WEBVTT - The Cutting Edge of Software Development in the AI Era    

0:00:02.520 --> 0:00:07.000
<v Speaker 1>Bloomberg Audio Studios, Podcasts, radio News.

0:00:08.200 --> 0:00:12.240
<v Speaker 2>You're listening to Bloomberg BusinessWeek with Carol Masser and Tim

0:00:12.280 --> 0:00:16.360
<v Speaker 2>Steneveek on Bloomberg Radio. Remember last week it was all

0:00:16.360 --> 0:00:19.080
<v Speaker 2>the way Yes. Last week Carol Alphabet, the parent company

0:00:19.079 --> 0:00:21.400
<v Speaker 2>of Google, reported a surge and demand for its cloud

0:00:21.440 --> 0:00:24.280
<v Speaker 2>and AI services. It pleased investors, who sent it shares up,

0:00:24.280 --> 0:00:26.680
<v Speaker 2>even as the company said capex for the year will

0:00:26.720 --> 0:00:30.120
<v Speaker 2>be even higher than expected. The company's investing record amounts

0:00:30.120 --> 0:00:32.920
<v Speaker 2>to try to push progress in AI and infuse answers

0:00:32.920 --> 0:00:37.879
<v Speaker 2>and assistance from its Lmgemini into its popular products, including search, and.

0:00:37.840 --> 0:00:40.200
<v Speaker 1>That's where Ryan J. Salva comes in. He is Senior

0:00:40.240 --> 0:00:43.280
<v Speaker 1>director of Product over at Google, where he builds AI

0:00:43.360 --> 0:00:47.800
<v Speaker 1>tools for developers, such as Gemini Cli. I think I'm

0:00:47.840 --> 0:00:50.480
<v Speaker 1>saying it correctly. We're talking about the command line interface.

0:00:50.760 --> 0:00:53.440
<v Speaker 1>It's an open source AI agent for developers, as well

0:00:53.479 --> 0:00:56.920
<v Speaker 1>as Gemini code Assist, Google's AI code assistant, Tim.

0:00:57.040 --> 0:00:59.040
<v Speaker 2>We've got Ryan Jay Salva with us. Also with us

0:00:59.080 --> 0:01:02.200
<v Speaker 2>Mandy Saying Bloomberg Intelligence, Global Head of Technology Research. He's

0:01:02.200 --> 0:01:05.279
<v Speaker 2>also those of the Tech Disruptors podcast. Ryan was featured

0:01:05.280 --> 0:01:07.440
<v Speaker 2>on an episode of the Tech Disruptors podcast that was

0:01:07.720 --> 0:01:11.440
<v Speaker 2>with Mandeep back in the spring. Welcome to both of you, Ryan,

0:01:11.480 --> 0:01:16.160
<v Speaker 2>our audience, some who code, probably more who don't. I'm

0:01:16.200 --> 0:01:18.920
<v Speaker 2>wondering though, if you can explain for everybody out there

0:01:19.280 --> 0:01:22.400
<v Speaker 2>how an AI assist in, including those from Google, how

0:01:22.440 --> 0:01:25.039
<v Speaker 2>they work right now with programmers, and the vision that

0:01:25.080 --> 0:01:26.520
<v Speaker 2>you have in the future.

0:01:27.600 --> 0:01:30.040
<v Speaker 3>Yeah. Absolutely, and first, thank you so much for having me.

