WEBVTT - How to Capitalize on Artificial Intelligence

0:00:00.120 --> 0:00:03.960
<v Speaker 1>This is Bloomberg Business Week with Carol Messer and Tim

0:00:04.000 --> 0:00:06.280
<v Speaker 1>Stenebek on Bloomberg Radio.

0:00:06.840 --> 0:00:10.080
<v Speaker 2>All Right, when someone says AI, they can be referring

0:00:10.119 --> 0:00:11.719
<v Speaker 2>to a lot of different things, you know of that, right,

0:00:11.720 --> 0:00:14.320
<v Speaker 2>They can be talking about chatbots such as chat Gibt

0:00:15.040 --> 0:00:18.599
<v Speaker 2>or Google Bard, which are examples of LM's or large

0:00:18.720 --> 0:00:21.120
<v Speaker 2>language models. We talked. We talked about this a little

0:00:21.120 --> 0:00:21.960
<v Speaker 2>bit with Mendeep earlier.

0:00:22.079 --> 0:00:24.640
<v Speaker 1>Exactly. Okay, what about when it comes to AGI, artificial

0:00:24.640 --> 0:00:27.280
<v Speaker 1>general intelligence. It's the holy grail of AI. It's not

0:00:27.400 --> 0:00:29.680
<v Speaker 1>around yet, but everyone is working on it. What that

0:00:29.760 --> 0:00:32.160
<v Speaker 1>means is that AI can perform as well or better

0:00:32.200 --> 0:00:34.839
<v Speaker 1>than humans on most tasks.

0:00:35.000 --> 0:00:37.280
<v Speaker 2>All right, so machine learning is something else. It's the

0:00:37.320 --> 0:00:39.680
<v Speaker 2>focus of our next guest, new book. Eric Siegel is

0:00:39.720 --> 0:00:42.519
<v Speaker 2>a consultant and former Columbia University professor. He's also the

0:00:42.560 --> 0:00:45.879
<v Speaker 2>author of a new book. It is entitled The AI Playbook,

0:00:45.920 --> 0:00:48.960
<v Speaker 2>Mastering the Rare Art of Machine Learning Deployment. It is

0:00:49.120 --> 0:00:53.159
<v Speaker 2>out today. Eric joins us on Zoom from the Bay Area. Eric, congratulations,

0:00:53.600 --> 0:00:56.200
<v Speaker 2>Let's start with the basics, because I think sometimes we

0:00:56.280 --> 0:01:00.720
<v Speaker 2>all throw around phrases assuming everybody gets it and they

0:01:00.720 --> 0:01:03.840
<v Speaker 2>don't necessarily. How do you define machine learning?

0:01:05.400 --> 0:01:09.319
<v Speaker 3>Well, thanks Carol. Machine learning is technology that learns from

0:01:09.520 --> 0:01:13.480
<v Speaker 3>experience in order to make predictions in order to target

0:01:13.800 --> 0:01:17.000
<v Speaker 3>and improve large scale operations. So who to target to

0:01:17.360 --> 0:01:21.640
<v Speaker 3>for marketing based on predicting who's going to buy which transaction?

0:01:21.720 --> 0:01:24.360
<v Speaker 3>To audit for fraud based on predicting which is going

0:01:24.440 --> 0:01:27.240
<v Speaker 3>to be turned out to be fraudulent. Who to approve

0:01:27.280 --> 0:01:29.840
<v Speaker 3>for credit application based on who's going to be the

0:01:29.880 --> 0:01:35.399
<v Speaker 3>most reliable debtor, etc. Etc. Which satellite to investigate for

0:01:35.440 --> 0:01:38.000
<v Speaker 3>potentially running out of a battery based on whether it's

0:01:38.000 --> 0:01:41.440
<v Speaker 3>going to where to drill for oil? Pretty much this

0:01:41.520 --> 0:01:43.280
<v Speaker 3>is the type of AI, and you could call it

0:01:43.360 --> 0:01:47.800
<v Speaker 3>predictive AI or predictive analytics to differentiate it from generative AI.

0:01:47.880 --> 0:01:50.680
<v Speaker 3>This is the type of AI you turn to when

0:01:50.720 --> 0:01:54.040
<v Speaker 3>you want to improve pretty much any and all of

0:01:54.080 --> 0:01:58.960
<v Speaker 3>your large scale operations, which are which consists of many decisions,

0:01:59.040 --> 0:02:02.720
<v Speaker 3>and prediction is the holy grail for improving decisions.

0:02:02.760 --> 0:02:07.040
<v Speaker 1>Would you consider the way that content is surfaced on

0:02:07.880 --> 0:02:12.239
<v Speaker 1>a social media platform like x slash, Twitter or Instagram?

0:02:12.440 --> 0:02:14.120
<v Speaker 1>Would you consider that machine learning?

