AI has advanced rapidly over the past decade – and exponentially so in the most recent couple of years. The rise of large language models and the many tools they can train has led to AI evolving into an enabling, platform technology.
Just as previous technology transitions of mobile and cloud impacted the nature of businesses, AI is beginning to shift nearly every category we invest in at Greylock, from cybersecurity and vertical saas to cloud infrastructure and consumer networks.
While it’s true that most companies and investors are exercising restraint in today’s economic environment, a fair number of early-stage startups are still attracting capital and partnerships. In many cases, these are companies that are developing tools using artificial intelligence – and for applications across a range of consumer and enterprise sectors.
We discussed this in depth at Morgan Stanley’s TMT Conference, where more than 3,000 investors and tech companies were in attendance (unsurprisingly, this year’s theme was AI). During our conversation with Morgan Stanley’s Global Head of Tech Private Equity and Venture Capital Investment Banking Umi Mehta, we discussed the impact AI is having on venture investing across enterprise and consumer sectors; the various ways AI is expected to impact nearly every profession by providing a “copilot” for various job functions; and the importance of developing the technology safely.
You can listen to the interview at the link below or wherever you get your podcasts.
Episode Transcript
Umi Mehta:
Thanks everybody for joining. I’m Umi Mehta from Morgan Stanley.
Before we jump in on what I know is everybody’s favorite topic, AI, I want to spend a few minutes and just talk about the overall VC ecosystem. The tone has changed, the investor sentiment has changed, what people are looking for has changed.
Just give me a little sense of where you think we are in the VC ecosystem today. Saam, why don’t you kick it off.
Saam Motamedi:
I’ll kick it off – and it’s good timing, we actually just came off a limited partner meeting last week.
I think it’s no surprise to anyone here that we’re in a reset moment in the venture capital industry. From the limited partner perspective, I think they’re going to be a lot more discerning going forward on the managers they work with. They’re going to orient towards people with a track record of actually building enduring businesses that can scale over time.
On the company side, we’re also seeing a significant reset moment. You have a number of companies that got ahead of their skis in terms of capital raises, the cost structures they’re operating their businesses at. We’re seeing that normalized. We’re seeing a lot of restructuring going on in mid to late stage tech, and I think that work is just beginning and it’s going to play out in the second half of this year and next year in particular.
On the positive side, I would note, early stage continues to be very robust and it might surprise people here that we are as active as we’ve ever been at Greylock. I think this week we had three companies into our full partner meeting on Monday. A lot of them have something to do with AI, which we’re going to talk a lot more about in a bit. We’re investing early, we’re investing with 10-year time horizons and we continue to be very active.
Reid Hoffman:
And I think part of the thing is, while obviously when you’re looking at these things and obviously we’re in a large language model AI moment (which Saam and I wrote about last year and a bunch of other things), it doesn’t mean you stop looking at marketplaces, networks, people.
One of the things that Greylock has been, for decades, one of the leading firms on is various enterprise SaaS, cloud security, et cetera. So you keep doing that. Of course, you’re also always asking in a new platform, okay, what’s the new platform you mean for this? What does AI as a new platform mean? And you’re always asking that question both of your current portfolio and also of course prospective portfolio.
UM:
So I heard some sub-verticals there that you’ve historically invested in. Outside of AI, are there other new verticals that the VC community and Greylock in particular is interested in?
SM:
On the enterprise side, it’s interesting, because AI is an enabling technology wave and it’s shifting every category that we invest in.
So for example, cybersecurity is an area that we have a prolific history and then we continue to invest in quite actively. And as you see the perimeter shift to the cloud and increasingly to home into mobile, that whole stack is being reinvented. And so that’s one example.
We continue to look at systems of engagement, systems of record. On the application side, there’s new themes and infrastructure. So at a high 30,000 foot view, the categories are the same, but I think for the first time in a long time they’re going through very material disruption. And a lot of these markets are for grabs.
“AI is an enabling technology wave and it’s shifting every category that we invest in.”
RH:
The other thing I would add to that is in addition to the classic areas that we look at – which, on the consumer side tends to also be marketplaces, Airbnb, a bunch of others networks, LinkedIn, Facebook, et cetera –some of the stuff is also pure play AI stuff.
So in terms of new categories, the first thing since LinkedIn I co-founded was a company Inflection last year. Saam led an investment in Adept, which also has a very heavy enterprise side but is not a classic category. It doesn’t fit in a category. So there’s some stuff also where that transition happens.
And I think one of the things we will see from AI as a platform technology is I think some new categories will emerge and we’re looking at a lot of seed and series A deals trying to figure out which will be those, what will be the value creators.
Because the short answer is now more or less if you go to a cocktail party that’s for the real estate industry and you ask what they’re talking about and the answer is ChatGPT. So it’s like, okay, everyone’s talking about this. It’s not news, but the question is now that everyone’s gaze is on it, what does the next couple of years look like?
UM:
I was trying to use the first five minutes to not talk about AI. We’ve failed terrifically.
RH:
I broke the plan (laughs).
UM:
But let’s just go to AI.
So from our account there were 800 companies identified as AI companies. The growth has 10Xed in terms of just volume both in terms of funding itself. And so I guess individually, I’d love to hear from both of you, what makes you so confident, so certain that this is the platform shift and it’s as big as you are saying and thinking and investing behind?