0:01:30.480 --> 0:01:32.800
<v Speaker 3>You know, really, what we see today is that a

0:01:32.840 --> 0:01:36.320
<v Speaker 3>lot of developers are really caught in kind of the

0:01:36.880 --> 0:01:41.280
<v Speaker 3>labor of writing if then l's statements, getting caught up

0:01:41.360 --> 0:01:45.319
<v Speaker 3>in little tiny logical loops, and so often developers and

0:01:45.440 --> 0:01:49.880
<v Speaker 3>organizations are really just trying to deliver user requirements. They're

0:01:49.880 --> 0:01:52.320
<v Speaker 3>trying to deliver real value to their customers, and so

0:01:52.920 --> 0:01:56.040
<v Speaker 3>they're able to use AI in large language models to

0:01:56.440 --> 0:02:00.920
<v Speaker 3>write those requirements and natural language translate that code, and

0:02:01.000 --> 0:02:05.520
<v Speaker 3>through that ultimately accelerate their space of their pace of iteration,

0:02:05.640 --> 0:02:09.240
<v Speaker 3>their pace of learning, so that developers can focus more

0:02:09.280 --> 0:02:13.000
<v Speaker 3>on building features rather than on the syntax of the

0:02:13.040 --> 0:02:13.680
<v Speaker 3>code itself.

0:02:14.120 --> 0:02:18.760
<v Speaker 4>And so what kind of productivity benefits you think you've

0:02:18.840 --> 0:02:22.560
<v Speaker 4>seen both internally as well as with clients like maybe

0:02:22.600 --> 0:02:25.680
<v Speaker 4>talk us about one of the best use cases that

0:02:25.760 --> 0:02:27.359
<v Speaker 4>you've come across with Gemini.

0:02:28.080 --> 0:02:31.440
<v Speaker 3>Oh my gosh, there's so many, you know, So I'll

0:02:31.480 --> 0:02:35.200
<v Speaker 3>maybe first talk a little bit about from a metrics standpoint,

0:02:35.280 --> 0:02:37.959
<v Speaker 3>what we tend to see. So one of the teams

0:02:37.960 --> 0:02:43.160
<v Speaker 3>within Google is the Door Research Team. Dora effectively surveys

0:02:43.280 --> 0:02:46.840
<v Speaker 3>thousands and thousands of engineers every year, follows that up

0:02:46.880 --> 0:02:50.400
<v Speaker 3>with hundreds of hours of qualitative interviews. One of the

0:02:50.440 --> 0:02:53.079
<v Speaker 3>things that we're seeing is that today roughly ninety percent

0:02:53.120 --> 0:02:57.720
<v Speaker 3>of developers are integrating AI into their everyday work. They're

0:02:57.800 --> 0:03:02.760
<v Speaker 3>using AI for roughly two hours of so that tidal

0:03:02.800 --> 0:03:06.720
<v Speaker 3>wave of adoption has already swept over us all and

0:03:06.840 --> 0:03:11.320
<v Speaker 3>now we're swimming in the ocean of AI. At Google.

0:03:11.440 --> 0:03:15.000
<v Speaker 3>What we see is today roughly fifty percent of our

0:03:15.120 --> 0:03:18.239
<v Speaker 3>code is being written by AI. And I want you

0:03:18.280 --> 0:03:19.680
<v Speaker 3>to stop and maybe a.

0:03:19.720 --> 0:03:22.040
<v Speaker 1>Match second, say that one more time.

0:03:22.760 --> 0:03:27.040
<v Speaker 3>Five zero fifty percent of code is being written by AI.

0:03:27.720 --> 0:03:31.520
<v Speaker 3>That is a tremendous amount of code. And this is

0:03:31.720 --> 0:03:35.360
<v Speaker 3>in all of Google's products, from search to YouTube, to

0:03:35.640 --> 0:03:40.080
<v Speaker 3>cloud to you name it. And so this is allowing

0:03:40.160 --> 0:03:43.760
<v Speaker 3>our developers to really iterate again at a much much

0:03:43.800 --> 0:03:48.560
<v Speaker 3>faster pace to experiment, to learn, to test out new ideas,

0:03:48.640 --> 0:03:51.920
<v Speaker 3>and ultimately to be just a little bit less precious

0:03:52.080 --> 0:03:54.760
<v Speaker 3>about every line of code they write. Because they're able

