remember having these discussions 16 years ago, right?

[00:08:20] And back in those days, you know,

[00:08:22] this kind of state-of-the-art machine learning was maybe

[00:08:25] large-scale logistic regression, gradient boosted decision trees.

[00:08:29] And at the time we were probably training on a trillion examples,

[00:08:32] had models with 10 billion parameters,

[00:08:35] and they were used for, you know,

[00:08:37] performance advertising, web search, recommendation systems.

[00:08:42] And some people wondered,

[00:08:44] are we going to standardize machine learning

[00:08:46] the way we standardize databases,

[00:08:48] where we have SQL for databases?

[00:08:50] And people running real will be a bunch of, you know,

[00:08:53] similar tools that everyone knows how to use.

[00:08:56] And then deep learning came around and everything really changed a lot.

[00:08:59] So I think it's extremely hard to predict

[00:09:01] how machine learning is going to evolve.

[00:09:03] In my opinion, you know, I think machine learning is going to evolve

[00:09:06] kind of like software has over the last 50 years.

[00:09:08] And so in that perspective, I think of machine learning

[00:09:11] as really the future of software, right?

[00:09:14] This doesn't mean that every engineer

[00:09:15] is going to spend all their time just building models,

[00:09:18] but instead we're going to rethink sort of our design patterns,

[00:09:21] how we build software to support machine learning powered products.

[00:09:26] Right? And this will change the way we manage our data,

[00:09:29] the way we test and evaluate products,

[00:09:32] the way we do iterations, learn from our users.

[00:09:36] And we've kind of already seen this of the big tech companies

[00:09:39] who have been doing this for I think about 10 to 20 years.

[00:09:41] And I think now with more

[00:09:44] machinery technologies being open sourced,

[00:09:46] more compute capabilities being available in the cloud,

[00:09:49] many of the smaller companies, you know, particularly like Maloko,

[00:09:51] are really building these operational ML systems.

[00:09:55] And it takes a lot of work

[00:09:56] and we're gradually making it easier and easier over time.

[00:10:00] And you strike me as someone with a strong eye for detail.

[00:10:03] And one of the reasons I say that is I read that you mentioned before

[00:10:07] that there's a general lack of appreciation

[00:10:09] for what it truly means to do quality work in ML engineering.

[00:10:13] So can you discuss what quality looks like in this field

[00:10:17] and how it can be assessed and maintained?

[00:10:20] Because I get the impression

[00:10:21] this is something you're very passionate about, right?

[00:10:24] Yeah. No, I think ML engineering is a discipline

[00:10:27] is a super set of software engineering.

[00:10:29] We talked about that earlier.

[00:10:30] So let me just kind of illustrate this with a few examples.

[00:10:33] So probably the biggest learning experience

[00:10:35] I had about machine learning was when I started working

[00:10:39] on web search, trying to convince everyone that they should

[00:10:42] use machine learning to rank results rather than thousands

[00:10:45] of lines if else logic.

[00:10:47] And of course, most people weren't bought into this back then.

[00:10:50] And they had really good reasons.

[00:10:51] They said, what happens if the results are bad?

[00:10:53] How do we change the system to get to do what we want to do

[00:10:56] so that has a better user experience, higher quality results?

[00:10:59] So let me give you a concrete example that's kind of amusing.

[00:11:02] So if you search for the query miserable failure

[00:11:06] a long, long time ago, if you did that,

[00:11:07] you'd probably get some definition of miserable failure.

[00:11:10] Maybe there's a video called miserable failure.

[00:11:13] And then later you actually got a result

[00:11:16] that was the homepage of George Bush.

[00:11:18] Now, how is it that the query miserable failure returns

[00:11:21] the webpage of George Bush?

[00:11:23] Is that a good result or a bad result?

[00:11:25] Well, it turns out that users were able to reverse engineer

[00:11:29] how the Google Raking System worked and were able to create

[00:11:33] web pages and links with the word visible failure

[00:11:36] that pointed back to the website of George Bush

[00:11:39] and that and the system thought, oh, that looks like

[00:11:41] a really relevant result based on what I see in the web.