0:02:16.000 --> 0:02:16.960
<v Speaker 3>Oh? I see? Yeah?

0:02:17.120 --> 0:02:19.480
<v Speaker 1>As far as the ordering of your feed, yeah, like

0:02:19.520 --> 0:02:20.880
<v Speaker 1>if I log in, you know, if I start to

0:02:20.880 --> 0:02:22.800
<v Speaker 1>follow someone new, Like if I follow someone new on

0:02:22.840 --> 0:02:25.800
<v Speaker 1>Instagram and then I log in Instagram again, that like

0:02:25.800 --> 0:02:27.640
<v Speaker 1>an old post from that person is going to be

0:02:27.680 --> 0:02:29.839
<v Speaker 1>like the first thing that comes up because Instagram thinks

0:02:29.880 --> 0:02:30.960
<v Speaker 1>I'm now interested in that.

0:02:31.840 --> 0:02:33.720
<v Speaker 3>Right, that's a prediction task. And the same thing with

0:02:33.760 --> 0:02:36.600
<v Speaker 3>the ordering of your Facebook feed the default feed, assuming

0:02:36.639 --> 0:02:38.639
<v Speaker 3>you leave it at that, and the same thing in

0:02:38.720 --> 0:02:41.520
<v Speaker 3>the ordering of your Google search results. It's all based

0:02:41.560 --> 0:02:44.760
<v Speaker 3>on predictive models. That's what machine learning generates from data.

0:02:44.840 --> 0:02:48.160
<v Speaker 3>Is a model that captures the patterns, that's the discoveries

0:02:48.160 --> 0:02:50.560
<v Speaker 3>it's made from data. That helps it predict. Predict is

0:02:50.560 --> 0:02:52.760
<v Speaker 3>the action. So you're predicting in order to say which

0:02:52.760 --> 0:02:54.679
<v Speaker 3>of these content ten items is going to be of

0:02:54.800 --> 0:02:57.880
<v Speaker 3>most interest or most relevant, whether it's for Internet search

0:02:57.960 --> 0:03:01.040
<v Speaker 3>like with Google, or the ordering of your news feed,

0:03:01.120 --> 0:03:04.200
<v Speaker 3>or the ordering of your search results for properties on Airbnb.

0:03:04.720 --> 0:03:08.880
<v Speaker 2>All right, whiteboard it for me. So AI big umbrella

0:03:09.040 --> 0:03:12.520
<v Speaker 2>machine learning a form of AI. How do we stack

0:03:12.600 --> 0:03:15.880
<v Speaker 2>up and make sense in this environment of because we

0:03:15.919 --> 0:03:18.280
<v Speaker 2>do throw around AI, which has been around for a

0:03:18.320 --> 0:03:21.560
<v Speaker 2>long time, but now we're talking about you know, generative AI.

0:03:22.440 --> 0:03:24.520
<v Speaker 2>Give me the whiteboard on it. You're teaching a class, like,

0:03:24.800 --> 0:03:26.400
<v Speaker 2>how do you lay it all out in terms of

0:03:26.560 --> 0:03:29.280
<v Speaker 2>AI what it means or machine learning how it'll play

0:03:29.280 --> 0:03:29.639
<v Speaker 2>a role.

0:03:31.080 --> 0:03:34.800
<v Speaker 3>Well, the typical hierarchy is that machine learning is part

0:03:34.800 --> 0:03:37.560
<v Speaker 3>of AI, and that's a more bigger umbrella term. But

0:03:37.800 --> 0:03:40.440
<v Speaker 3>my opinion varies from a lot of the mainstream, although

0:03:40.440 --> 0:03:43.240
<v Speaker 3>more people are jumping on board that AI is really

0:03:43.280 --> 0:03:46.440
<v Speaker 3>an amorphous, ill defined term. We're trying to ascribe the

0:03:46.520 --> 0:03:50.120
<v Speaker 3>word intelligence to a machine. That's quite problematic to nail down.

0:03:50.280 --> 0:03:53.320
<v Speaker 3>But if you don't define something well, you can't pursue it.

