The AI opportunity: Sequoia Capital's AI Ascent 2024 opening remarks
00:00:00.000 | >> My name is Pat Grady.
00:00:03.120 | I'm one of the members of Team Sequoia.
00:00:05.480 | I'm here with my partners,
00:00:06.880 | Sonia and Konstantin,
00:00:09.480 | who will be your MCs for the day.
00:00:11.480 | Along with all of our partners at Sequoia,
00:00:13.840 | we would like to welcome you to AI Ascent.
00:00:17.800 | There's a lot going on in the world of AI.
00:00:22.220 | We have an objective to
00:00:23.880 | learn a few things while we're here today.
00:00:25.680 | We have an objective to meet a few people who can be
00:00:28.440 | helpful in our journey while we're here
00:00:30.000 | today and hopefully, we'll have a little bit of fun.
00:00:32.400 | So just to frame the opportunity, what is it?
00:00:36.160 | Well, a year ago,
00:00:38.680 | it felt like this magic box
00:00:40.800 | that could do wonderful, amazing things.
00:00:43.320 | I think over the last 12 months,
00:00:44.740 | we've been through this contracted form of the hype cycle.
00:00:47.640 | We had the peak of inflated expectations,
00:00:49.800 | we have the trough of disillusionment,
00:00:51.320 | we're crawling back out into the plateau of productivity.
00:00:53.920 | I think we've realized that what LLMs,
00:00:56.200 | what AI really brings to us today are
00:00:57.720 | three distinct capabilities that can
00:01:00.240 | be woven into a wide variety of magical applications.
00:01:02.920 | The first is the ability to create,
00:01:05.280 | hence the name generative AI.
00:01:06.680 | You can create images, you can create text,
00:01:08.260 | you can create video, you can create audio,
00:01:09.840 | you can create all sorts of things.
00:01:11.240 | Not something software has been able to do before.
00:01:13.440 | So that's pretty cool.
00:01:14.720 | The second is the ability to reason,
00:01:17.520 | could be one-shot, could be multi-step agentic type reasoning.
00:01:20.800 | But again, not something software has been able to do before.
00:01:24.340 | Because it can create,
00:01:26.460 | because it can reason,
00:01:28.100 | we've got the right brain and the left brain covered,
00:01:30.740 | which means that software can also for the first time,
00:01:34.660 | interact in a human-like capacity.
00:01:38.180 | This is huge because this has
00:01:39.780 | profound business model implications
00:01:41.620 | that we're going to mention on the next slide.
00:01:43.860 | So what? A lot of times we try to
00:01:47.740 | reason by analogy when we see something new.
00:01:49.940 | In this case, the best analogy that we can come up with,
00:01:52.740 | which is imperfect for a million reasons,
00:01:54.820 | but still useful, is the Cloud transition.
00:01:57.500 | Over the last 20 years or so,
00:01:59.140 | that was a major tectonic shift in
00:02:00.700 | the technology landscape that led to new business models,
00:02:03.100 | new applications, new ways for people to interact with technology.
00:02:06.940 | If we go back to some of the early days of that Cloud transition,
00:02:10.060 | this is circa about 2010,
00:02:12.260 | the entire pie, the entire global TAM for software is about 350 billion,
00:02:16.800 | of which this tiny slice,
00:02:18.720 | just $6 billion is Cloud software.
00:02:20.900 | Fast forward to last year,
00:02:22.940 | the TAM has grown from about 350 to 650,
00:02:25.580 | but that slice has become 400 billion of revenue.
00:02:29.300 | That's a 40 percent CAGR over 15 years.
00:02:32.420 | That's massive growth.
00:02:34.300 | Now, if we're going to reason by analogy,
00:02:37.980 | Cloud was replacing software with software.
00:02:41.740 | Because of what I mentioned about the ability to
00:02:43.900 | interact in a human-like capability,
00:02:46.580 | one of the big opportunities for AI is to replace services with software.
00:02:52.380 | If that's the TAM that we're going after,
00:02:54.380 | the starting point is not hundreds of billions,
00:02:56.420 | the starting point is possibly tens of trillions.
00:03:00.220 | So you can really dream about what this has a chance to become.
00:03:05.820 | We would posit, and this is a hypothesis,
00:03:08.420 | as everything we say today will be,
00:03:10.280 | we would posit that we are standing at the precipice of
00:03:12.860 | the single greatest value creation opportunity mankind has ever known.
00:03:16.900 | Why now? One of the benefits of being part of Sequoia is that we have
00:03:23.300 | this long history and we've gotten to study
00:03:26.100 | the different waves of technology and understand how they
00:03:28.620 | interact and understand how they lead us to the present moment.