[00:11:44] So let me show that after that happened.

[00:11:46] Of course, people change the algorithm manually to avoid that.

[00:11:50] And then we search for Google miserable failure.

[00:11:52] You got a new result which is called Google bombing,

[00:11:54] which was the technique users used to find this.

[00:11:57] So the point is that for the query miserable failure,

[00:11:59] the best results kept changing over time as the world changed.

[00:12:02] And that's really a quality problem.

[00:12:04] It's not that there's an engineering challenge of just how

[00:12:06] to build a system to work correctly.

[00:12:07] It's what is the intent of the user and how do I give them

[00:12:10] the right results?

[00:12:11] And when I don't give those right results, how do I change

[00:12:14] the features, the algorithms, whatever might be to get that?

[00:12:18] And pretty much I think every modeling problem has that.

[00:12:21] You know, we see this today with with chat box, right?

[00:12:24] Are hallucinations good or bad?

[00:12:25] Well, it really depends on the context.

[00:12:28] Is the system sounding too robotic or is it interesting

[00:12:31] and creative?

[00:12:31] I want to spend more time talking with it.

[00:12:33] And those are really quality problems because you can't

[00:12:35] just get a list of requirements, build a system

[00:12:37] and say, I'm done.

[00:12:38] You really need to sort of iterate with people, with the real

[00:12:41] world to gradually improve the quality of the overall

[00:12:45] experience.

[00:12:46] And if we scroll down our news feeds right now, there are

[00:12:49] so many different stories and we're bombarded daily with

[00:12:52] talk of AI, ML and even AGI is adoring the conversation

[00:12:57] now.

[00:12:57] But the term machine intelligence is not as commonly

[00:13:01] discussed compared to ML and AI, for example.

[00:13:04] So what is your definition of machine intelligence?

[00:13:08] And why do you think the tech community should shift its

[00:13:11] focus more towards this concept?

[00:13:13] Because something we don't hear about enough.

[00:13:16] Yeah, this is a very interesting question.

[00:13:19] You know, technically AI artificial intelligence is

[00:13:22] a field that's been defined from a long time ago as a

[00:13:25] subset of computer science.

[00:13:29] But dream your hand and take it more literally, which is

[00:13:31] artificial intelligence.

[00:13:32] What does intelligence mean?

[00:13:33] According to Google, it's the ability to acquire and apply

[00:13:36] knowledge and skills.

[00:13:37] And clearly the way that humans do that and the way

[00:13:40] that machines do that is different.

[00:13:42] Similarly, you know, it's the definition of artificial,

[00:13:44] you know, well, according to Google, it's made or produced

[00:13:47] by human beings rather than occurring naturally.

[00:13:50] This makes me wonder, well, yeah, we consider we

[00:13:52] consider playing chess to be some form of intelligence,

[00:13:55] but I can tell you from personal experience that there

[00:13:57] was nothing natural about becoming a chess

[00:13:59] grandmaster.

[00:13:59] It took a lot of very hard work, very deliberate

[00:14:03] practice to kind of reach that.

[00:14:06] And you know, I think I would say that humans are

[00:14:08] designed to do that.

[00:14:09] Well, it was sort of out of my way to learn how to

[00:14:11] do that.

[00:14:12] And so to me, I think really the distinction I

[00:14:14] think about is, you know, human intelligence versus

[00:14:17] machine intelligence.

[00:14:19] And the point here is that humans and machines,

[00:14:21] they learn in very different ways.

[00:14:23] And similarly, they have very different strengths.

[00:14:25] For example, computers have almost infinite storage

[00:14:28] and almost, you know, very, very fast computational

[00:14:31] capabilities, much so more than humans.

[00:14:34] In fact, then we see this, right?

