Who Is Bob McGrew? The Epicenter of the AI Revolution

Bob McGrew is one of the key players in the current Artificial Intelligence wave. He implemented the tech and techniques needed for ChatGPT to absolutely dominate the generative AI industry. And as the VP of Research at OpenAI, he has a front row seat to the progress, power, and potential of AGI. 

Given how motivated and resilient he was ever since he was young, his multiple achievements and contributions should come as no surprise. So let’s take a quick look at this overachiever’s journey thus far. 

A Natural Born Problem-Solver

Ever since he was young, Bob loved solving problems.

His parents were both college professors. His dad, in particular, taught Computer Science. One day, when Bob was in 3rd grade, his dad bought him a book about two kids using computer programs to help detectives solve cases. 

Since then, Bob has been hooked. 

He continued buying the books in these series and writing computer programs inspired by them. This was in “small town Oklahoma,” as he put it, during the first big web boom. “I always knew I wanted to go to Stanford,” Bob says in an interview on American Optimist. “I never actually thought I’d get in. But I did!”

He got into Stanford, got his PhD in Artificial Intelligence, and continued doing what he loved most: writing computer programs to solve problems.

From PayPal to Palantir

Bob worked cryptography in PayPal from 2001 to 2002. He eventually left to pursue a career in AI. Unfortunately, during this time, AI wasn’t the powerhouse it is now. It wasn’t as feasible, accessible, or even as remotely refined. The data, tech, and knowledge that could possibly help Bob in his career were simply not available.

Regardless, Bob pushed on. He joined Palantir–an independent research firm–and became one of their key leaders. 

Shyam Sankar, Palantir’s Chief Technology Officer, describes Bob as “legitimately brilliant.” He says Bob’s excitement when he’s faced with a new and interesting challenge to solve is visually palpable.

Joe Lonsdale, the host of American Optimist and an old colleague of Bob’s in Stanford, even points out how Bob’s shift to Palantir still allows him to practice his old hobby of solving cases with computer programs. 

Bob admits to the crazy coincidence but definitely isn’t complaining. 

Palantir Gotham

Bob’s run with Palantir was fulfilling. But it definitely wasn’t easy.

One of his biggest challenges early on was Palantir Gotham–an operating system that is “commercially-available” and “AI-ready.” It takes data from any number of systems and presents it in multiple visual formats to help analysts interpret the information quickly and more efficiently. 

The idea is to help them produce actionable intelligence based on the integrated data, thereby “improving and accelerating decisions for operators across roles and all domains.”

But before it became the refined, cutting-edge tech platform it is today, it started out as a concept Bob and his team struggled with–simply because they didn’t know what it was supposed to be. Ideally, Palantir Gotham was meant to be a platform for the intelligence community (or “software for spies,” as Bob puts it). 

Unfortunately, since most of the information that they needed from their clients was “classified,” they didn’t know what problems they needed help solving. Ergo, Bob and his team could only build programs based on what they thought the client wanted.

And, most of the time, they guessed wrong.

The Need for Resilience

Another issue Bob faced in Palantir was uncertainty. 

According to him, it took three whole years for their team to finally release a working product that the clients actually used. And there were many, many people who quit right before this milestone was achieved. 

Bob didn’t blame them, either. As the VP of Engineering, he spent most of his time figuring out if the right people were working on the right problems. He would work on his own projects from time to time (mostly a lot of backend) but, for the most part, he was focused on other people. So he could see their frustration first-hand. 

“A really great engineer wants certainty,” he pointed out. If someone does the same thing for three years, it stands to reason that all they’re going to see are the innate problems.

Bob also remembers sitting down and having a conversation with one of his best engineers. He managed to convince him to stick to the project for nine more months. And that engineer ended up building the very security system Bob’s team needed to deploy their product. It was a critical piece to the entire project. 

Yet the engineer still ended up quitting before the final working product was launched! 

The Journey Towards AGI

In 2016, Bob joined OpenAI. He recalls that their first “crazy research project” was to find a way to beat Dota 2–a multiplayer online battle arena game by Valve. 

They were stuck on it for a while until Jakub Pachocki (the Director of Research in OpenAI) developed a technique that allowed engineers to put a huge amount of data and compute into solving a single problem. 

Since it was a huge help to the Dota 2 problem, Jakub brought this technique over to the robotics department, where Bob was.

The Robot and the Ball

Bob’s on-going project at that time was trying to get a two-fingered robot claw to grab a ball. Unfortunately, despite multiple attempts, he couldn’t get it to work.

He recalls Ilya Sutskever, Co-Founder and Chief Scientist of OpenAI, sitting him down and telling him that they were going to build AGI (Artificial General Intelligence) programs that were going to run at human-level intelligence. He admits to being skeptical.

“Why are we talking about AGI if we can’t even have a claw grab a ball?” Bob remembers thinking. “This is really easy. We should be able to do this.”

And then Jakub introduced his new Dota-2-beating technique and Bob immediately saw the potential. He implemented it right away.

“In two weeks, we went from not being able to grab a ball with a claw to having a five-fingered humanoid robot hand solving [a] rubik’s cube in simulation.” He said excitedly. “And I was like, wow!”

This breakthrough is what convinced Bob that AI and AGI were possible. 

Neural Networks: The Possible Key to OpenAI

Bob saw the potential of general-access AI even before it was technologically feasible. He realized early on that the key to AI was big data. 

He credits two papers for AI’s massive leap forward. The first one, published by Ilya Sutskever, was about reinvented neural networks. They had found a way to run them on GPUs so that large amounts of compute and data could be plugged in. 

Bob even states this on American Optimist. “If you can pour enough compute and pour enough data into a neural network,” he says “there’s no limit to how good it can get.”

This was OpenAI’s thesis when Bob first joined. They predicted that neural networks would be the final architecture that could AI all the way to human-level intelligence. All they needed was a way to beef it up. And they found one.

The second paper was written by someone at Google. They’d come up with the idea of “transformers,” or neurons that could basically allow programs to have a better understanding of context. Instead of functioning by remembering the data they were fed–which is how standard networks worked–transformers made networks pay attention to the data they were being fed so that they could predict the next best action. 

This was in 2017. And it was a game-changer.

Ushering in the Future

Bob’s implementation of big data and transformers into neural networks revolutionized AI. He and his team at OpenAI have redefined society and the tech landscape with programs like ChatGPT3, ChatGPT4, and DALL-E.

And Bob believes that this is just the beginning.

According to him, OpenAI is still a ways away from achieving human-like intelligence. What it currently does is predictions. Predictions per element, per character, happening in less than a second. When put together, it resembles something an awful lot like intelligence. 

But it’s not. And Bob won’t be satisfied until they achieve the genuine thing–and then maybe even surpass it.

“It’s very clear to me,” he says, “and I think it’s clear to people who are working in the field that progress is going to continue to happen.” GPT4 may be the most advanced model to date, but GPT5, GPT6 and so on are definitely going to happen. And it’s possible, Bob says, that we may eventually need AI to teach itself. Because it can get to a point where learning by mimicking humans no longer works.

“I think that’s the open problem that everybody sees right now. The current GPTs are just mimicking what humans do. How do you unlock creativity?”

A loaded question to be sure. But if there’s anyone who can find the answer to that, we’re certain it would be Bob McGrew.