00:03:31.300 | We're going to take a quick trip down memory lane.
00:03:33.300 | So 1960s, our partner,
00:03:35.940 | Don Valentine, who founded Sequoia,
00:03:37.580 | was actually the guy who ran the go-to-market for Fairchild Semiconductor,
00:03:40.980 | which gave Silicon Valley its name with Silicon-based transistors.
00:03:44.140 | We got to see that happen.
00:03:45.500 | We got to see the 1970s when systems were built on top of those chips.
00:03:49.940 | We got to see the 1980s when they were connected up by
00:03:53.300 | networks with PCs as the endpoint and the advent of package software.
00:03:57.620 | We got to see the 1990s when those networks went public-facing in the form of the Internet,
00:04:02.020 | changed the way we communicate,
00:04:03.340 | changed the way we consume.
00:04:04.700 | We got to see the 2000s when the Internet matured to
00:04:07.820 | the point where it could support sophisticated applications,
00:04:10.380 | which became known as the Cloud.
00:04:12.260 | We got to see the 2010s where all those apps showed up in
00:04:15.500 | our pocket in the form of mobile devices and changed the way we work.
00:04:19.700 | So why do we bother going through this little build?
00:04:23.220 | Well, the point here is that each one of these waves is
00:04:25.260 | additive with what came before.
00:04:28.180 | The idea of AI is nothing new.
00:04:31.100 | It dates back to the 1940s.
00:04:32.860 | I think neural nets first became an idea in the 1940s.
00:04:36.180 | But the ingredients required to take AI from idea,
00:04:40.900 | from dream, into production,
00:04:43.780 | into reality, to actually solve real-world problems in
00:04:47.620 | a unique and compelling way that you can build a durable business around.
00:04:52.020 | The ingredients required to do that did not exist until the past couple of years.
00:04:56.500 | We finally have compute that is cheap and plentiful.
00:04:59.980 | We have networks that are fast and efficient and reliable.
00:05:03.500 | Seven of the eight billion people on
00:05:05.100 | the planet have a supercomputer in their pockets.
00:05:07.380 | Thanks in part to COVID,
00:05:08.900 | everything has been forced online,
00:05:10.780 | and the data required to fuel
00:05:12.900 | all of these delightful experiences is readily available.
00:05:15.700 | So now is the moment for AI to
00:05:19.380 | become the theme of the next 10, probably 20 years.
00:05:23.900 | So we have as strong conviction as you could possibly
00:05:27.740 | have in a hypothesis that is not yet proven,
00:05:30.620 | that the next couple of decades are going to be the time of AI.
00:05:35.020 | What shape would that opportunity take?
00:05:38.180 | Again, we're going to analogize to
00:05:40.020 | the Cloud transition and the mobile transition.
00:05:41.820 | These logos on the left side of the page,
00:05:44.380 | those are most of the companies born as a result of
00:05:46.860 | those transitions that got to a billion dollars plus of revenue.
00:05:49.740 | The list is not exhaustive,
00:05:51.140 | but this is probably 80 percent or so of the companies
00:05:54.140 | formed in those transitions that got to a billion plus of revenue,
00:05:57.300 | not valuation, revenue.
00:05:59.540 | The most interesting thing about this slide is the right side.
00:06:03.100 | It's not what's there, it's what isn't there.
00:06:07.380 | The landscape is wide open.
00:06:10.380 | The opportunity set is massive.
00:06:14.820 | We think if we were standing here 10 or 15 years from today,
00:06:19.660 | that right side is going to have 40 or 50 logos in it.
00:06:23.660 | Chances are, it's going to be a bunch of
00:06:25.620 | the logos of companies that are in this room.
00:06:28.260 | This is the opportunity,
00:06:29.860 | this is why we're excited.
00:06:31.700 | With that, I will hand it off to Sonia.
00:06:34.740 | >> Thanks Pat. Wow, what a year.
00:06:43.940 | ChatGPT came out a year and a half ago.
00:06:46.980 | I think it's been a whirlwind for everybody here.
00:06:49.340 | It probably feels like just about all of us have been going
00:06:51.860 | non-stop with the ground shifting under our feet constantly.
00:06:55.220 | So let's take a pause, zoom out,
00:06:57.060 | and take stock on what's happened so far.
00:06:59.580 | Last year, we were talking about how AI was going to
00:07:02.220 | revolutionize all these different fields
00:07:04.140 | and provide amazing productivity gains.
00:07:06.020 | A year later, it's starting to come into focus.