[00:14:35] If you want to have a system that can look at pictures

[00:14:38] of dogs and classify them into the right breeds,

[00:14:41] that's hard for humans.

[00:14:42] It's, you know, nowadays it's almost trivial for a

[00:14:44] machine, right?

[00:14:45] It's a different kind of problem.

[00:14:46] Same thing like basic computation, right?

[00:14:48] So to me, I think these are just really different

[00:14:50] kind of intelligences.

[00:14:52] And what we've been working on, you know, over the

[00:14:54] last, I mean, I've been involved in the last 20

[00:14:56] years is how do you teach or use machine learning

[00:15:00] to leverage the different kinds of intelligence

[00:15:02] that computers have so that they can solve or make

[00:15:06] progress on very interesting problems, particularly

[00:15:08] problems that humans are interested in.

[00:15:09] Right?

[00:15:10] And that's what we're seeing now is we're seeing

[00:15:12] many advances in machine learning technologies.

[00:15:15] And that's enabling us to better leverage,

[00:15:19] you know, the intelligence of computers to

[00:15:21] create new forms of machine intelligence that

[00:15:23] now humans can engage with an interesting way,

[00:15:25] which we see through generative AI and other

[00:15:28] kind of applications.

[00:15:29] But behind the scenes in the past, machine

[00:15:31] learning was created, was creating a ton of value,

[00:15:34] again, in performance advertising and

[00:15:37] recommendation systems and classification and

[00:15:41] other optimization fields.

[00:15:42] But those are often behind the scenes in a way

[00:15:44] that were a surface to like everyone, like

[00:15:47] we are regenerative AI.

[00:15:48] And considering the unique attributes of

[00:15:51] computers versus human intelligence, what do

[00:15:54] you think are some of the key areas where

[00:15:57] machine intelligence could possibly uniquely

[00:16:00] contribute beyond human capabilities?

[00:16:02] Is anything that excites you there that you

[00:16:04] can talk about and share?

[00:16:06] Yeah.

[00:16:07] So, I mean, we're already seeing a lot of this

[00:16:09] because I think, as I mentioned, the ability

[00:16:12] to manage a lot more information and

[00:16:15] being able to compute much more quickly,

[00:16:17] we're seeing computers or machine

[00:16:19] intelligence being able to solve problems

[00:16:22] or assist humans in solving problems better

[00:16:24] than humans can do on their own.

[00:16:26] Right?

[00:16:26] And I'll give an example again for one from

[00:16:28] chess, just because I was very close to home.

[00:16:31] Right?

[00:16:31] So when I started playing chess, there weren't

[00:16:33] really any good chess computers.

[00:16:35] You know, I had a few, I learned how to

[00:16:37] beat them and then it got kind of boring.

[00:16:39] Right?

[00:16:39] Then over time through machine learning,

[00:16:41] we generated computers that learned how to

[00:16:44] play chess through self-play and became

[00:16:46] extremely strong, much stronger than the

[00:16:48] strongest humans.

[00:16:49] And what we're seeing today is that

[00:16:51] because we have these chess-playing computers

[00:16:53] that are much stronger than before,

[00:16:54] humans themselves are becoming much better

[00:16:56] at chess at earlier ages.

[00:16:58] And as they get older, they become

[00:17:00] even better chess players now than

[00:17:02] we had in the past.

[00:17:02] Right?

[00:17:03] And why is this?

[00:17:04] Well, it turns out that, you know,

[00:17:06] because chess computers are able to

[00:17:07] play better moves and sort of intuitively

[00:17:09] spline those moves to a human by

[00:17:11] playing out different ideas,

[00:17:13] humans are actually able to learn

[00:17:15] chess better and faster than they were before.

[00:17:17] Right?

[00:17:18] So to make a simple case of, you know,

[00:17:20] computers or machine intelligence

[00:17:22] helping humans become better at something.

[00:17:24] But I think what's much more interesting than

[00:17:25] me, you know, again, this is from

[00:17:27] playing games is other fields.

[00:17:29] Right?