0:03:54.800 --> 0:03:59.040
<v Speaker 3>to use the large language models to experiment, it's real

0:03:59.160 --> 0:04:01.680
<v Speaker 3>easy for them to to try out an idea on

0:04:01.720 --> 0:04:04.280
<v Speaker 3>a Tuesday, put it in front of a couple of

0:04:04.400 --> 0:04:07.440
<v Speaker 3>users on a Wednesday, and get a feel for whether

0:04:07.520 --> 0:04:11.600
<v Speaker 3>or not it provides real value. This is the real

0:04:11.880 --> 0:04:13.840
<v Speaker 3>magic and the real value that I feel, like AI

0:04:13.960 --> 0:04:14.920
<v Speaker 3>on Locks, I love that.

0:04:14.880 --> 0:04:18.239
<v Speaker 1>Idea less precious because it almost to me is akin

0:04:18.320 --> 0:04:21.640
<v Speaker 1>to when we got like digital cameras on our phone, right,

0:04:21.720 --> 0:04:24.480
<v Speaker 1>and we used to take pictures with film and everyone

0:04:24.600 --> 0:04:26.599
<v Speaker 1>Like I used to think about everyone, how many more

0:04:26.640 --> 0:04:27.440
<v Speaker 1>photos did I have?

0:04:27.560 --> 0:04:27.800
<v Speaker 4>Left?

0:04:27.960 --> 0:04:29.960
<v Speaker 1>Now I don't even care, right, I just take a

0:04:30.000 --> 0:04:33.680
<v Speaker 1>million photos Ryan. I do wonder, though, if we're less precious,

0:04:34.320 --> 0:04:36.680
<v Speaker 1>we're more efficient, We're more productive, which is what I'm

0:04:36.760 --> 0:04:38.880
<v Speaker 1>kind of getting from this conversation. What does it mean

0:04:38.920 --> 0:04:40.080
<v Speaker 1>for developer jobs?

0:04:41.120 --> 0:04:44.040
<v Speaker 3>Oh so, I mean, let me tell you this right now,

0:04:44.440 --> 0:04:47.800
<v Speaker 3>within my team, we are hiring more engineers, we are

0:04:47.880 --> 0:04:50.240
<v Speaker 3>hiring more product managers. And I see this when I

0:04:50.279 --> 0:04:54.400
<v Speaker 3>talk to so many other enterprises and organizations today. It's

0:04:54.440 --> 0:04:58.839
<v Speaker 3>not so much that the developer's job is any less important,

0:04:59.040 --> 0:05:01.640
<v Speaker 3>but what it does mean is that our job requirements

0:05:01.680 --> 0:05:03.960
<v Speaker 3>are changing. The skills that we need are a little

0:05:03.960 --> 0:05:06.880
<v Speaker 3>bit different. Because developers are spending a little bit less

0:05:06.920 --> 0:05:11.400
<v Speaker 3>time writing syntax, they're spending more time thinking about requirements.

0:05:11.920 --> 0:05:15.640
<v Speaker 3>We're really asking developers to think more like architects, to

0:05:15.800 --> 0:05:20.360
<v Speaker 3>think about systems design, to think about negotiating the contract

0:05:20.440 --> 0:05:25.920
<v Speaker 3>between components. And it means that ultimately, as our next

0:05:26.000 --> 0:05:30.760
<v Speaker 3>generation of creators and developers and builders are coming up,

0:05:31.160 --> 0:05:33.839
<v Speaker 3>we're asking them to think not just about can they

0:05:34.120 --> 0:05:38.320
<v Speaker 3>speak the language of programming, can they speak Java or

0:05:38.360 --> 0:05:42.520
<v Speaker 3>Python or c sharp, but rather can they do good

0:05:42.640 --> 0:05:46.599
<v Speaker 3>basic problem solving and can they think about large systems

0:05:46.680 --> 0:05:48.480
<v Speaker 3>level design. That's where the magic is at.