00:07:09.220 | Who here has seen this tweet from Sebastian at Klarna?
00:07:12.780 | Show of hands. It's pretty incredible.
00:07:16.740 | Klarna is now using OpenAI to
00:07:18.380 | handle two-thirds of customer service inquiries.
00:07:20.980 | They've automated the equivalent of 700 full-time agents jobs.
00:07:25.460 | We think there are tens of millions of call center agents
00:07:28.100 | globally and one of the most exciting areas
00:07:30.500 | where we've already seen AI find
00:07:31.900 | product market fit is in this customer support markets.
00:07:35.740 | Legal services. A year ago,
00:07:38.340 | the law was considered one of the least tech forward industries,
00:07:41.460 | one of the least likely to take risks.
00:07:43.980 | Now, we have companies like Harvey that are automating away
00:07:46.740 | a lot of the work that lawyers do from day-to-day grunt work and
00:07:50.140 | drudgery all the way to more advanced analysis.
00:07:53.340 | Or software engineering. I'm sure a bunch of people in
00:07:55.980 | this room have seen some of
00:07:56.900 | the demos floating around on Twitter recently.
00:07:59.660 | It's remarkable that we've gone from a year ago,
00:08:02.220 | AI theoretically writing our code to
00:08:05.420 | entirely self-contained AI software engineers.
00:08:08.900 | I think it's really exciting. The future's
00:08:10.420 | going to have a lot more software.
00:08:12.420 | AI isn't all about revolutionizing work.
00:08:16.260 | It's already increasing our quality of life.
00:08:18.060 | Now, the other day, I was in a Zoom with Pat,
00:08:20.580 | and I noticed that he looked a little bit suspicious,
00:08:23.660 | didn't speak the entire time.
00:08:25.900 | Having reflected on it more,
00:08:27.460 | I'm pretty sure that he actually sent in
00:08:29.500 | his virtual AI avatar and was actually hitting the gym,
00:08:32.740 | which would explain a lot.
00:08:34.260 | >> Hi, this is Pat Grady. This is definitely me.
00:08:37.020 | I'm definitely here and not at the gym right now.
00:08:40.220 | >> It even gets the facial scrunches.
00:08:44.660 | >> This is courtesy of Haygen.
00:08:46.580 | It's pretty amazing.
00:08:48.620 | This is how far technology has come in a year.
00:08:51.300 | It's scary and exciting
00:08:55.100 | to think about how this all plays out in the coming decade.
00:08:57.940 | All kidding aside, two years ago,
00:09:01.500 | when we thought that Generative AI might usher
00:09:03.820 | in the next great technology shift,
00:09:06.260 | we didn't know what to expect.
00:09:08.140 | Would real companies come out of it?
00:09:10.100 | Would real revenue materialize?
00:09:12.460 | I think the sheer scale of user pull and
00:09:15.060 | revenue momentum has surprised just about everybody.
00:09:18.620 | Generative AI, we think,
00:09:20.260 | is now clocking in around $3 billion of revenues in aggregate,
00:09:23.780 | and that's before you count all the incremental revenue
00:09:26.180 | generated by the FANG companies and the Cloud providers in AI.
00:09:29.660 | To put three billion in context,
00:09:31.740 | it took the SaaS market nearly a decade
00:09:34.220 | to reach that level of revenue.
00:09:36.020 | Generative AI got there its first year out the gate.
00:09:38.780 | The rate and the magnitude of
00:09:40.460 | the sea change make it very clear to us
00:09:42.140 | that Generative AI is here to stay.
00:09:44.740 | The customer pull in AI isn't restricted to one or two apps.
00:09:49.580 | It's everywhere. I'm sure everyone's aware of
00:09:51.980 | how many users ChatGPT has.
00:09:54.020 | But when you look at the revenue and the usage numbers,
00:09:56.620 | for a lot of AI apps,
00:09:58.060 | both consumer companies and enterprise companies,
00:10:00.820 | startups, and incumbents,
00:10:02.940 | many AI products are actually striking a chord with
00:10:05.660 | customers and starting to find
00:10:07.060 | product market fit across industries.
00:10:09.140 | We find the diversity of use cases that are starting to
00:10:11.580 | hit really exciting.
00:10:14.020 | The number one thing that has surprised me,
00:10:17.020 | at least, about the funding environment over
00:10:19.540 | the last year has been how
00:10:20.700 | uneven the share of funding has been.
00:10:22.780 | If you think of Generative AI as
00:10:24.740 | a layer cake where you have foundation models on the bottom,
00:10:27.820 | you have developer tools and infra above,
00:10:29.860 | and then you have applications on top.