RH:
So first, Saam and I wrote a piece six months ago now on something along those lines that said every profession will have a co-pilot within five years. It’s really within two to five years. Every profession will have a co-pilot that will be between useful and essential for what you’re doing. Once you think that, that doesn’t even get to the [question of] well, how does this change infrastructure, or how does this change services, or how does this change productivity software, or how does this change on and on and on security, on and on and on. Even that’s changing the industries. And then you obviously have how the product is reconstituted as well.
And so part of the reason why this is a wave that couldn’t happen without the internet, couldn’t happen without mobile, couldn’t happen without cloud. But it’s building upon all three. It’s the crescendo of them. And that’s the reason why the impact is very, very broad. When the individuals come up to me and say, well, how should I start planning my AI strategy? The answer is start playing with it. It will evolve rapidly.
And one of the mistakes people make is they say, “Well, I play with ChatGPT, I know exactly what it is.” Well actually, in fact, things will be a little different three months from now, things will be a little bit different six months from now.
“This is a wave that couldn’t happen without the internet, couldn’t happen without mobile, couldn’t happen without cloud. But it’s building upon all three. It’s the crescendo of them.”
UM:
Keep holding that, don’t give it away.
RH:
And what’s more, when you look at what’s being built on the APIs, even right now it’s like, well, it goes anything from new ways of constituting how you’re doing marketing to how you’re doing sales to how you’re doing customer service. And you think, well those are fundamental to how companies operate.
And that’s part of the reason why it’s a really interesting transformative moment. And the thing to start playing is not to think it’s fixed right now, but you say, well, what do I think people won’t get is that the slowness adoption will be a little bit, most people still don’t really fully know how to use search in order to solve all their problems.
There’s still a lot of cases where people don’t realize you can actually solve problem X by using search (usually) to solve a problem they don’t understand. If you type in six words or 10 words as a way of actually getting there, it’s the same thing as if you say, “Well, what should I think about this question? Well, can I put it in as a prompt that will get me something useful as my output?”
So literally I was talking to a friend and he was like, “Well, I’m going to resign from my job.” So I went to ChatGPT and I typed in, “Give me a resignation letter with the following three characteristics.” And he is like, “Oh my god, that’s a lot better than I would’ve done, and it took three seconds.”
UM:
Saam, before you go, I want to double click on the copilot concept. And so we have a lot of investors in the room. Give me the range of applications and tools that could be a copilot for this audience.
RH:
So for this audience, think for example, questions of at least, well, okay, what are the current trends in technology X? What are the current challenges that technology X is having? What company should I also look at if I’m investigating a company? This is all analytic due diligence questions.
I think one of the things is it’s a research assistant that, by the way, sometimes hallucinates and sometimes gets it wrong. You get it. That’s part of the things you learn and get things, but it delivers immediately as opposed to the, “Hey, go and please do this. Come back in two or three days.”
If I’m thinking about company X and I’m thinking about how it’s going to transform, say the market for databases, what are the other companies that would be really interested in this thing? And if I was considering it to be this new object-oriented database or database intersected with quantum computing, what would be the other things that I would look at in terms of doing and what, or I’m meeting with an entrepreneur, what questions should I ask?
Now? Right now – by the way, most of having done all of these things as an investor, everything I’m telling you, I have actually sat in front of the prompts and done – [asking] which questions [to ask] are too vanilla. So far GPT is not good at generating the question to ask. If I were to say, “I’m interviewing you, what questions should I ask?” And it said, “Well, ask about the markets.” I was like, “Okay…”
UM:
Yeah, they weren’t that good. I asked.
RH:
Yes, exactly, yes, exactly. Not good at, what questions should I ask? But it’s like an instant Wikipedia page on whatever prompt you can think of that has some general characteristics.
Now the other one to be a little bit wary of is the training of these things is to be as interesting as possible. So when you’re asking overview, synthetic, et cetera, it actually hits it really well. If I said, give me Umi’s biography, it might invent something, [for example] a PhD from Cambridge University. And you’re like, “Okay.”
So when you go to the specific, you have to understand which areas is it good on. So anything in that arena, and that’s only beginning to scratch it.
SM:
I’d add two things just around your initial question, which is what’s the motivation for the optimism and the excitement? So one is we’ve been investing in AI for a long time at Greylock and this new wave of AI for several years now. The thing that continues to surprise even us is how quickly and how significantly these models are improving. And so it’s just hard to fully have intuition for exponential growth. And I think this is one of the areas where two quarters from now, four quarters from now, eight quarters from now, everyone in this room is going to be surprised by what these models can do and the applications they enable. That’s one thing I’d note.
The second is we’re still early days in terms of modalities. So the way most of us have interacted with these models is primarily via text with things like ChatGPT and image with things like DALL-E and Stable Diffusion.
But later this year or next year, we’re going to see many more modalities come out. Reid just referenced our example in Adept. For those of you who aren’t familiar, Adept is a model to take actions. So instead of generating text, Adept can use something like [the fact that] Jeff [Lawson] was on stage right before us. Use something like Twilio to take actions on behalf of the end user. And you begin to think about, okay, if you have these generative models that can be interacted with in natural language and use any tool that we as humans can use, what types of things does that unlock?
And that’s just one example, but I think these additional modalities as they come out, many, many more applications are going to get built.
“The thing that continues to surprise even us is how quickly and how significantly these models are improving.”