0:05:49.600 --> 0:05:52.400
<v Speaker 2>We're speaking with Ryan J. Salva, Senior director of Product

0:05:52.480 --> 0:05:56.520
<v Speaker 2>at Google. Ryan, you must remember that New York Times

0:05:56.640 --> 0:05:59.120
<v Speaker 2>article from August Goodbye one hundred and sixty five thousand

0:05:59.120 --> 0:06:01.000
<v Speaker 2>dollars tech job, as it went through all the entry

0:06:01.080 --> 0:06:04.360
<v Speaker 2>level tech job attentry that you're laughing, but the edgry

0:06:04.400 --> 0:06:06.320
<v Speaker 2>level tech jobs that were drying up and people work,

0:06:06.440 --> 0:06:09.640
<v Speaker 2>you know, compside graduates essentially working at Chipotle because they

0:06:09.640 --> 0:06:12.440
<v Speaker 2>couldn't find those entry level jobs. When you say you're

0:06:12.480 --> 0:06:15.839
<v Speaker 2>hiring engineers, are you hiring entry level engineers? Or is

0:06:16.080 --> 0:06:18.400
<v Speaker 2>entry level just dried up because of LMS.

0:06:19.080 --> 0:06:21.800
<v Speaker 3>Yeah. And by the way, I don't mean to laugh,

0:06:21.960 --> 0:06:25.679
<v Speaker 3>because every job is really important and I want folks

0:06:25.680 --> 0:06:27.960
<v Speaker 3>to be able to discover it. But I laughed because

0:06:27.960 --> 0:06:31.600
<v Speaker 3>I do think that the mem is sometimes the headlines

0:06:31.600 --> 0:06:34.479
<v Speaker 3>a little bit easier to grab attention than the ground

0:06:34.560 --> 0:06:38.000
<v Speaker 3>level reality. You know, I have, So that headline's wrong,

0:06:38.960 --> 0:06:39.440
<v Speaker 3>I'm sorry.

0:06:39.600 --> 0:06:41.239
<v Speaker 1>Is so that headline Ryan is wrong?

0:06:42.440 --> 0:06:44.600
<v Speaker 3>You know what? I think that I'm not saying that

0:06:44.680 --> 0:06:48.360
<v Speaker 3>an individual use case or an individual company doesn't go

0:06:48.440 --> 0:06:51.560
<v Speaker 3>through periods where they may let go of workers or

0:06:51.560 --> 0:06:53.880
<v Speaker 3>they may make different hiring decisions. But what I am

0:06:53.960 --> 0:06:58.320
<v Speaker 3>saying is that writ large across the industry, I'm still

0:06:58.360 --> 0:07:02.360
<v Speaker 3>seeing a very, very healthy the engineering ecosystem, and I'm

0:07:02.360 --> 0:07:06.160
<v Speaker 3>still seeing companies really prize and value the developers who

0:07:06.200 --> 0:07:10.920
<v Speaker 3>can come bringing skills that are more appropriate for this

0:07:10.960 --> 0:07:15.520
<v Speaker 3>new AI era. And that does mean less kind of again,

0:07:15.840 --> 0:07:19.040
<v Speaker 3>just being able to speak programming, to be able to

0:07:19.080 --> 0:07:23.520
<v Speaker 3>speak Java or JavaScript or typescript, is not enough anymore.

0:07:23.640 --> 0:07:25.680
<v Speaker 3>The developers really need to think about how they solve

0:07:25.720 --> 0:07:25.960
<v Speaker 3>the pro.

0:07:26.080 --> 0:07:28.760
<v Speaker 4>So, Ryan, one of the stats from Google that has

0:07:28.760 --> 0:07:32.560
<v Speaker 4>caught my attention is the increase, the exponential increase in

0:07:32.600 --> 0:07:35.840
<v Speaker 4>their token count, you know, to almost one point three

0:07:36.000 --> 0:07:39.920
<v Speaker 4>quantillion tokens. Where does coding assistance as well as that

0:07:40.000 --> 0:07:44.240
<v Speaker 4>number one point three contillion? So it's like that.