00:10:31.740 | A year ago, we'd expected
00:10:33.300 | that there would be a Cambrian explosion in
00:10:35.500 | the application layer due to
00:10:37.060 | the new enabling technology in the foundation layer.
00:10:39.540 | Instead, we've actually found that
00:10:41.340 | new company formation and capital
00:10:43.060 | has formed in an inverse pattern.
00:10:44.980 | More and more foundation models are popping
00:10:47.220 | up and raising very large funding rounds,
00:10:49.660 | while the application layer feels
00:10:51.060 | like it is just getting going.
00:10:53.220 | Our partner, David, is right here,
00:10:55.780 | and posed a thought-provoking question last year with
00:10:58.540 | his article, AI's $200 billion question.
00:11:01.860 | If you look at the amount of money
00:11:05.300 | that companies are pouring into GPUs right now,
00:11:08.300 | we spent about $50 billion on NVIDIA GPUs just last year.
00:11:12.940 | Everybody's assuming if you build it, they will come.
00:11:15.820 | AI is a field of dreams.
00:11:17.660 | But so far, remember on the previous slide,
00:11:19.700 | we've identified about $3 billion or so of
00:11:22.060 | AI revenue plus change from the Cloud providers.
00:11:25.100 | We've put 50 billion into the ground,
00:11:26.940 | plus energy, plus data center costs and more.
00:11:29.500 | We've gotten three out.
00:11:31.220 | To me, that means the math isn't mathing yet.
00:11:34.380 | The amount of money it takes to build
00:11:36.540 | this stuff has vastly exceeded the amount of money
00:11:38.820 | coming out so far.
00:11:40.140 | So we've got some real problems to fix still.
00:11:43.900 | And even though the usage and--
00:11:46.780 | even though the revenue and the user numbers in AI
00:11:48.740 | look incredible, the usage data says
00:11:50.740 | that we're still really early.
00:11:52.420 | And so if you look at, for example, the ratio of daily
00:11:54.820 | to monthly active users, or if you look at one month
00:11:57.660 | retention, generative AI apps are still
00:12:00.380 | falling far short of their mobile peers.
00:12:03.580 | To me, that is both a problem and an opportunity.
00:12:06.100 | It's an opportunity because AI right now
00:12:08.540 | is a once a week, once a month kind of tinkery phenomenon
00:12:12.420 | for the most part for people.
00:12:14.260 | But we have the opportunity to use AI to create apps
00:12:16.620 | that people want to use every single day of their lives.
00:12:20.180 | When we interview users, one of the biggest reasons
00:12:22.780 | they don't stick on AI apps is the gap between expectations
00:12:26.540 | and reality.
00:12:27.780 | So that magical Twitter demo becomes a disappointment
00:12:30.940 | when you see that the model just isn't smart enough
00:12:33.080 | to reliably do the thing that you asked it to do.
00:12:36.200 | The good thing is with that $50 billion
00:12:37.820 | plus of GPU spend last year, we now
00:12:40.520 | have smarter and smarter base models to build on.
00:12:42.960 | And just in the last month, we've seen Sora.
00:12:44.800 | We've seen Cloud 3.
00:12:45.840 | We saw Grok over the weekend.
00:12:47.760 | And so as the level of intelligence of the baseline
00:12:49.920 | rises, we should expect AI's product market fit
00:12:52.320 | to accelerate.
00:12:53.320 | So unlike in some markets where the future of the market
00:12:55.640 | is very unclear, the good thing about AI
00:12:57.840 | is you can draw a very clear line to how those apps will
00:13:00.160 | get predictably better and better.
00:13:03.720 | Let's remember that success takes time.
00:13:05.480 | We said this at last year's AI Ascent, and we'll say it again.
00:13:08.600 | If you look at the iPhone, some of the first apps in the V1
00:13:12.800 | of the App Store were the beer drinking app or the lightsaber
00:13:15.640 | app or the flip cup app or the flashlight--
00:13:19.760 | kind of the fun, lightweight demonstrations
00:13:22.040 | of new technology.
00:13:23.480 | Those eventually became either native apps--
00:13:25.920 | AKA the flashlight, et cetera--
00:13:27.960 | or utilities and gimmicks.
00:13:30.060 | The iPhone came out in 2007.
00:13:31.920 | The App Store came out in 2008.
00:13:34.120 | It wasn't until 2010 that you saw Instagram and DoorDash 2013.
00:13:39.360 | So it took time for companies to discover and harness
00:13:42.720 | the net new capabilities of the iPhone in creative ways
00:13:45.160 | that we couldn't just imagine yet.