RH:
And to give you an example of the applications that will be live basically this year is another of our portfolio companies that I’m on the board of, Coda. [The application will] take notes in the meeting, give us a summary, give you what your action items are, the triggers go to, like go now send an email to Umi about this, right? Et cetera, “Inform so-and-so.” All of that stuff now becomes the assistant. It all becomes doable, and that’s the reason why there’s a stack of companies because that’s essentially an application company in productivity.
UM:
Well, I think you’re bringing up a great point because I do think there are theories in the market that it’s about elimination, not enhancement, and so you just described the enhancement of the position and the job and the to-dos as opposed to elimination. There’s also some themes around just critical thinking. And so when I think of the limitations around critical thinking, will we have to be less critical in the way that we do our work? Will these LLMs catch up to us? Give me the critical thinking concept to this.
RH:
Let me take this one first. So for example, take ChatGPT. It comes out and the education establishment has a collective aneurysm going, “Oh my God, we can’t assign essays anymore.” It’s like, we’ve got to run this still as if it’s the 1940s, which is when we had all this stuff. So I sat down for three minutes to think about how could I say, “GPT, teach essay writing.” That’s one version of critical thinking.
Here’s two. One is to say your essay is the argument’s for and against, and then you have to make both arguments, and then you have to conclude as to what’s there. That’s something that you can be using GPT to help you with, but you have to essentially still get to that conclusion or. So you don’t want to do the pro side, antagonist side. Well, okay, so I’m an English teacher and say the assignment is Jane Austin and colonialism.
I go to whatever the current ChatGPT is, I generate five different essays of my prompts. I hand out those essays to the students saying, I’ve done this with GPT, this is a D plus. You have to hand in something better. So it sets the new benchmark. But then you go use GPT, and by the way, I can use GPT to help grade and all the rest, but if you just typed in a prompt saying give me an essay about Jane Austen and British colonialism in Africa, you’re going to get one of the five essays. You have to figure out how to make the essay better than what you got. And that’s the reason I do that as a lead in using the education thing is critical thinking will still be there. It’s still going to be amplifying.
You’re looking at this thing. When I was saying what I learned from having been using GPT for a year and a half now in various ways, is I learned which things where I go, “Oh, that I don’t want to just rely on, I want to crosscheck.”
UM:
How many times do you use GPT in a day?
RH:
I’d say at least four and sometimes 20.
UM:
Okay. Specific?
RH:
Yes.
SM:
The thing I would add is it’s remarkable how effective ChatGPT is at Socratic reasoning. And so actually I use it this way, which is if you’re thinking through a problem as a thought partner to help you reason through it, it’s actually very, very effective.
RH:
Hence again, critical thinking. I think the anti-critical thinking thing is a lack of thought on the part of the critics. It’s clearly just [possible to] use it as an amplifier and you still have to go further down that path.
UM:
Saam, you mentioned there’s been massive adoption, but when does it become ubiquitous within certain sectors and what’s the adoption curve look like? Which sectors will be affected the fastest? Maybe which ones won’t be at least until there’s some other version out? Comments on some of those things.
SM:
Sure. So I’d start by saying it’s already ubiquitous. So if we’re using Google search, parts of Google search are publicly discussed as being powered by large models. Those of us use different social media applications that have recommendation systems powered by large models. A number of the existing application companies that have data gravity. So whether it’s Salesforce with their announcement of Einstein GPT or Workday and things they’ll do on top of HR data. I actually think this is proliferating very, very quickly.
The other thing that’s remarkable about these companies (and a data point I can share from the early stage perspective) is because these tools really feel like magic. The first time I used ChatGPT, it certainly felt like magic to me and the level of value they deliver to end users is a step function over prior generations of software. The velocity with which they get adopted is also unlike anything we’ve seen. So for example, just to give you one example of a concrete case, we incubated a company called Tome at Greylock. Tome is a generative AI way to tell stories or to build presentations or decks. This company reached a million monthly active users faster than any productivity tool we’ve seen.
UM:
The analysts and associates in here look very worried right now, by the way.
SM:
That’s a co-pilot for analysts and associates.
RH:
It’s a co-pilot.
UM:
There you go.
RH:
But for example, I’m starting an additional podcast called Possible, which is how do we imagine what future is possible? How do we build towards the technology?
I sat down in front of Tome and I said, “Write me a deck about a new podcast by Reid Hoffman Possible that says technology is between 30 and 80% of the solution of any scale thing. The problems can include criminal justice, economic justice, climate, et cetera will involve AI. Give me a deck.” In five minutes. I had the beginnings of a workable deck with 10 slides, graphics, text, et cetera. It was not the deck I would start with, but I was there.
UM:
You saved a week, or you saved a couple of days.
RH:
Yes.
SM:
The other thing I would add is just, so for example, one area Reid and I talk about is customer service. So if you think about areas inside the enterprise that are more of a cost center where there’s more of an orientation towards efficiency and where often the types of things happening are more repetitive. I pick on the contact center, but you can look at different categories of work to be done that today might be serviced by BPOs or different automation players. Those areas lend themselves really well to augmentation and eventually for some subsets automation with these technologies. So I think that might be another area where you see things begin to take off quite quickly in coming quarters.
UM:
Lots of positives. Let’s talk about some of the potential negatives, whether it’s using data to your advantage, some of the data privacy, who actually owns the content at some point in time. And so maybe there’s a lot there to unpack, but I’d love to hear how there should be some agency set up within the government at some point in time to regulate what comes out of ChatGPT or some of these other models?
