The Future of Image Models Is Multimodal

**Mohammad Norouzi** (0:00)
I do realize some conversations back then, okay, putting text to image into different productivity tools that Google had at the time, like for example, Google Slide. And the question was, okay, is this a hundred million dollar business? But it's harder to see it when it's not a hundred million dollar business. So I think fundamentally, it's just harder to do that kind of product innovation instead of big corporation, because there's a lot of flow hanging through it.
You might as well spend your time improving ads by like 0.001%.

**Derrick Harris** (0:34)
Hi, this is Derek Harris, and you're listening to the A16z AI Podcast, where we're focused on the people building at the cutting edge of artificial intelligence. That's a description that perfectly describes this week's guest, Ideogram co-founder and CEO, Mohammad Norouzi. He spent about a decade as an AI researcher at Google, working on MLP translation and computer vision, ultimately culminating his work on Google's imagined text-to-image model. Joining me to speak with Mohammad is A16z general partner, Jennifer Lee. We cover the gamut of Mohammad's journey, from drawing pictures as a young child in Iran, to competitive programming in college, to what he's learned from starting a company that lets anybody create their own pictures. It's an insightful conversation, and if you're interested in learning, a good introduction to some of the key concepts behind text-to-image models.
Enjoy the discussion.
As a reminder, please note that the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a16z fund. For more details, please see a16z.com/disclosures.

**Mohammad Norouzi** (1:55)
When I was a little kid, I didn't go to kindergarten or daycare. All I remember is wandering around my grandparents' backyard, listening to stories and making drawings.
In high school, I fell in love with math and programming. And the context is we have this national entrance exam at the university, and if you make it to the Olympia teams in informatics or mathematics, you get to skip the exam. And so all I did was really compete with a bunch of other high achieving high school kids to solve programming and math problems. And it was a lot of combinatorics, geometry and graph theory, but I ended up going more towards programming. And then I made it to the team.
I did some competitive programming at the time, and then made it to the university because I got to skip that exam. And then in the university, I think I had a lot of the algorithms and data structures backgrounds. I didn't do much of the studying. Most of the things I did was building these web applications for competitive programming again. So that was being used for the next generation of high school kids trying to compete in programming. It was mostly a solo project. I didn't think of turning that into a business. And then towards the end of university, I was thinking about what I want to do next. Most of my friends went into theoretical computer science just because of the math background, but I felt like I want to do something more practical. And I self-taught myself neural networks, backpropagation.
And actually, I read Jan LeCun's paper on convolutional neural nets back then when it wasn't popular at all.
And I started implementing backpropagation in Java, actually, and trained models on MNIST. So this was kind of interesting that I was just reading academic papers for the first time and then figuring it out all by myself.

**Derrick Harris** (3:59)
What time frame was that, can I ask?

**Mohammad Norouzi** (4:00)
This is 2007

**Derrick Harris** (4:04)
Okay, so you're several years ahead of the curve.

**Mohammad Norouzi** (4:07)
Yeah, this is way back then.
And then I was thinking, maybe I want to get into cognitive science and figure out more about how the brain works because a lot of the commentary at the time was a similarity between the neural nets and the brain.
And I tried that. I got into a master's program in Canada and tried doing a bit of cognitive science, but it wasn't for me. So I switched quickly to computer vision, machine learning, stuff like that. And from the very beginning, people were asking me, are you interested in computer vision or natural language processing? Because these were the two applied fields around machine learning. It didn't make sense for me to kind of separate the two. So I was doing a bit of an LP, a bit of computer vision, and then did a PhD in Toronto and kind of application of neural nets for larger scale similarity search. But then I went to Google and it was early days again of building deep neural nets. I was on the brain team and I got to work on a whole bunch of exploratory projects, but I have to dive into that.