[00:17:30] So, you know, in advertising,

[00:17:33] I think that this is really the ability

[00:17:34] to match users with products

[00:17:37] that they'd be interested in.

[00:17:39] And this is a challenging problem.

[00:17:40] There are so many products out there.

[00:17:42] People are so busy.

[00:17:43] People spend time on the web searching for stuff.

[00:17:45] It's hard to discover everything.

[00:17:47] And so advertising is kind of a discovery

[00:17:49] mechanism to help these users

[00:17:51] find those products that they'll be interested in.

[00:17:53] And the number of products out there

[00:17:55] and the way you personalize to users,

[00:17:57] it's just immense.

[00:17:57] You know, this is a problem that humans

[00:17:59] can't really solve.

[00:18:00] But we saw this right?

[00:18:01] They try to do brand advertising,

[00:18:02] audience based targeting.

[00:18:04] And those approaches simply just do not work

[00:18:06] as well as performance advertising.

[00:18:07] And I personally have gotten better

[00:18:09] products as a result of this.

[00:18:10] And so is my wife.

[00:18:11] You know, another area I think looking

[00:18:12] forward that I'm really excited about,

[00:18:14] I would say, you know, is a health care

[00:18:17] and education. Right?

[00:18:17] I mean, we know with health care that

[00:18:20] there's so much data out there

[00:18:22] about different kinds of diseases,

[00:18:24] illnesses, symptoms.

[00:18:25] And there's no doctor that's able

[00:18:27] to understand all that and retrieve

[00:18:29] that information instantly and apply it

[00:18:31] to any kind of any patient they see.

[00:18:34] But machine learning and, you know,

[00:18:36] machine intelligence can actually do this over time.

[00:18:38] I think and we're making progress.

[00:18:40] Right? We're already seeing many cases

[00:18:41] in terms of understanding DNA,

[00:18:43] making diagnoses and even some stories

[00:18:46] about, you know, chatbots helping diagnose

[00:18:48] problems for dogs or people, you know,

[00:18:50] better than doctors could or vets.

[00:18:52] And so I think health care is going to be

[00:18:54] an area where this ability to process

[00:18:57] all kinds of data and make better sort

[00:18:59] of inferences on on diagnosis

[00:19:01] is really going to be transformed.

[00:19:03] Another one is education.

[00:19:04] Right? You know, I think many studies

[00:19:06] have shown that, you know,

[00:19:08] every person is different.

[00:19:09] If you teach through the average,

[00:19:11] you're probably not teaching any single

[00:19:12] person as well as possible for how they learn.

[00:19:15] And in fact, I think private tutoring is one

[00:19:17] of the best ways that people can learn.

[00:19:18] And I went through that.

[00:19:20] I mean, I'm obviously in chess,

[00:19:21] but even when I was younger, you know,

[00:19:22] I think I had trouble reading

[00:19:24] and I had some private tutoring there.

[00:19:25] I just I think my mind was wired a bit differently.

[00:19:28] But imagine what we can do if we had a,

[00:19:32] you know, machine intelligence or, you know,

[00:19:34] intelligent system that could customize

[00:19:36] the way they teach to every single person or kid.

[00:19:39] I think people could learn much faster

[00:19:41] and learn the way that's best for them.

[00:19:43] So I think those are just two areas.

[00:19:44] But I think really there's I think every industry

[00:19:47] is going to be transformed through kind of new

[00:19:49] capabilities of machine intelligence,

[00:19:51] especially humans and machines working together.

[00:19:55] What's your percent with you?

[00:19:56] Love that.

[00:19:57] At Maloko, your mission is to empower

[00:19:59] businesses of all sizes to grow

[00:20:01] through operational machine learning.

[00:20:03] So I've got to ask, how does Maloko

[00:20:06] integrate these advanced concepts of ML

[00:20:09] and machine intelligence into your products

[00:20:11] and services?

[00:20:12] And what kind of challenges have you encountered

[00:20:14] on this integration?