SM:
Yeah, so I’d say you’re absolutely right that these are an important set of challenges and those of us who have played with these products, we see the challenges. Reid referenced hallucination, there’s questions of data ownership, data privacy.
In an enterprise context for these models to really get deployed in mission-critical applications where efficacy and accuracy matters, there’s a gap that has to be closed. That gap will be closed and it’ll be closed in two ways. One is you’ll see companies like OpenAI and others find ways to adapt these models to enterprise environments that preserve the integrity of the enterprise’s data and give better guarantees around performance and not hallucinating. So imagine ChatGPT for Morgan Stanley on top of Morgan Stanley sensitive data in a way where the answers it comes back with are Morgan Stanley-specific. We’re going to see a lot of that type of adaptation happen.
RH:
With regulatory logging and all this stuff that’s a hassle in your life.
SM:
And then the second is tooling that gets built from an explainability fairness bias perspective that helps make these generations and predictions more tractable and more explainable and that also inspires confidence and allows them to be used more and more in mission-critical applications.
RH:
One of the things I think that people don’t realize is that part of what happens is these large language models are trained on massive amounts of data. It’s the massive amounts of data and the general model that gets created there, which doesn’t involve any of the things that people are worried about, is high-value data. And so actually it doesn’t even involve any of that. You can then tune them to high-value data and then once you tune them – because we don’t really have explainability yet – you would then have to keep it contained to whatever that tuning is. If it’s tuned to Morgan Stanley data, you have to have an authentication for using it and we’d have to log everything because this is a 200 or 300 billion parameter model. You can’t look under the hood and go, oh, “This is where the data is,” and so forth.
But people then say, “Well, if it was training on publisher, did it train on publisher X data?” Actually all the base models almost don’t do that. And it isn’t like, “Oh, it’s publication X because it has a really good columnist.” It’s often a massive size of the data and it’s tuned. And the tuning is part of what gets the co-pilot thing that Saam and I were talking about, which is, well, you can tune it for writing code, you can tune it for writing poetry, you can tune it for doing medical stuff, you can tune it for, there’s a bunch of different things you can do that once you have that general model.
SM:
Specialized verticals.
RH:
Yes, you can do specialized verticals. Sorry, you were about to say something.
SM:
No, I was just going to add, and I think when people think about tuning, there’s different levels of specificity. So we talk about code we’re all familiar with Co-pilot, it’s having amazing adoption, but if you just take code as a modality and you drill deeper, for example, you look at code quality, code security, infrastructure instrumentation from a monitoring perspective, you will see models that will be tuned to those specific use cases and that’s how you’ll see applications get built that really have a chance to go disrupt these software markets.
UM:
Where do you think AI ChatGPT will play into the election next year? And where is it going to get its information from?
RH:
Well, one of the things that’s cutting edge – part of what Bing’s announcements have been and so forth – is that these are trained on a bunch of data and then they’re fixed. So how do you bring in live and current information?
That’s a work that’s very rapidly and iteratively progressed. That’s part of what the February Bing announcements are gesturing at. So I think there’s intersections between that that will be important to how those play out.
Now when it gets to the elections, I think that part of the question will be, “Will we have a lot of generated misinformation?” It’s like, well, we already have a lot of generated misinformation and some of it’s from actual people, some of it’s from paid people, some of it’s from other things.
It’s not that I don’t pay attention to it, but I’m actually somewhat optimistic that we might be able to figure out in these timeframes and the actors are saying, “How do we help?”
For example, you say, “Well I have a co-pilot that helps me ascertain things relative to, let’s say, medical knowledge.” And I’m looking at [the question of whether] vaccines cause Ebola. And he goes “BEEP” [wrong answer indicator] right? But then that’s part of where you say, well look, it’s actually in fact also part of the solution.
Now you have to have people who then say, “Look, I care about something that’s trained and tuned on a set of expertise. I’m using it as a copilot in some manner.” But if you get that and you have that deployed in various ways, and it makes maybe some product warranties or claims or other things that maybe have teeth to them, maybe not, whatever – well, misinformation has been a very common subject around AI for X months. Well how would you solve it? And I’m just like, sit down for three minutes and say, “How would I deploy the technology and try to help and solve doesn’t mean solved to zero Because by the way, there’s tons of misinformation on the internet. There’s tons of misinformation within Google search. There’s tons, but it’s within the parameters of how we’re operating.
UM:
Maybe the next question for me is a little around, I talked about what goes wrong – and I just like this question. Utopia or dystopia? Saam, maybe Utopia and I’ll let you take dystopia.
SM:
I’ll start with Utopia. You’re going to hear a recurring theme because it’s how important we think this paradigm is, which is the co-pilot paradigm. I think if you take both what we’re all doing in our personal lives, whether booking a trip, thinking through a problem in your personal life and in the enterprise context, what we’re doing at work, we are going to have hundreds of tuned co-pilots that make us 10 times more effective and 10 times more efficient at everything that we do.
And I’ll just give you one concrete example. We talked about customer service earlier. We’re investors in a company called Cresta. It’s a co-pilot for the contact center. So if you call in to Verizon today and you’re talking to an agent in Tempe, Arizona, they have a co-pilot that’s steering them through that conversation and helping that person sell a new cell phone subscription, earn more money and be more effective at their job. And that’s one example, but it is going to proliferate everywhere and I think it’s going to be a massive step function in knowledge, work productivity.