00:13:47.080 | We think the same thing is playing out in AI.
00:13:50.880 | We think we're already seeing a peek
00:13:52.520 | into what some of those next legendary companies might be.
00:13:55.840 | Here are a few of the ones that have captured our attention
00:13:58.880 | recently, but I think it's much broader than the set of use
00:14:01.900 | cases on this page.
00:14:03.340 | As I mentioned, we think customer support
00:14:05.060 | is one of the first handful of use cases that's really hitting
00:14:07.340 | product market fit in the enterprise.
00:14:09.260 | As I mentioned with the Klarna story,
00:14:10.820 | I don't think that's an exception.
00:14:12.260 | It's the rule.
00:14:12.820 | I think that is the rule.
00:14:14.220 | AI Friendship has been one of the most surprising
00:14:17.060 | applications for many of us.
00:14:18.340 | I think it took a few months of thinking for us
00:14:20.620 | to wrap our heads around.
00:14:22.980 | But I think the user and the usage metrics in this category
00:14:26.100 | imply very strong user love.
00:14:30.160 | And then Horizontal Enterprise Knowledge.
00:14:32.620 | We'll hear more from Glean and Dusk later today.
00:14:35.620 | We think that enterprise knowledge is finally
00:14:37.460 | starting to become unlocked.
00:14:40.500 | So here are some predictions for what
00:14:42.000 | we'll see over the coming year.
00:14:43.500 | Prediction number one, 2024 is the year
00:14:46.920 | that we see real applications take us
00:14:48.900 | from co-pilots that are kind of helpers on the side
00:14:51.920 | and suggest things to you and help you,
00:14:54.180 | to agents that can actually take the human out
00:14:56.540 | of the loop entirely.
00:14:58.180 | AI that feels more like a co-worker than a tool.
00:15:01.140 | We're seeing this start to work in domains like software
00:15:03.780 | engineering, customer service.
00:15:06.040 | And we'll hear more about this topic today.
00:15:07.820 | I think both Andrew Ng and Harrison Chase
00:15:09.900 | are planning to speak on it.
00:15:12.540 | Prediction number two, one of the biggest knocks against LLMs
00:15:15.980 | is that they seem to be parroting
00:15:17.420 | the statistical patterns in text and aren't actually
00:15:20.180 | taking the time to reason and plan through the tasks
00:15:22.300 | at hand.
00:15:23.460 | That's starting to change with a lot of new research,
00:15:26.300 | like inference time compute and gameplay-style value
00:15:29.140 | iteration.
00:15:30.140 | What happens when you give the model the time
00:15:32.180 | to actually think through what to do?
00:15:34.380 | We think that this is a major research
00:15:36.940 | thrust for many of the foundation model companies.
00:15:39.460 | And we expect it to result in AI that's
00:15:41.540 | more capable of higher-level cognitive tasks like planning
00:15:45.580 | and reasoning over the next year.
00:15:47.300 | And we'll hear more about this later today
00:15:49.060 | from Noam Brown of OpenAI.
00:15:52.820 | Prediction number three, we are seeing an evolution
00:15:55.760 | from fun consumer apps or prosumer apps,
00:15:58.860 | where you don't really care if the AI says something
00:16:02.180 | wrong or crazy occasionally, to real enterprise applications,
00:16:06.720 | where the stakes are really high,
00:16:08.260 | like hospitals and defense.
00:16:09.900 | The good thing is that there's different tools
00:16:11.860 | and techniques emerging to help bring these LLMs sometimes
00:16:15.220 | into the 5.9's reliability range,
00:16:17.060 | from RLHF to prompt training to vector databases.
00:16:19.660 | And I'm sure that's something that you guys can
00:16:21.540 | compare notes on later today.
00:16:22.740 | I think a lot of folks in this room
00:16:24.020 | are doing really interesting things
00:16:25.480 | to make LLMs more reliable in production.
00:16:28.860 | And finally, 2024 is the year that we
00:16:31.060 | expect to see a lot of AI prototypes and experiments
00:16:33.700 | go into production.
00:16:35.300 | And what happens when you do that?
00:16:36.740 | That means latency matters.
00:16:38.380 | That means cost matters.
00:16:39.640 | That means you care about model ownership.
00:16:41.420 | You care about data ownership.
00:16:43.300 | And it means we expect the balance of compute
00:16:45.180 | to begin shifting from pre-training over
00:16:47.220 | to inference.
00:16:48.500 | So 2024 is a big year.
00:16:50.020 | There's a lot of pressure and expectations
00:16:52.220 | built into some of these applications
00:16:53.740 | as they transition into production.