0:07:45.920 --> 0:07:46.840
<v Speaker 2>Is that what he's going to ask?

0:07:47.520 --> 0:07:50.880
<v Speaker 4>Look, I mean, these numbers are staggering, but when it

0:07:50.920 --> 0:07:54.800
<v Speaker 4>comes to use cases, I think there's a big variance

0:07:54.880 --> 0:07:58.120
<v Speaker 4>between you know, a simple chat bot Q and A

0:07:58.320 --> 0:08:02.040
<v Speaker 4>versus coding agent or and AI agent running for days.

0:08:02.440 --> 0:08:06.920
<v Speaker 4>How would you characterize the contribution of coding assistant and

0:08:07.000 --> 0:08:10.400
<v Speaker 4>the products that you oversee to the overall token consumption

0:08:10.480 --> 0:08:11.000
<v Speaker 4>at Google?

0:08:11.920 --> 0:08:15.600
<v Speaker 3>Sure? Sure? So, I mean I'll start here. We don't

0:08:15.640 --> 0:08:19.840
<v Speaker 3>necessarily count if a token is used for a Google

0:08:19.880 --> 0:08:24.960
<v Speaker 3>search versus software development problem versus someone doing their homework.

0:08:25.480 --> 0:08:28.000
<v Speaker 3>Having said that, what I can tell you is that

0:08:28.160 --> 0:08:32.040
<v Speaker 3>perhaps nowhere better than in software development have I seen

0:08:32.360 --> 0:08:35.880
<v Speaker 3>product market fit better between large language models and a

0:08:35.920 --> 0:08:39.440
<v Speaker 3>particular use case. There are a lot of reasons for this.

0:08:39.520 --> 0:08:43.600
<v Speaker 3>I think probably the biggest one is that large language models.

0:08:44.600 --> 0:08:47.800
<v Speaker 3>You know. You know this when you use Gemini or

0:08:47.920 --> 0:08:51.160
<v Speaker 3>use chat, GPT or any other kind of large language

0:08:51.200 --> 0:08:54.520
<v Speaker 3>model out there, if you're asking it to help you

0:08:54.559 --> 0:08:58.240
<v Speaker 3>write an email or help you write a document of

0:08:58.280 --> 0:09:02.360
<v Speaker 3>some kind. Often the response, the quality of the response

0:09:02.440 --> 0:09:05.720
<v Speaker 3>depends an awful lot upon your personal judgment and your

0:09:05.760 --> 0:09:11.040
<v Speaker 3>personal taste. Whereas with software development, we have decades of

0:09:12.200 --> 0:09:16.199
<v Speaker 3>deterministic quality measures that let us know whether the software

0:09:16.360 --> 0:09:19.880
<v Speaker 3>is good and safe and useful or not. We have

0:09:20.080 --> 0:09:23.840
<v Speaker 3>unit tests and static analysis and all these other ways

0:09:23.880 --> 0:09:27.440
<v Speaker 3>of validating the quality of software. And so what I

0:09:27.520 --> 0:09:32.560
<v Speaker 3>see is a lot of organizations using AI, using agents,

0:09:32.640 --> 0:09:35.839
<v Speaker 3>using large language models to accelerate their engineering life cycle

0:09:35.920 --> 0:09:39.600
<v Speaker 3>because they can deterministically say this is of good quality,

0:09:39.800 --> 0:09:42.160
<v Speaker 3>this is of bad quality, this is something I want

0:09:42.160 --> 0:09:44.200
<v Speaker 3>to use, this is something I don't. That's how I

0:09:44.240 --> 0:09:48.199
<v Speaker 3>see it really accelerating, particularly in the software development space.