RH:
So on the dystopia side, unfortunately, I think a lot of the discussion is a little bit misleading. You have the, “Oh God, it’s going to become self-aware and I’ve watched the movie Terminator and so forth!”
Part of the reason I tell people to go play with ChatGPT is you go play with it and you realize, “Oh, well these are the things it’s good at and these are the things it’s not good at.”
For example, right now it can’t solve, say, you’re an expert in investing and you ask it a question you’re an expert at, it will not give you a better answer than the one you have in your head.
So then what do you say, where are the downsides we should be tracking? Well, in a transformation with all the co-pilots, I think many jobs will actually not be downsized. I think that we’ll take infinite of those jobs. We’ll take infinite engineers, infinite sales people, infinite marketing people, et cetera. Customer service I think will be one of the ones that’ll be a little bit more challenging.
For example, once we solve autonomous vehicles, we’ll want all the vehicles to be autonomous. We won’t have the gridlock that at least I experienced getting here. You might have too, the other issues, [the impact] it’ll have on climate change so forth. So that will also be more of a replacement, but there will also be creation of new jobs. It isn’t just like when you get 10 x more efficient as a graphic designer, that doesn’t mean there’s fewer graphic designers, there’s a lot of things too. But the dystopia that I’m gesturing at and saying “This is the transition,” not, “Oh my God, I’m worried that it’s going to be like Elysium the movie.”
Because obviously Hollywood films do a lot better with dystopias than they do with Utopias and go, oh, there’s going to be this small group that lives on the moon and all the rest of the people are whatever, and it’s a robot police force. I really think that’s grounds to zero in our grandchildren’s lifetimes. And who knows, I don’t know how to make projections beyond that in various ways just because look, the world changes and I’m generally optimistic. But the transition will be things that we need to be intelligent about.
When we did the transition from agrarian to industrial, it was very difficult on society.Now, even truck driving, you say, “Oh my god, all the truck drivers jobs are going to go.” One, there’s a huge number of empty listings right now for truck drivers. Two, if all manufacturers on the planet started making autonomous trucks right now it’s 10 years before you replaced about half of them. So it’s not like, it’s like, “Oh my God, tomorrow is going to be totally different.” So again, transitions I think are really important.
The other area that I would go that I think we haven’t seen anything yet is cybersecurity. People say, well, “Oh, I’m worried about the machine becoming super smart.” And you’re like, “Okay, I’m much more worried about these tools in the hands of malicious humans doing things that malicious humans do.”
And so the questions about what that means, and what you do there, and what happens and are you ready to deal with that is another area that I think is a legitimate area of dystopia that we have to be careful about.
So to go all the way down to one of the reasons why OpenAI maintains a very high ethics standpoint on this stuff. DALL-E was ready for four months before it launched. I was playing with DALL-E for four months before it actually launched. Why? Because they were like, “Oh, it could be used for other really terrible things that would happen. Let’s serve it through the API and let’s tune it to safety for all that.” That’s the kind of thing, it’s the humans using it for very bad things is the thing you have to track.
Now in many cases, to your earlier regulatory question: Look, I think dialogue and goodness for societies are important, but this will be redefining industries and what we want as societies, as investors, as inventors, as workers, we want to be in the industries of the future. We don’t want to be Luddites. If you say, well no, no, all looms should be hand done and we should just stay with hand done looms. That position is a very bad position in a very quick amount of time.
“The world changes and I’m generally optimistic. But the transition will be things that we need to be intelligent about.”
UM:
Yeah, it’s back to my [question of] who’s going to regulate and all that.
RH:
Yes. But I think that the right answer is not to say, oh, regulators should go drink Mai Tais on the beach and look the other way, which is sometimes how they hear the [idea of] “Don’t do anything rash right now.” It’s like, “No no, no. Get engaged in conversation. The thing that’s like, what’s the outcome you’d like to see?
I’ll give it a parallel. Say, well, I’d like to see very few acts of terrorism and online video. Should I then say, “Well, I’m a regulator and so I’m going to create a process. You have to delay broadcasting it by 15 minutes and da da da, and this is now how we’re going to lock in what you have to do and that’s how we’re going to do it.” And you’re like, “Well, that’s probably not even going to work.” And is a regulator a good technology design?
What I should do is I should go say, okay, I realize getting to zero is probably difficult, but let’s say the first 100 impressions of saying you have to run through your auditor, you turn off when you got to a 100 impressions, you’re fine. You get to 1,000, that’s a million dollars, you get to 10,000, that’s 100 million dollars. You figure out how not to do this. This is what I care about. I want you not to be doing this. You figure out how to do it, we’ll give you an economic incentive to do it, and then you improve it as you do it. That’s a much, much better way to do this.
And that’s the kind of thing everyone goes, okay, what should we be doing in AI? It’s like, look, we should be embracing all the things. For example, what I see in my line of sight right now is an AI tutor and an AI doctor on every cell phone. If you slow that down, think of the human suffering that you’re creating by not having that, right? That’s super important to get to.
UM:
How far away are we from that?
RH:
You can build it on today’s technology. It is doable today. Now there’s work to do, interface, a bunch of other stuff, but it is doable on today’s technology. That’s the thing, if I could just say, well, I’m a government, I want that for all my citizens, please, that would be really good for the wellbeing of my citizens. I want that. That’s what you want.
UM:
There’s probably lots of questions, so maybe we can get some mics around. First question.