0:03:53.360 --> 0:03:56.920
<v Speaker 3>For engineering, AI is the story we hear about machine

0:03:56.960 --> 0:03:59.240
<v Speaker 3>learning is the technology that we have, and machine learning

0:03:59.320 --> 0:04:01.480
<v Speaker 3>in all those ways. I just describe where you're predicting

0:04:01.640 --> 0:04:05.200
<v Speaker 3>for each individual customer healthcare client, as far as their

0:04:05.200 --> 0:04:08.080
<v Speaker 3>disease progression, or where to drill for oil which I'll

0:04:08.160 --> 0:04:11.080
<v Speaker 3>like to investigate in which transaction to audit on that

0:04:11.200 --> 0:04:15.080
<v Speaker 3>individual level. It's the same core technology that drives generative

0:04:15.080 --> 0:04:19.440
<v Speaker 3>AI for its ability to generate first drafts of writing

0:04:19.800 --> 0:04:22.960
<v Speaker 3>or code, or of images, and in those cases what

0:04:22.960 --> 0:04:25.440
<v Speaker 3>it's doing is predicting what should the next word be Okay,

0:04:25.440 --> 0:04:27.320
<v Speaker 3>well it's actually the next token, but it's on that

0:04:27.400 --> 0:04:30.040
<v Speaker 3>level of detail. What should the next word be? How

0:04:30.040 --> 0:04:33.360
<v Speaker 3>should I change this individual pixel in an iteration? As

0:04:33.400 --> 0:04:36.799
<v Speaker 3>I'm rendering this image, I being the computer in this case.

0:04:37.680 --> 0:04:40.880
<v Speaker 3>It's the same core technology learning from data to predict.

0:04:41.320 --> 0:04:44.560
<v Speaker 1>You know, it's funny, Carol, when you're so just a

0:04:44.600 --> 0:04:48.039
<v Speaker 1>little behind the scenes. Eric, It's like, we use this

0:04:48.080 --> 0:04:50.520
<v Speaker 1>Google doc to prepare the show, and we both work

0:04:50.520 --> 0:04:53.039
<v Speaker 1>on it and work on different things. And Google now

0:04:53.160 --> 0:04:56.560
<v Speaker 1>has predictive texts, And when I was writing this one

0:04:57.839 --> 0:05:01.200
<v Speaker 1>for Eric's intro, it actually got some stuff wrong, which

0:05:01.200 --> 0:05:02.960
<v Speaker 1>I thought was really funny because here we are talking

0:05:02.960 --> 0:05:05.560
<v Speaker 1>about the technology that it's getting wrong. When I'm like,

0:05:05.640 --> 0:05:07.839
<v Speaker 1>you know, how do you writing a question in the doc?

0:05:07.920 --> 0:05:08.839
<v Speaker 1>I mean, have you noticed that?

0:05:09.000 --> 0:05:11.120
<v Speaker 2>Yeah? No, you're absolutely right if you're very careful, because

0:05:11.120 --> 0:05:13.640
<v Speaker 2>it jumps ahead right if you don't catch it. So, yeah,

0:05:13.640 --> 0:05:15.560
<v Speaker 2>how do we make sure all this stuff that were

0:05:16.040 --> 0:05:19.480
<v Speaker 2>it feels like increasingly moving towards relying on make sure

0:05:19.520 --> 0:05:23.520
<v Speaker 2>whether it's machine learning, that it's accurate and predicting right

0:05:23.560 --> 0:05:27.240
<v Speaker 2>things or the right outcomes or the smart outcomes. How

0:05:27.240 --> 0:05:28.640
<v Speaker 2>do we make sure I know. Is it just the

0:05:28.720 --> 0:05:30.719
<v Speaker 2>data sets that go into it? Is it that simple

0:05:31.520 --> 0:05:32.560
<v Speaker 2>and that complicated? No?

0:05:32.640 --> 0:05:37.560
<v Speaker 3>I mean no, no, because we're not headed with definitiveness

0:05:37.720 --> 0:05:42.520
<v Speaker 3>towards reliability when it comes to that type of generated text. Look,

0:05:42.800 --> 0:05:46.200
<v Speaker 3>these models are so seemingly human like. They're amazing. I

0:05:46.279 --> 0:05:48.880
<v Speaker 3>spent six years of my career in the Natural Language

0:05:48.880 --> 0:05:51.719
<v Speaker 3>Processing Research Group at Columbia, and I never thought I

0:05:51.720 --> 0:05:54.680
<v Speaker 3>would see what we're seeing today. It's so amazing. The

0:05:54.720 --> 0:05:58.320
<v Speaker 3>way it creates often cohesive content, can talk about anything,

0:05:58.680 --> 0:06:01.800
<v Speaker 3>use expressions the humans use, and because it's trained over

0:06:01.880 --> 0:06:06.200
<v Speaker 3>so much data and the actual modeling itself is so advanced. However,

0:06:06.800 --> 0:06:08.960
<v Speaker 3>what it's trained to do is essentially on that per

0:06:09.040 --> 0:06:11.480
<v Speaker 3>word level of detail, which really gives it a human

0:06:11.640 --> 0:06:14.680
<v Speaker 3>like aura. But that doesn't mean that it was developed

0:06:15.240 --> 0:06:19.440
<v Speaker 3>to pursue higher order human goals like being correct. That's

0:06:19.480 --> 0:06:22.800
<v Speaker 3>a whole nother thing. The fact that it's seemingly human

0:06:22.920 --> 0:06:26.760
<v Speaker 3>like doesn't mean it's a step towards general human behavior.