00:16:55.780 | And it's really important that we get it right.
00:16:58.980 | With that, I'll transition to Konstantin, who
00:17:00.940 | will help us dream about AI over an even longer time horizon.
00:17:04.580 | [APPLAUSE]
00:17:09.460 | Thank you, Sonia.
00:17:10.380 | And thank you, everyone, for being here today.
00:17:12.340 | Pat just set up the "so what?"
00:17:14.860 | Why is this so important?
00:17:16.020 | Why are we all in the room?
00:17:17.700 | And Sonia just walked us through the "what now?"
00:17:20.300 | Where are we in the state of AI?
00:17:22.260 | This section is going to be about what's next.
00:17:25.700 | We're going to take a step back and think
00:17:27.700 | through what this means in the broader
00:17:29.980 | concept of technology and society at large.
00:17:34.900 | So there are many types of technology revolution.
00:17:38.380 | There are communication revolutions, like telephony.
00:17:42.340 | There are transportation revolutions,
00:17:44.860 | like the locomotive.
00:17:46.740 | There are productivity revolutions,
00:17:48.940 | like the mechanization of food harvest.
00:17:52.900 | We believe that AI is primarily a productivity revolution.
00:17:58.340 | And these revolutions follow a pattern.
00:18:01.580 | It starts with a human with a tool.
00:18:04.020 | That transitions into a human with a machine assistant.
00:18:07.860 | And eventually, that moves into a human with a machine network.
00:18:12.460 | The two predictions that we're going to talk about
00:18:14.580 | both relate to this concept of humans
00:18:17.340 | working with machine networks.
00:18:19.300 | Let's look at a historical example.
00:18:21.300 | The sickle has been around as a tool for the human
00:18:23.700 | for over 10,000 years.
00:18:26.300 | The mechanical reaper, which is a human and a machine assistant,
00:18:29.340 | was invented in 1831, a single machine system
00:18:33.900 | being used by a human.
00:18:35.860 | Today, we live in an era where we have a combine harvester.
00:18:39.620 | The combine harvester is tens of thousands
00:18:42.540 | of machine systems working together as a complex network.
00:18:48.740 | We're starting to use language in AI to describe this.
00:18:51.460 | Language like individual machine participants in the system
00:18:54.220 | might be called an agent.
00:18:55.340 | We're talking about this quite a bit today.
00:18:57.540 | The way that topology and the way
00:18:58.980 | that the information is transferred
00:19:00.500 | between these agents, we're starting
00:19:01.980 | to talk about as reasoning, for example.
00:19:04.060 | In essence, we're creating very complicated layers
00:19:07.260 | of abstraction above the primitives of AI.
00:19:11.220 | I'll talk about two examples today,
00:19:12.660 | two examples that we're experiencing right in front
00:19:14.900 | of us in knowledge work.
00:19:16.540 | The first is software.
00:19:18.300 | So software started off as a very manual process.
00:19:21.140 | Here's Ada Lovelace, who wrote logical programming
00:19:24.580 | with pen and paper, was able to do these computations,
00:19:27.380 | but without the assistant of a machine.
00:19:30.480 | We've been living in an era where
00:19:31.820 | we have significant machine assistants for computation,
00:19:35.780 | not just the computer, but the integrated development
00:19:38.020 | environment, and increasingly more and more technologies
00:19:40.340 | to accelerate development of software.
00:19:42.900 | We're entering a new era in which
00:19:45.460 | these systems are working together
00:19:46.820 | in a complex machine network.
00:19:50.180 | What you see is a series of processes
00:19:53.420 | that are working together in order to produce
00:19:56.020 | complex engineering systems.
00:19:57.740 | And what you would see here is agents working together
00:20:00.060 | to produce code, not one at a time,
00:20:02.020 | but actually in unison and harmony.
00:20:04.260 | The same pattern is being applied
00:20:06.000 | in writing very commonly.
00:20:07.460 | Writing was a human process, human and a tool.
00:20:09.780 | Over time, this has progressed to human and a machine
00:20:11.980 | assistant.
00:20:13.020 | And now we have a human that's actually
00:20:14.780 | leveraging not one, but a network of assistants.
00:20:17.440 | I'll tell you in my own personal workflow,
00:20:19.940 | now any time I call an AI assistant,
00:20:22.060 | I'm not just calling GPT-4, I'm calling Mistral-Large,
00:20:24.940 | I'm calling Cloud-3.
00:20:26.140 | I'm having them work together and also against each other
00:20:28.980 | to have better answers.