Audience Member:
At Greylock, where do you all think most of the value capture is going to happen? Is it going to be at the application layer, the data layer, the closed source, open source model layer, proprietary models or for that matter, even the compute? And also how does one monetize? I mean you spoke about ethics, biases, whatever might creep in. And also I’m pretty sure the data of mine that’s being used to create some model somewhere, I wouldn’t be surprised. How do I get monetized and are there unique models that are emerging to help monetize that data that you’re seeing?
SM:
Yes. Maybe I’ll take the first part and hand it to you Reid for the second.
So it’s a good question. There are three layers, and I wish I had an either or answer, but the answer is all of the above. So we think of it in three layers. There’s the underlying foundation models themselves and our view is there’s a select number of players and as sub 10 who have the ability understanding capital compute required to actually build these foundation models.
We’re nowhere close to the end of returns on scaling. And those players will create value, there will be differences, some will focus on one modality versus another, but there will be that layer and that layer will monetize through an API model, like Twilio and others that we’re all used to. Then there’s a layer on top, which we think of as the infrastructure and orchestration layer.
And so for example, you want to now take one of those things and adapt it to an environment like Morgan Stanley. There’s a lot of things you have to do. There’s a data management piece, there’s a serving piece, there’s integrating it into my internal application piece and there’s interesting opportunities for new companies there, new tooling companies there, and we’re investing in those. And then the top is the application layer.
And I think there you’re going to see both existing applications that own the end user, have some proprietary data asset that has gravity, that’s a consequence of their workflow. They’re going to incorporate these LMS into their businesses and it might be a turbocharge for existing.
And then you’re going to see categories. Tome is a good example, where the whole ergonomics of how you interact with software shift because of these models and that creates new opportunities for new applications. And some of them will be in existing categories, some of them will be in categories we don’t really know how to name because the ergonomics weren’t previously possible and those businesses will monetize application companies today do.
And so I think our view is very optimistic around all three layers, recognizing each of those layers has different dynamics in terms of what’s required for success.
“We’re nowhere close to the end of returns on scaling.”
RH:
And then the other thing is I think there will be in business models, I think a bunch of the standard business models, I think subscriptions, I think ads, I think a bunch of those things work. I think we will also see some new invention business models that will be unique to technology. Just like to some degree AdWords is unique ad technology for this stuff. And so I think that’s one of the areas we pay attention to.
Now the data thing is one of the places where a lot of people are just misconstruing what’s the value of this data? And well, the value of the data is in what context? So for example, you’re in a search and you’re typing “mesothelioma super valuable”, I have a possible intent and issue and so forth or Sony camera, super valuable. I mean word documents typing it not that valuable and they may even be signals and so on.
And so part of the thing that people don’t understand is there isn’t like the fact that Reid is wearing a blue jacket is not an intrinsic, that’s a two-set item piece of data. It’s a context of which it is and part of what these systems build as they go. We try to provide you systems of value by which you’re engaging data with us sometimes by the way, that’s SEO things because I want to be discovered in various ways and then it’s in public indices. Because that’s valuable to me in some way. And then other folks that trade off has figured out how to use that data in things.
And so for example, I had a very funny conversation five or six years ago with someone in Silicon Valley who said, what does Google give me for my data? I’m like, well research. So the question is what that context is. So I’m quite certain that all of us data is being used in multiple contexts because we’re engaging in various things where our engagement in that is somehow the data is part of what we’re being attracted into doing this thing that, for example, posting a web post on and SEOing it because I want people to discover it. Well then it’s also out there publicly.
SM:
And I’d add just one thing, hearing Reid talk about new business models, I think there’s a new business model invention, but another thing we talk about is evolution. So if you take something like ads and just to give a concrete example, take a software developer who’s building a piece of software and goes to a search engine and makes a query and there’s ads around different software packages the person could use while in a world with copilot for developers, you still may have that flow happen. It just doesn’t look like a query and a set of links. It looks like autocomplete in your IDE with a suggestion around a promoted package. Now you still can have that same concept of advertising. It’s just going to evolve in the way we consume it.
Audience Member:
I have a question on compute usage. So over the period 1960s to 2010, the compute usage doubled every two years, but in 2010 to 2020 it doubled every three to four months. So if you can conjecture 10 years from now if you’re sitting on that stage, what do you think that doubling time would be? And are we as a human society constrained by either hardware or data to continue that progress in this AI model?
UM:
And let me just throw one in there. Does capital then become the competitive advantage to have all this compute power?
RH:
Oh, it certainly is in some context. It depends on which. So I think there’s, on the scale models, I think there’s the players who are going to as absolute great scale as they can and you say, why is it still valuable? It’s one of the things you and I were chatting about. We just say, well, if you actually can say I spent a billion dollars and I have a 20% better engineering aid, that’s worth it. 20% better, doctor, that’s worth it. 20% better lawyer, that’s worth it. It’s worth a billion dollars. It’s fine. In order to do that, you got the internet and all the rest. And it’s a great question because I think that computers are well-thought of as like energy, which is that we have infinite demand for it at a certain price. And so as long as we can deliver on that price, we will have infinite demand for compute.
And so I think the constraint of your acceleration curve is actually an economic constraint, not a natural law constrain. And by the way, of course part of the cost of providing compute is you have to build data centers and places to do it.