[00:20:16] Yeah.

[00:20:16] So this is a great question.

[00:20:18] I mean, this really gets at the core

[00:20:20] of what it means to be an ML first company.

[00:20:23] Right.

[00:20:24] So so there's a few parts, right?

[00:20:25] There's, you know, we can talk about

[00:20:27] what we actually do from a machine learning

[00:20:28] perspective to deliver performance advertising.

[00:20:31] And, you know, roughly speaking,

[00:20:33] we want to show, you know, the right add

[00:20:35] to the right user at the right time for the right price

[00:20:38] to kind of maximize, you know,

[00:20:40] the relevance to the user so they get value out of

[00:20:43] possibly putting on that ad and buying that product

[00:20:46] and also to the advertise themselves

[00:20:48] so they get more high value users.

[00:20:50] Right.

[00:20:51] So how do we do this?

[00:20:52] We need to build models.

[00:20:53] We need to build models that understand

[00:20:55] what that product is about, that understand

[00:20:57] the interests of the users that can predict

[00:21:01] whether if we show this ad to this user

[00:21:04] will they in our case install that app.

[00:21:06] And if they install that app, will they engage

[00:21:09] with that app a lot, find value with that app

[00:21:11] and maybe take actions on that app.

[00:21:13] If they do that, how much will they use

[00:21:16] that app over time?

[00:21:16] How much value would they create for themselves?

[00:21:18] And also how much value will they create for the

[00:21:20] for the advertiser?

[00:21:21] And so we build models that make all of those predictions

[00:21:24] and by making those predictions, we're able to

[00:21:27] target relevant ads to those users

[00:21:29] and ensure that users will find a lot of value

[00:21:32] from those the products that they advertise.

[00:21:35] So yeah, so that's kind of concretely

[00:21:36] how we use machine learning to create value

[00:21:39] and again, and then enable those businesses to grow.

[00:21:42] But beyond that, there's how do you

[00:21:44] organize your teams?

[00:21:46] How do you build the right kind of environment,

[00:21:48] the right infrastructure, the right processes

[00:21:50] so that you can build the ML powered products

[00:21:52] really quickly and efficiently?

[00:21:54] And that's something else that I put a lot of

[00:21:56] thought in having, you know, worked at many companies,

[00:21:59] you know, not just Google, but also Lyft and Snowflake.

[00:22:02] I think this is the challenge that

[00:22:04] I guess they've been working on for the last 20 years.

[00:22:07] You know, it started in a project

[00:22:09] called Cyple at Google where we want to just

[00:22:12] build the world's best large scale

[00:22:14] machinery system and we realized that

[00:22:16] we could build more of a machine learning platform,

[00:22:19] an environment that makes it easy for people

[00:22:21] to develop machine learning powered products.

[00:22:24] And that that became a system that is used by

[00:22:27] hundreds of teams.

[00:22:28] And then when deep learning came around,

[00:22:29] we've all that to tangible extended.

[00:22:31] But that was really built primarily for Google, right?

[00:22:34] I think now that I'm outside the Google bubble,

[00:22:36] working with open source,

[00:22:38] cloud technologies, it's really interesting to think about,

[00:22:41] well, how do we build the right kinds of

[00:22:44] machine learning platforms

[00:22:46] for companies like Malogo to build ML powered products

[00:22:50] and do that really quickly and efficiently?

[00:22:52] And so I think that's sort of at a more extract level.

[00:22:55] Really what I'm excited about, what I've been working on for 20 years

[00:22:58] and you know, one of the really interesting things about

[00:23:01] the work at Malogo is I'm doing that,

[00:23:02] but focusing initially in the context of performance advertising.

[00:23:05] But over time, hopefully that would span to other

[00:23:09] really interesting applications and machine learning

[00:23:10] that will create a lot of business value for future customers.

[00:23:14] Well, we started this podcast talking about your origin story,

[00:23:18] 20 year old professional chess player before you embarked

[00:23:22] on your journey in tech.