00:20:30.620 | This is the future that we're seeing right in front of us.
00:20:34.260 | So what?
00:20:34.840 | What does this type of revolution
00:20:36.140 | mean for everyone in this room, and frankly, everyone
00:20:38.380 | outside of this room?
00:20:40.100 | In cold, hard economic terms, what this means
00:20:44.740 | is significant cost reduction.
00:20:47.780 | So this chart is the number of workers
00:20:49.540 | needed at an S&P 500 company to generate 1 million of revenue.
00:20:53.420 | It's going down rapidly.
00:20:54.860 | We're entering an era where this will continue to decline.
00:20:57.700 | What does that mean?
00:20:58.860 | Faster and fewer.
00:21:01.020 | The good news is it's not so that we can do less.
00:21:03.100 | It's so that we can do more.
00:21:04.460 | And we'll get to that in the next set of predictions.
00:21:07.080 | Also fortunate is all the areas where
00:21:09.460 | we've had this type of progress in the past
00:21:11.660 | have been deflationary.
00:21:13.100 | I'll call out computer software and accessories.
00:21:15.400 | The process of computer software,
00:21:16.820 | because we're constantly building on each other,
00:21:18.820 | has actually gone down in cost over time.
00:21:21.300 | Televisions are also here.
00:21:22.420 | But some of the most important things to our society--
00:21:27.060 | education, college tuition, medical care, housing--
00:21:31.220 | they've gone up far faster than inflation.
00:21:33.700 | And it's perhaps a very happy coincidence
00:21:36.220 | that artificial intelligence is poised
00:21:37.900 | to help drive down costs in these and many other crucial
00:21:40.940 | areas.
00:21:42.380 | So that's the first conclusion about the long-term future
00:21:44.840 | of artificial intelligence.
00:21:46.340 | As a massive cost driver, a productivity revolution
00:21:49.420 | that's going to be able to help us
00:21:50.940 | do more with less in some of the most
00:21:52.820 | critical areas of our society.
00:21:56.340 | The second is related to, what is it really doing?
00:21:59.940 | One year ago on the stage, we had Jensen Huang
00:22:02.820 | make a powerful prediction.
00:22:04.860 | He said that in the future, pixels
00:22:08.020 | are not going to be rendered.
00:22:09.500 | They're going to be generated.
00:22:10.800 | Any given image, even information, will be generated.
00:22:13.900 | What did he mean by this?
00:22:16.020 | Well, as everyone in this room knows,
00:22:18.340 | historically, images have been stored as rote memory.
00:22:21.980 | So let's think about the letter A, ASCII character number 97.
00:22:25.900 | That is stored as a matrix of pixels,
00:22:28.420 | either the presence or absence, if we
00:22:29.940 | use a very simple black and white, presence
00:22:31.780 | or absence of those pixels.
00:22:33.900 | Well, we're entering a period in which we already
00:22:36.340 | are representing concepts, like the letter A,
00:22:39.300 | not as rote storage, not as a presence or absence of pixels,
00:22:42.700 | but as a concept, a multidimensional point.
00:22:45.380 | I mean, the image to think about here
00:22:47.100 | is the concept of an A which is generalizable to any given
00:22:50.260 | format for that letter A. So many different typefaces
00:22:53.560 | in this multidimensional space.
00:22:55.240 | We're sitting at the center.
00:22:57.100 | And where do we go from here?
00:22:58.740 | Well, the powerful thing is the computers
00:23:00.500 | are now starting to understand not just
00:23:03.060 | this multidimensional point, not just how
00:23:05.100 | to take it and render it and generate that image,
00:23:07.740 | like Jensen was talking about.
00:23:09.420 | We are now at the point where we're
00:23:11.380 | going to be able to contextualize that understanding.
00:23:13.900 | The computer is going to understand the A,
00:23:15.660 | be able to render it, understand it's an alphabet,
00:23:17.780 | understand it's an English alphabet,
00:23:19.660 | and understand what that means in the broader context
00:23:22.020 | of this rendering.
00:23:23.340 | Computer is going to look at the word multidimensional
00:23:24.780 | and not even think about the A, but rather
00:23:26.780 | understand the full context of why that's being brought up.
00:23:30.140 | And amazingly, this future is how we think,
00:23:32.860 | how humans think.
00:23:34.180 | No longer are we going to be storing the rote pixels
00:23:37.540 | in a computer memory.
00:23:38.420 | That's not how we think.
00:23:39.580 | I wasn't taught about the letter A
00:23:41.180 | as the presence or absence of a pixel on a page.
00:23:44.780 | Instead, we're going to be thinking about that as a concept.