One of the ones that you didn’t mention in your variables is power. Power is a huge variable on all of this because compute is not the only customer for power as we sit here in this hotel and so forth because part of this AI revolution is a scale compute revolution. It’s part of the reason why when I was answering the earlier questions – like it’s internet plus mobile plus cloud, what we’re doing is going, “Oh my gosh, we now figure out how to apply a whole bunch of scale compute to creating amazing things that hadn’t existed before.” Well, we don’t yet see that if you have 20X of flops versus 10X of flops, 20X of flops is still better. And so that the compute demand will be very high.
“Computers are well-thought of as like energy, which is that we have infinite demand for it at a certain price. And so as long as we can deliver on that price, we will have infinite demand for compute.”
Audience Member:
Double clicking on your first layer, you talked about sub-10 foundational systems. You’ve been interchangeably using ChatGPT or GPT OpenAI’s product. And so go a little deeper onto that first layer. Are there sub-10 large language models or we’re talking about some additional natural language processing models, what’s in there? Or is this one revolution you’re talking about today purely about large language models?
SM:
It’s a great question. So when I talk about that layer, I think about the distinct companies that can train these foundation models around different data modalities. So it could be language and text with something like GPT ChatGPTT being an interface on top of how you actually interact with that underlying technology. It could be image like the DALL-E model from OpenAI or the Stable Diffusion models that are open source and from Stability, it could be audio speech. We’re going to see all sorts of these.
RH:
Or Adept in action.
SM:
Adept in action. We’re investors in a company called Atomic that’s building a foundation model for RNA structure design that can be used for therapeutic design. So we’re going to see a proliferation, but these models connect to the prior point, order of magnitude a 100 people in the world that have actually scaled these models to the levels of scale that we’re at today. And certainly it’s going to grow from there.
So in terms of capability and then capital access to compute and data, there’s likely sub 10 companies that will have offerings at that layer. So that’s what I mean when I say sub 10. And there’s OpenAI, there’s Anthropic, there’s Adept, Inflection Google, to name five.
RH:
And the other thing is part of the reason why we’re so focused on large language models. There is somewhere on the internet, 10 to 14 trillion tokens of data, how we can train these things to be really interesting offers tokens data. Many of these models are trained on about a trillion tokens of data, maybe a trillion and a half for the large ones. And so there’s a bunch of headroom there and all kinds of things.
And you’re not even getting into one of the questions we dealt with earlier, which is how could you fine tune within an enterprise’s data or other kinds of things.
So anyways, there’s a stack of things there and the large language models are the first that are generating this thing like, oh my gosh, I can see the co-pilot, I can see how it’s working. There will be other scale compute techniques, large language models I think are now certainly one of the important tools in the tool chest, maybe in three years still the most important tool in the tool chest, but maybe it’s not.
For example, if you look back to the very early kickoff with what DeepMind was doing, self-play is another scalable compute thing. And so there’s various ways to put that self-play comes into these things that also generates interesting things that when you move from twoX flops to five to 10 to 20, you get an increasing performance curve.
The thing I’m nearly certain will be is that within three years another interesting major scale tool will be, “Oh yeah, we’re using that one too”. And which one is this? Well, I’m spending, I’m scratching around trying to figure it out because that’s my job, but who knows.
Audience Member:
Hi, I wanted to ask about the co-pilot thesis and the part that the mobile phone is going to play in it. Obviously we’re seeing the models becoming bigger and bigger, optimizing but still becoming bigger. I want to ask, how much do you think the barriers of the mobile phone to actually have the models locally installed versus API calls is important for the co-pilot paradigm to actually exist? And how far do you think we are away from that?
SM:
So I can share a quick thought on that and Reid please add, or maybe two quick thoughts. So one is it’s very use case specific. So there are use cases where you can use the mobile device to interact with models on the cloud and the latency is fine, you get things back and the speed you need for the use case. And candidly, I think a lot are going to play out that way. So that’s one thing I’d say we’re there today.
But the second thing I’d say is it comes back to this concept of smaller models being tuned for very specific use cases. So while you’re right that these foundational models that are solving really horizontal tasks are getting larger and larger, pick your favorite example, but if you take an AI tutor for fifth grade math, you can actually build a small model tune to that and that’s actually highly performing at that and that could run on your mobile device today. So I think the combination of those two things, depending on the use case, will solve the mobile challenge.
RH:
Yeah, it’ll be more, I think what are the ones that are specific that can’t have the latency of the network that then have to be on it. So for example, you said, well I want to use my mobile phone as a driver assist that will automatically quickly recognize there’s something and then do something – I don’t have time for the latency. So it has to run in the context of I’m putting my mobile phone on.
UM:
Fifth grade algebra, I think you’re okay.
RH:
Yes, but it’s algebra, it doesn’t matter, oh wait, it was 500 milliseconds or even two seconds, who cares?
And then so much more I would say that we probably haven’t really taken, just to give you a sense, is that we don’t really spend a lot of time at the Greylock partnership going, “What are the models that can only run on the phone?” Because there are so many interesting super valuable things where the network latency is not a problem. And so you get cloud-plus-phone.
Audience Member:
We talk about AI being the next platform shift or next shift of computing. And I guess if you look over the last few, whether it goes to cloud mobile, seems like they’ve all been won by existing winners and we haven’t really had a huge win for new players since the internet. So why is this any different where the key winners aren’t Amazon, Microsoft and Google who already seem to be really far ahead?
