"Martin Casado is a general partner at venture-capital company Andreessen Horowitz, where he focuses on enterprise investing. Mr. Casado started his career at Lawrence Livermore National Laboratory, where he worked on large-scale simulations for the Defense Department. His work, first as a researcher and now as an investor, gives him insight into the development of artificial-intelligence products and usage.
Wall Street Journal reporter Berber Jin spoke with Mr. Casado at The Wall Street Journal's CIO Network Summit about the current capabilities, and the possible future of, AI. Edited excerpts follow:
WSJ: What excites you most about generative artificial intelligence?
MR. CASADO: With machine-learning AI, basically you have data, you create a model, and use that model for predictive stuff. We've been using this forever, for fraud detection, pricing and all sorts of stuff.
Generative AI is a very different type of system. The cost of creating something new now is very, very cheap.
For example, look at these large image models. If you're like, "OK, so I'm Martin." If you enter the prompt, "Martin in the Alps, drinking a martini with an alpaca," the cost to create that is about one one-hundredth of a penny. This is relatively new technology, and relatively unoptimized technology.
If you compare that generation, that one one-hundredth of a penny, to hiring a designer or going on Photoshop or, God forbid, flying to the Alps, you're talking about four or five orders of magnitude in difference in cost, the time to create that. And that's very rare in my 20 years' experience doing -- to have four of five orders of magnitude of improvement on something people care about.
The way that I think about it is: The creation of the microchip brought the cost of compute, the marginal cost of compute, to zero. And with the internet, and with software, the marginal cost of distribution went to zero, meaning you can reach anybody. And that's why software businesses are these 80%-margin businesses.
I think there's a chance that with these generative things, not traditional AI, the marginal cost of creation can converge on zero. And I don't think we have a clue what that means. But we have some glimpses. And it's interesting to talk about those glimpses.
WSJ: Venture capitalists are putting a lot of money into AI companies, trying to see what sticks. What do you think is going to be an enduring business?
MR. CASADO: This is the question. If you bought a computer in 1994, you would get a central processing unit made by Intel. But then there were all these other special-purpose ASICs [microchips] that were on the board that did stuff that the CPU couldn't do. What you didn't know is: Are each one of those coprocessors going to grow to be these huge companies? Or are they going to be consumed into the CPU? Is the CPU going to basically have everything?
We have the very same question now with these foundation models. I don't know the answer.
WSJ: A new AI-based large-language model, GPT-4, was just released by OpenAI. It's much more powerful than 3.5. How do you think businesses will incorporate GPT-4 into their everyday work streams?
MR. CASADO: Let's just talk to the generative space. There are certain spaces where it's clearly directly applicable. The easy one is the creation of images. People like images. There are markets for images, whether that's clip art, or for a book, or illustration, or whatever. And now we can do that very, very cheaply.
When it comes to the language models, that's not so clear. You have actual natural languages. And then you've got code. And one has to be correct. And the other really doesn't have to be correct.
So here's what I am pretty comfortable with now. I work with a lot of relatively junior analysts. I actually ran this experiment because I was curious. We get early access to these things like GPT-4. So I got some write-ups of companies. And then I asked the junior analyst to summarize it. And I asked ChatGPT-4 to analyze it. And both were kind of wrong. So I think a lot of the things that we tend to hate about these language models, they're kind of stuff we hate about ourselves.
I do think that there's a set of tasks, like summarization, text generation, brainstorming, that it's already kind of like the co-processor that we have as humans. But it has a lot of the failing that humans do.
And I will tell you, I think it's going to creep into our lives in ways we never expected. I was just at a wedding. I've known this guy forever. And I've known his friend. And he just gave the most compelling groomsman speech I've every heard in my life. People were in tears. I'm like, "How can he come up with this thing?" And then after he did it, he admitted it. "Oh. This came from ChatGPT." So I do think it's like we're in a bit of a different world, right?
WSJ: Is Andreessen Horowitz using it for everyday tasks and stuff?
MR. CASADO: No. What's interesting about things like the internet, and mass changes like this, is they tend to follow shifts in consumer behavior, as opposed to business behavior. And so a lot of individuals are changing their behavior. I am.
I write code at night. It's kind of like how I relax. I work on a stupid game. It's not interesting. It's just literally a way to relax. I use GitHub Copilot [a program that offers autocomplete suggestions as a person writes code]. And it's fantastic. It's great. It just gives me hints. It saves me time from googling. That's what it does.
That's a consumer behavior that's shifting. Many people use it to brainstorm. Like, "I've got a talk tomorrow. ChatGPT, give me some ideas." And then a lot of people are using it for summarization. It's all consumer-driven.
And I remember this is kind of how the PC happened. And this is definitely how the internet happened, too. And I think that's what's going to happen for this level of change.
WSJ: I don't know if I would trust it to run larger-scale business operations, because of these concerns around accuracy, and hallucination, and that sort of stuff.
MR. CASADO: I'm saying something very different: that it's enough of a different life form that we don't have an understanding how to consume it in an organization. Organizations have a way of operating. But it's powerful enough that we, as users, are changing our behaviors already. It has infiltrated many of our lives. It has infiltrated mine in very significant ways.
I'll give you an example. I like writing a videogame. It's fun for me. It's silly. I'm 46 years old. What a waste of time. I have a friend. Do you guys remember the game "Where in the World Is Carmen Sandiego?" In a weekend, at a hackathon, my friend made a generative Carmen Sandiego where you literally put in a prompt, and out comes a full game of Carmen Sandiego. So the prompt could be like, "Madrid 1940s. Pre-war jitters. Kind of dusky. Neo-noir."
And then out will come the full game based on that. I really think this is cold-fusion powerful, but we as organizations don't know how to consume it.
WSJ: What do you make of the speed at which a lot of companies are starting to already integrate it into their businesses?
MR. CASADO: This is actually a great question. There's a lot of use that, I think, is like ChatGPT-washing. It's like, "Oh. It's a cool new thing." We saw it with the Internet of Things and we saw it with AI. They'll kind of include it, it's kind of bolted on.
I know of a few instances where it has been bolted onto old businesses in ways that are actually quite useful. A few people have figured that out.
And then for sure we are seeing net new businesses. This is one of the reasons why it's different now. We're seeing net new businesses with hundreds of millions in annual recurring revenue, that drive just on the generative quality. It's not bolted onto an old thing.
So I think we're kind of seeing all of these things. The least mature, in my opinion, is existing organizations introducing it into their operations. I don't think that we're at that point at a large scale yet." [1]
1. Workplace Technology (A Special Report) --- A VC Looks at Generative AI -- and Can Only Begin to Imagine the Impact: For one thing, says Martin Casado of Andreessen Horowitz, it will bring the cost of creating something new to near zero
Berber, Jin. Wall Street Journal, Eastern edition; New York, N.Y. [New York, N.Y]. 20 Mar 2023: R.9.
Komentarų nėra:
Rašyti komentarą