'Hard to Predict' Where AI Will Be in 6 Months

Fortune

Lili Cheng, Corporate Vice President of Microsoft AI and Research, chats with Fortune about building vs. buying AI systems, AI copilots, and the future of the advancing technology.

Transcript

00:00Lily, I think what everyone is excited to hear about is you being on the AI front lines

00:05at Microsoft.

00:06So I want to start by talking about the big trends that you're seeing in AI models.

00:10So what new capabilities in AI systems are coming soon that all of the people here, CIOs,

00:16top tech people should really know about?

00:18Well, it was great actually talking before, you know, being up here.

00:22I think everybody's just kind of immersed in AI and how quickly things are moving.

00:28And so, you know, in a way it's hard to predict what's going to happen in the next six months

00:33even.

00:34I mean, we see a lot of trends that you've probably seen over the last week or two, you

00:38know, more types of conversation and media coming in, more asynchronous conversation

00:44and the ability to do more complex tasks.

00:49I think so far we've seen a lot of progress in some of our co-pilots at Microsoft, like

00:56the GitHub co-pilot.

00:57We've seen developers, I think today over 50% of all code in GitHub is AI generated,

01:04which is kind of mind-blowing for a product.

01:06It was one of the first co-pilots.

01:09And we see the developers, probably 90% of developers, find that they can do their repetitive

01:15tasks much more easily.

01:17But still maybe 60% or 70% of what a developer does isn't coding.

01:22It's really around collaboration and coordinating work, doing deeper thought.

01:28And so I think one of the things that you're going to see AI looking at is how do we address

01:32some of the real challenges that people have around getting work done, more of the complex

01:37tasks.

01:38And we'll continue to see progress, I think, in the, you know, more tactical tasks.

01:43But I think you're going to see more workflows and workspaces, people collaborating and some

01:48of these challenges being addressed.

01:51What are some of the biggest challenges you have faced and that you're trying to solve

01:54for at Microsoft?

01:57It's interesting, like in a way, even at Microsoft, I think AI cuts across so many different groups.

02:03So it's causing people to work together in a really different way.

02:08So our whole company, I feel like, has just collapsed into, you know, a group of people

02:13who have to take on more disciplines.

02:16So the difference between like what a product manager does and a developer does, a machine

02:22learning expert, a designer, somebody working with customers, we're just seeing a lot of

02:28things merging together.

02:29And I think that's partly because of the way AI is so general purpose and it's merging

02:35like workflows, things that were vertically for, you see things wanting to combine.

02:42So an example is we have a product that does sales, we have a product that does service

02:46and customers are saying, well, sometimes my happy customers want to use a customer

02:51service product, but then they want to buy something with the sales product.

02:55And so how can we, how can I, as a, you know, somebody looking at technologies, make sure

03:01that the systems interoperate a lot more because I don't want to have one system for this and

03:05then a whole nother system for that.

03:08Just the way companies are doing business, I think, is really being disrupted.

03:12But even the way we're developing things at Microsoft, I can feel that myself, like who

03:16should I hire?

03:17How come this team, which felt way over here before is now, you know, right here and how

03:22do we kind of move quickly and get everybody to coordinate better?

03:26Yeah.

03:27I think one of the themes that I've been hearing just from talking with people in this room

03:30earlier today is that their companies are still not sure whether they want to train

03:34the AI models from scratch, whether they want to build or buy.

03:37What advice are you giving people these days?

03:40Yeah.

03:41So I run a product which is called Copilot Studio.

03:43So that lets people create their own copilots.

03:46And what we're advising companies is maybe start with a tool that is more of a SaaS service

03:52that you can get up and running really quickly and try with your customers.

03:57And then as you need to extend that, you can go to building your own systems.

04:02One of the reasons I think we're advising companies to do that is, you know, it's hard

04:07to hire AI talent.

04:09And a lot of times by the time you build something, the system has changed and then you're kind

04:13of reworking it.

04:15So I think that would probably be my advice is, unless you have a really big AI team,

04:21I would probably start with some of the more pre-built services and then customize that

04:26as you need, because you'll get, you know, as we update things, you'll just get that

04:30for free.

04:31Yeah.

04:32Now, we talked a lot also about company culture and branding and trust.

04:37Trust being a big thing, company culture being huge, especially as you're saying, people

04:40talking about skilling, hiring, upskilling.

04:43How are you advising people on how to keep brand and company culture intact while shifting

04:48to automation, both from an employee perspective and a consumer experience?

04:54How are you building AI in a way that builds trust rather than breaks it down?

04:58I think there's kind of a couple parts to that.

05:00I mean, first of all, security.

05:03I mean, someone here, we were talking about security.

05:07Security and AI, I mean, there's so much happening, I think, on the security front.