RH:
Well, if you mean cloud like AWS then yes, but a few clouds is a revolution. There’s obviously Workday and Salesforce and a whole stack of companies there. I’d say that the short answer is both are going to win a lot, not an either or. It’s like false dichotomies, you tend to say, well, which one’s going to win more and so forth. And I think we are going to see great things by the tech companies that do the revolution.
For example, like you say, mobile, well, it’s what resurrected Apple, right? Apple wasn’t Apple before the thing. And so I think that’s just literally this platform thing is, it depends on which. Now if you say, well, what I’d really like to do is I’d really like to build a billion-dollar computer and be competing with them, then well you got a bit of a challenge to start up. You better have something pretty unique and an interesting idea.
But that’s for example. Also, if you had come to me 10 years ago and said, “I want to launch a new desktop search engine…” okay, you need new platforms, new technologies, new market opportunities that are not being taken in order to do that. So anyway, I think that it’s going to be a huge amount of opportunity across both and growth.
SM:
Yeah, and I mean two things I’d add. One is you’re going to see dramatic business model disruption that creates the opportunity for new companies. I was just thinking about that. We’re talking about search and advertising as one example. And then the second is, I think one thing that may not be fully priced in yet is how different some of these categories are going to feel because the interface changes and we’re all now used to interacting with ChatGPT and other tools with natural language. Natural language becomes a very ergonomic and flexible interface to do different workflows. I’d argue it’s a much more material shift than the client shift from desktop to mobile and how these categories of application software are consumed. And so what it means to build a PowerPoint or tell a story is going to be very, very different built on top of these technologies. And I think that might create more of a window for disruption.
RH:
Actually, just one amplification. Because I loved this comment I read a couple weeks ago, which is what’s the biggest programming language on the planet? English.
SM:
Yeah, exactly, exactly.
RH:
Right. As a function of that.
Audience Member:
On the monetization end, I guess you can think perhaps on the copilot thesis, the upper limit could be what humans or what companies are willing to pay humans to get that work done. So as we think about moving this to more of a technological model or moving it more to machines, are there constraints that you think that now come into play in terms of that could change where we land on that monetization gap of starting either from free or going up to what we’d pay humans? Any way to think about how the monetization could land in between there.
RH:
It’s a very complicated question that will change a lot year by year and in different functions, but that I was gesturing at principles when I was saying, compare and contrast sales and customer service. Sales, when you go to a company, you go, well, you have a 10 x amplifier on your sales capabilities. Great. We still like to have our whole sales force. We’d love to have that, make that better for us. If you say you have a 10X amplifier in your customer service, they’ll go, oh, we can reduce our cost line. And so those are places where the variables will play off on this and that’s the reason why there will be some replacement. But I think that the thing that is, when you really begin to look at it, you’ll find that even under existing many job cases, when you look at it, you’re actually looking at things where the amplification and the use of it is much more interesting than that.
Now, that’s the reason I’m fundamentally optimistic, and fundamentally as far as the dystopian case, I was saying “transition,” and what do the transitions look like? But I do think that that’s part of the economies of how we run. So it’s like, okay, well what do we pay a human for this? What do we pay for something else? What makes businesses really effective is they think, okay, can we provide this quality of product or service and can we do it in a much more competitive way at a cheaper price? That’s the iteration. And so figuring out how to do that.
UM:
We do have time for one more question.
Audience Member:
There was an interesting paper recently, I think some researchers in Japan, they hooked people up to FMRI machines, they showed them images and they took the data, they fed it into an LLM and it spat out the images with pretty decent fidelity. So I’m curious whether you spend any time thinking about novel sources of data signal that maybe were only kind of sort of useful like data exhaust coming off of all kinds of different things that maybe were not quite commercializable in a relevant way before the advent of large language models where now all of a sudden there’s all kinds of interesting stuff that you can do.
RH:
So the short answer is yes. That’s among the things that get to where are the interesting new companies or business models, or do you have a data source or an angle and a data source that causes that.
Now you can create a really interesting product that couldn’t exist before. Maybe it has a different business model. I know about that paper you make this large language model, this foundational model, and it’s 300 billion parameters and it’s trained off the patterns of what we encounter through human language and images and words and mapping them. Because part of what one of the things that people don’t fully realize is in order to do the image generators, you had to have the text things first because you had to be able to say, if I would like to have a panda bear in an astro suit on a rocket, and so forth, that has to understand what all those categories are in order to be able to classify them, be able to produce them, et cetera, and all that.
And so you have that and that’s a massive platform layer. Then you say, “Oh, actually in fact, I’m just doing a translation. I’m doing a translation of what this neural pattern looks like.” So it’s trained on all this human data. One of the things that we’ve been discovering is the larger and larger model, the easier it is to tune to what human concerns we want it to have, but it also means it tunes other aspects and human life. And so the short answer to your question is yes. And that’s one of the threads that we spent a long time looking at.
“Now you can create a really interesting product that couldn’t exist before.”
SM:
Yeah, and the one note I would add is if you just generalize that to data earlier, there’s a question around compute. We talked about compute, but data is an equally important dimension. And at all of these companies, that’s something that’s top of mind. And it’s both publicly available data sets, data sets that are proprietary, looking for exhaust from new sources that haven’t been considered. And then the last thing I note is also using the models themselves to generate more data. And there’s a lot of interesting work happening there.
UM:
Well, that was extremely interesting and very thought provoking conversation. So Reid and Saam, thank you again for joining us. Really appreciate it and thanks for the audience for the great questions. Thank you.