[00:23:23] And as we come full circle, I want you to look back now

[00:23:26] over your shoulder at your entire career

[00:23:29] and share the funniest or most interesting story

[00:23:32] that has happened in your career,

[00:23:34] because I would imagine that in that time

[00:23:36] you've picked up more than a few of those,

[00:23:37] but you probably can't share too many of them.

[00:23:39] Is there one that we can share today?

[00:23:42] You know, yeah, I'll share sort of

[00:23:46] my breakthrough into sort of like this,

[00:23:48] the Silicon Valley and high tech companies like Google.

[00:23:52] So, you know, as an undergrad,

[00:23:53] I went to the University of Arizona.

[00:23:55] My dad was a professor there.

[00:23:57] And the University of Arizona is

[00:23:59] known particularly for its computer science departments.

[00:24:01] Maybe it's known more for being a party school.

[00:24:04] And so it was a little bit difficult for me to kind of break

[00:24:07] into Silicon Valley and the top tech companies.

[00:24:10] I kept applying to different opportunities.

[00:24:13] And I'd say the closest I got was I did an internship

[00:24:15] at Tula Packard in Boise, Idaho in 2001.

[00:24:19] And in 2002, I applied to a bunch of places

[00:24:21] and again, I was about to go to your second internship at HP.

[00:24:24] And then one day someone from Intel called.

[00:24:27] They didn't reach me. They reached my mom.

[00:24:29] When I sent her to your son's resume,

[00:24:32] I noticed he's a professional chess player

[00:24:34] or was a professional chess player.

[00:24:36] I want to talk to him about internship.

[00:24:39] And my mom was super excited because she knew I wanted to go to Silicon Valley

[00:24:42] and actually talked to this person for a long time.

[00:24:45] She even started singing to this person in French,

[00:24:47] thinking that would help increase the chances that I'd get internship there.

[00:24:50] I don't know if it increased the chances.

[00:24:52] But anyway, eventually I got the message

[00:24:54] and I ended up talking to this person.

[00:24:56] And the person chatted with me, asked me a bunch of questions.

[00:24:58] This is Intel. We're doing chip design.

[00:25:01] They said, you know, tell you sound really smart.

[00:25:03] You know, you're a professional chess, you were a professional chess player.

[00:25:05] That's that's cool.

[00:25:07] But, you know, I've got all these other candidates

[00:25:09] that are doing PhDs and research in exactly this field.

[00:25:13] I just don't know if you're going to be the right fit for this.

[00:25:15] So we chatted for a while and the first thing that, you know,

[00:25:17] I'm going to think about on you.

[00:25:18] And so that was my internship,

[00:25:19] my first internship in Silicon Valley at Intel.

[00:25:22] And it was all because I'm being a chess player

[00:25:23] and that kind of differentiated me from other people.

[00:25:27] And then in 2003, pretty much the same thing happened.

[00:25:30] I applied to Google and Google was a small company back then.

[00:25:33] I think they had 30 interns total.

[00:25:36] And the mentor that I spoke with,

[00:25:38] that he actually became my mentor, he said, hey, you know,

[00:25:40] I'm a pretty serious go player.

[00:25:43] It's really cool that you've been a professional chess player.

[00:25:45] I mean, I've never had an intern like that before.

[00:25:47] We'd like to do an internship at Google.

[00:25:49] And that's kind of how I got my second opportunity

[00:25:52] at Silicon Valley at a top tech company.

[00:25:55] And so, you know, that was, I mean,

[00:25:57] chess was really differentiator.

[00:25:58] And what's amusing about this is, you know,

[00:26:00] things have changed so much now when I talked to kids

[00:26:02] trying to figure out to get into college and get internships.

[00:26:05] It's like, oh, everyone's getting 6,100 SATs.

[00:26:07] It's hard to differentiate myself.

[00:26:09] And so it's like, well, back when I did this,

[00:26:11] it was playing chess, you know, I don't know.