00:23:47.540 | Powerfully, this is how we've thought about it
00:23:49.380 | philosophically for thousands of years.
00:23:51.180 | Here's my fellow Greek Plato 2,500 years ago,
00:23:54.300 | who said this idea of a platonic form
00:23:56.380 | is what we all ascribe to, are all striving for,
00:23:59.100 | that you have this concept, in this case of a letter A,
00:24:01.420 | or this concept of software engineering
00:24:03.340 | that we actually are able to build a model around.
00:24:06.220 | So what?
00:24:07.020 | Now, we've talked about the second pattern, this idea
00:24:09.180 | that we're going to have generalization inside computing
00:24:11.220 | itself.
00:24:11.700 | What does that mean for each of us?
00:24:13.180 | Well, it's going to mean a lot for company building.
00:24:15.900 | Today, we're already integrating this
00:24:18.420 | into specific processes and KPIs.
00:24:20.380 | Sonia just mentioned how Klarna is
00:24:21.900 | using this in order to accelerate their KPIs
00:24:24.340 | around customer support.
00:24:25.860 | They know that they have certain KPIs that they can drive
00:24:28.140 | towards, and they can have a system that's actually
00:24:30.300 | delivering information, generating great customer
00:24:32.300 | experiences.
00:24:33.860 | Tomorrow-- and this is already happening alongside--
00:24:36.740 | new user interfaces.
00:24:37.780 | That might be a different interface
00:24:39.300 | for how the support is actually being communicated.
00:24:42.380 | And this is what I'm personally incredibly excited about,
00:24:45.000 | is because of this future in which concepts are rendered,
00:24:47.860 | because of this future in which everything is generated,
00:24:50.220 | eventually, the entire company might start
00:24:51.980 | working like a neural network.
00:24:53.740 | Let me break that down in a specific example.
00:24:57.300 | This is a caricature.
00:24:58.700 | As with everything in this presentation,
00:25:00.580 | in reality, everything is continuous.
00:25:02.140 | These are all discrete.
00:25:03.220 | This is a caricature of the customer support process.
00:25:06.500 | You have customer service that has certain KPIs.
00:25:09.220 | These are driven by text-to-voice,
00:25:10.820 | language generation, customer personalization, and the like.
00:25:14.060 | This feeds into subpatterns, subtrees
00:25:16.940 | that you're optimizing.
00:25:17.900 | And eventually, you're actually going
00:25:19.440 | to have a fully connected graph here.
00:25:21.180 | You're actually going to have feedback from the language
00:25:23.560 | generation to the end KPI for the servicing of the customers.
00:25:27.700 | This is going to be, at some point,
00:25:29.300 | a layer of abstraction, where customer support is managed,
00:25:31.960 | optimized, and improved by the neural network.
00:25:34.540 | Now, let's think about unique customers,
00:25:36.760 | another part of the important job of building a business.
00:25:39.980 | Well, again, you have the primitives
00:25:41.540 | of artificial intelligence, from language generation
00:25:43.740 | to a growth engine to add customization and optimization.
00:25:46.820 | This will all feed into each other, once again.
00:25:49.140 | The powerful conclusion here is, eventually,
00:25:51.180 | these layers of abstraction will become
00:25:53.660 | interoperable to the point where the entire company
00:25:57.260 | is able to function like a neural network.
00:26:00.260 | Here comes the rise of the one-person company.
00:26:04.920 | The one-person company is going to enable us not to do less,
00:26:07.500 | but to do more.
00:26:09.260 | More problems can be tackled by more people
00:26:11.300 | to create a better society.
00:26:13.740 | So what's next?
00:26:16.440 | The reality is, the people in the room here
00:26:18.220 | are going to decide what's next.
00:26:19.520 | You are the ones who are building this future.
00:26:22.100 | We personally are very excited about the future,
00:26:24.300 | because we think that AI is positioned
00:26:26.300 | to help drive down costs and increase productivity
00:26:28.780 | in some of the most crucial areas in our society--
00:26:31.740 | better education, healthier populations,
00:26:34.200 | more productive populations.
00:26:36.260 | And that's the purpose of convening this group today.
00:26:38.500 | You all are going to be able to talk about,
00:26:39.800 | how are we able to take our technologies,
00:26:41.740 | abstract away complexity, mundane details,
00:26:44.300 | and actually build something that's much more
00:26:46.300 | powerful for the future?
00:26:48.120 | I'll hand it off to Sonia to introduce our first speaker.
00:26:50.700 | [APPLAUSE]
00:26:53.740 | [BLANK_AUDIO]