0:09:48.400 --> 0:09:53.560
<v Speaker 4>So do you expect a big migration of legacy systems

0:09:53.640 --> 0:09:57.360
<v Speaker 4>to the modern architecture that you mentioned as a result

0:09:57.520 --> 0:10:00.720
<v Speaker 4>of you know, coding agents being that good, or do

0:10:00.760 --> 0:10:03.640
<v Speaker 4>you see limitations in terms of you know where the

0:10:04.080 --> 0:10:07.600
<v Speaker 4>practical use cases are versus you know where the legacy

0:10:07.640 --> 0:10:09.520
<v Speaker 4>technologies are just too hard to move.

0:10:10.280 --> 0:10:13.679
<v Speaker 3>Yeah, you know. Actually, migration and modernization is one of

0:10:13.720 --> 0:10:17.000
<v Speaker 3>the areas where I see the most interest among large

0:10:17.040 --> 0:10:20.240
<v Speaker 3>engineering teams today. There are a lot of reasons for that.

0:10:20.360 --> 0:10:23.560
<v Speaker 3>In some cases, the engineers who are maintaining those legacy

0:10:23.760 --> 0:10:29.240
<v Speaker 3>kind of applications are retiring or moving on. Skill sets

0:10:29.280 --> 0:10:32.360
<v Speaker 3>are atrophying, and there is a thing within software development

0:10:32.440 --> 0:10:36.960
<v Speaker 3>called code rot effectively when an application just sits around

0:10:37.000 --> 0:10:41.400
<v Speaker 3>so long that it atrophies over time and becomes less performed.

0:10:41.400 --> 0:10:44.320
<v Speaker 4>So AI is good in that without consuming too many

0:10:44.360 --> 0:10:46.280
<v Speaker 4>tokens or you know, increasing your bill.

0:10:47.080 --> 0:10:50.600
<v Speaker 3>So what I actually hear is a lot of organizations

0:10:50.679 --> 0:10:55.360
<v Speaker 3>are willing to dedicate waves and waves and waves of

0:10:55.360 --> 0:10:58.800
<v Speaker 3>tokens because the cost of maintaining those legacy applications is

0:10:58.880 --> 0:11:03.120
<v Speaker 3>so high. Often they're having to maintain entire data centers,

0:11:03.440 --> 0:11:05.680
<v Speaker 3>which means that you're paying not only the cost of

0:11:05.920 --> 0:11:08.800
<v Speaker 3>the engineers to maintain them, but you're also paying for

0:11:08.840 --> 0:11:13.280
<v Speaker 3>the facilities, for the hardware, for all of the extra

0:11:14.120 --> 0:11:18.080
<v Speaker 3>it that goes with maintaining those And honestly, if you

0:11:18.200 --> 0:11:22.480
<v Speaker 3>even just take the cost of maintaining them to the side,

0:11:22.760 --> 0:11:25.240
<v Speaker 3>the fact that you're not able to carry those applications

0:11:25.280 --> 0:11:27.640
<v Speaker 3>forward and innovate with them and do new things at them.

0:11:27.840 --> 0:11:29.359
<v Speaker 3>Often that's the real cost.

0:11:29.200 --> 0:11:31.480
<v Speaker 1>Ryan come back. We'd love to continue this. Ryan J.

0:11:31.640 --> 0:11:35.200
<v Speaker 1>Salva over at Director or senior director of product at Google,

0:11:35.240 --> 0:11:38.199
<v Speaker 1>and of course our own man Deep seeing a Bloomberg Intelligence.

0:11:37.760 --> 0:11:39.440
<v Speaker 2>Met For more insights from mand Deep in the Bloomberg

0:11:39.480 --> 0:11:42.440
<v Speaker 2>Intelligence team, check out the Tech Disruptors podcast. You can

0:11:42.480 --> 0:11:45.480
<v Speaker 2>find it on Apple, Spotify, or wherever you get your podcasts.