0:06:27.240 --> 0:06:31.280
<v Speaker 3>So you know you earlier Tim mentioned artificial general intelligence.

0:06:31.360 --> 0:06:33.039
<v Speaker 3>I'm actually sort of a disbeliever in that.

0:06:33.080 --> 0:06:35.520
<v Speaker 1>I don't think it's Why are you a disbeliever.

0:06:35.720 --> 0:06:39.279
<v Speaker 3>Yeah, I don't think it's technically impossible that someday, But

0:06:39.400 --> 0:06:41.840
<v Speaker 3>I do not believe that any of the advancements, as

0:06:41.880 --> 0:06:46.680
<v Speaker 3>impressive and valuable as they are, actually represent a concrete

0:06:46.720 --> 0:06:51.560
<v Speaker 3>step towards general human level capabilities. Where the machine is basically,

0:06:51.720 --> 0:06:54.240
<v Speaker 3>let's call it what it is in the story, it's

0:06:54.240 --> 0:06:57.880
<v Speaker 3>an artificial human. You can onboard them like an employee

0:06:57.920 --> 0:07:00.240
<v Speaker 3>at like a human employee, and let it rip. They

0:07:00.240 --> 0:07:03.360
<v Speaker 3>can run a fortune five hundred company, whatever it is.

0:07:03.680 --> 0:07:07.080
<v Speaker 3>That is a science fiction fantasy. I do not believe

0:07:07.400 --> 0:07:09.360
<v Speaker 3>that we're taking concrete steps in that direction.

0:07:09.520 --> 0:07:11.280
<v Speaker 1>So you're not concerned about the rise.

0:07:11.120 --> 0:07:16.000
<v Speaker 3>Of the bots, right, And that's they call it critic hype,

0:07:16.280 --> 0:07:18.000
<v Speaker 3>right where you say, hey, look this stuff is so

0:07:18.160 --> 0:07:21.160
<v Speaker 3>good it could kill all of us. It's really just

0:07:21.200 --> 0:07:25.480
<v Speaker 3>another way to sort of mismanage expectations. And there's a

0:07:25.560 --> 0:07:29.520
<v Speaker 3>variety of reasons why people do that. Some genuinely believe it.

0:07:29.840 --> 0:07:33.880
<v Speaker 3>I'm trying to calm the world down a little bit here.

0:07:34.080 --> 0:07:38.320
<v Speaker 3>The stuff is extremely valuable and in what it can

0:07:38.360 --> 0:07:42.360
<v Speaker 3>do today, and the story that it's becoming human like

0:07:42.800 --> 0:07:45.280
<v Speaker 3>is it over sells? In other words, it's hype and

0:07:45.320 --> 0:07:50.360
<v Speaker 3>that gap between what's real and what's plausible from the

0:07:50.440 --> 0:07:54.720
<v Speaker 3>stories is bad. And it's with mismanaged expectations that's when

0:07:54.760 --> 0:07:59.000
<v Speaker 3>you have the downfall, the disappointment. The disillusion meant also

0:07:59.400 --> 0:08:02.160
<v Speaker 3>in the more extreme case called an AI winter, and

0:08:02.480 --> 0:08:04.480
<v Speaker 3>the problem there is you throw the baby out with

0:08:04.520 --> 0:08:07.960
<v Speaker 3>the bathwater. You throw the value of generative AI for

0:08:08.040 --> 0:08:11.600
<v Speaker 3>first drafts and predictive AI, which by the way, is

0:08:11.640 --> 0:08:15.640
<v Speaker 3>still a much bigger industry right now, out with the bathwater.

0:08:16.360 --> 0:08:16.560
<v Speaker 1>You know.

0:08:16.640 --> 0:08:18.640
<v Speaker 2>Our producer Paul Brennan said, Yep, you're gonna want to

0:08:18.640 --> 0:08:20.120
<v Speaker 2>talk to this guy for a long time. We have

0:08:20.200 --> 0:08:22.600
<v Speaker 2>unfortunately run out of time, so promise you will come

0:08:22.600 --> 0:08:24.680
<v Speaker 2>back soon because I feel like this is a conversation

0:08:24.800 --> 0:08:28.680
<v Speaker 2>we need to continue. Just making so much sense, Eric,

0:08:28.720 --> 0:08:31.400
<v Speaker 2>thank you so much. Eric Siegel, he's a consultant, former

0:08:31.400 --> 0:08:33.760
<v Speaker 2>Columbia University professor. He's got a new book at the

0:08:33.760 --> 0:08:38.360
<v Speaker 2>AI Playbook, Mastering the rare art of Machine Learning Deployment.