[00:26:13] Today it's a completely different challenge,

[00:26:15] but I had it not been for chess.

[00:26:16] Yeah, who knows?

[00:26:17] I might be doing something very different,

[00:26:20] not in the machine learning space, not in Silicon Valley.

[00:26:22] Wow, what an incredible story.

[00:26:24] It reminds me of the Steve Jobs quote,

[00:26:26] you know, where you can't join up the dots looking forward.

[00:26:29] When you were a 20-year-old chess player,

[00:26:31] you wouldn't have been able to do that.

[00:26:32] But looking back now, you can join up those dots easily.

[00:26:35] A great story.

[00:26:36] And before I let you go, for anyone listening,

[00:26:38] just wanting to find out more information

[00:26:40] about Meloko, the work you're doing,

[00:26:42] connect with you or your team,

[00:26:44] where's the best starting point for everything?

[00:26:46] You know, in terms of Meloko,

[00:26:47] I mean we have, you know, a website online.

[00:26:50] I think it describes some of our products.

[00:26:51] In terms of reaching out to me,

[00:26:53] you know, I'm on LinkedIn,

[00:26:54] I've been engaged with other people.

[00:26:55] They're also quite interested in sort of the entrepreneurial space,

[00:26:59] you know, many different companies

[00:27:01] founding new, new companies,

[00:27:03] developing new ML-powered products.

[00:27:06] So I'm actually a bit of an angel investor

[00:27:08] in that space as well.

[00:27:09] And so happy to worry about what people are working on

[00:27:12] and seeing how I can help.

[00:27:13] Also, well, I will add links to both the website

[00:27:17] and your LinkedIn,

[00:27:18] so make it easy for anybody listening to connect with you

[00:27:21] and carry on this conversation we started today.

[00:27:23] And we covered so much from the ML machine learning

[00:27:27] as a discipline, the future of software,

[00:27:30] people talking about ML and AI,

[00:27:33] but no one talking about machine intelligence.

[00:27:35] It's just such a rich topic

[00:27:36] that we could talk about for hours.

[00:27:38] But more than anything,

[00:27:39] just thank you for taking the time out of your busy day

[00:27:41] to sit down and share your fantastic insights

[00:27:44] and great story too.

[00:27:45] Thanks again.

[00:27:46] Thank you, Neil.

[00:27:47] I really enjoyed the conversation.

[00:27:49] And yeah, we'll ask some really great questions.

[00:27:51] And as you can tell, I'm really passionate about this space

[00:27:54] and love engaging with anyone on these topics.

[00:27:57] So as we conclude today's deep dive,

[00:28:00] I think we touched on a few provocative ideas.

[00:28:04] Ideas about the future of machine learning

[00:28:06] and its distinction from traditional software practices.

[00:28:10] And I think his insights remind us that technology evolves.

[00:28:14] That as technology evolves,

[00:28:17] so must our understanding and methodologies.

[00:28:20] So the big questions are what implications

[00:28:23] does this shift have for future innovations?

[00:28:26] How should industries adapt

[00:28:28] to embrace the full potential of machine intelligence

[00:28:31] without falling into the trap of human centric biases?

[00:28:36] It's a big topic, this one,

[00:28:37] and one we can't solve in a 30 minute podcast.

[00:28:39] But I'll happily carry on the conversation with you.

[00:28:42] You can email me techblogwriteroutlook.com,

[00:28:45] Twitter, LinkedIn, Instagram, just at Neil Seaheuse.

[00:28:49] Let's keep this conversation going.

[00:28:51] But I'm afraid we've reached the end of our time together today.

[00:28:54] So share your views.

[00:28:56] Join me tomorrow morning on Tech Talks Daily as once again,

[00:29:00] we'll continue to explore the cutting edge intersections

[00:29:03] of technology and most importantly, human ingenuity.

[00:29:08] Well, thank you for listening.

[00:29:10] And until next time, don't be a stranger.