05:13And so you definitely want to make sure that your systems are secure.

05:17That's kind of the foundation of trust, I think.

05:19And then there's, you know, just, you know, there's kind of brand honesty, which might

05:26kind of be a different variation of trust, you know?

05:31So I think that with security, even with all our focus on AI, that would definitely be

05:42our company's number one focus, just because we see so many bad actors using AI as well

05:49as good actors.

05:50So just a lot of focus there.

05:52And again, I think that's kind of why I would recommend people, you know, you don't want

05:56to build all of the systems yourself, because there's just a lot of work that goes into

05:59managing secure systems.

06:02And then as far as, you know, kind of maybe on the other side, one of the things that

06:06we see a lot of people doing is wanting to have an agent or an assistant as your front

06:11door to your company.

06:12So, you know, in the old days, you had your website, or maybe the front door of your company,

06:18which I don't think people really think of that as anymore.

06:21So you might have, you know, an assistant that's out there helping people ask questions

06:28or answer questions, and just what is the persona of that?

06:31How does it build trust?

06:33How does it keep, you know, data secure and things like that?

06:38You know, how are those experiences engaging and meaningful to people?

06:42So there's just a lot of different aspects from low level security to privacy and responsible

06:49AI to just the way your system's interacting.

06:52You mentioned bad actors, and I want to talk about that because I think that's a lot, there's

06:57a lot of scariness out there.

06:58What are some of the biggest examples of bad actors or any pitfalls you really see for

07:03relying too heavily on AI?

07:05I mean, there's probably a couple different levels.

07:11I mean, probably the thing that you hear a lot about with consumer systems is just, you

07:16know, what are you doing with my data?

07:17That's the thing that we hear the most from companies, like, are you leaking my data out

07:22to a model?

07:25Are the things that my customers say, does that go into making a model smarter?

07:29So like at Microsoft, one of the first things that we did is we took the open AI model and

07:33we definitely made it so that companies could upload their own proprietary information and

07:39their customer data and keep that separate from the public models.

07:43And that combination, I think, has been really great for people and really critical.

07:49For us, like, you know, we manage companies' email data and things like that.

07:53So obviously, if we were looking at your data or using it, like, people wouldn't use our

07:57tools.

07:58So that's just been a really foundational piece of the work that we've been focusing

08:04on.

08:05So I love that you have guardrails around that.

08:07How do you advise companies that are trying to make these guardrails and these regulations

08:13and rules within how they look at AI?

08:16How are you thinking about that at Microsoft?

08:18Well, I think what we would do is, again, it kind of depends on what level of the service

08:23you do.

08:24So if you use the SaaS services, which is kind of what I was recommending earlier, a

08:28lot of those are kind of built into the system.

08:31So, you know, what we would allow an administrator maybe to control if you were using these services

08:37inside your company, like we're definitely, you know, letting admins control that.

08:43And then as you kind of go down the stack, you can customize that more yourself.

08:48And I know this dinner is about Gen AI, and we're going to talk about that.

08:51But what is what is next?

08:53What's up and coming?

08:54What's on your radar?

08:56I think the thing that I would really love to see are more systems like the system that

09:01we built with GitHub, which is, you know, where we're really seeing real productivity

09:07and job satisfaction coming from the people that use the system.

09:11So in GitHub, we're seeing that both from the programmers using the system, as well

09:15as their managers.

09:17And like I said before, for like easy tasks that you would just want to do more repetitive

09:21tasks that you want to do more easily and more sophisticated tasks, GitHub is kind of

09:26unique because it has a lot of public data and people have kind of put a lot of their

09:32work history into a system already.

09:34So you're not starting fresh.

09:36We're starting to see that also with other verticals like customer support or for a whole

09:42company.

09:43We're starting to see that where you have all your documents, your email, your meetings

09:47and things like that.

09:48How do you how do you really well, maybe let me take a step back, like I think the advantage

09:53of Gen AI is that it can deal with conversation data, very unstructured information that you

10:00couldn't do before.

10:01So, you know, I worked on conversational ad when you kind of had to manually do a lot

10:06of this.

10:07And so it's kind of a miracle in a way that it exists and it's just been a crazy last

10:12two years.

10:13But I think a benefit for a company is taking all that information that was very unstructured

10:17either to do a sales call or to just run a meeting in your company and be able to

10:22take to get more information out of that.

10:25So people can be more productive.

10:27And so we've just been focused on that.

10:29It's it's it's it's been great to see, but we're definitely just starting.

10:33Yeah, I know.

10:34And there's so much more to do, I feel like, and as Andrew said, it's going so fast.

10:38So thank you so much, Lily.

10:39And we will leave time for questions for Lily once we sit.

10:43So thank you.

Category

Transcript

Recommended