Q: Why didn't you get into AI sooner?
Since college, I thought AI would change the world in a significant way.[158]
I didn't go all-in on AI then because I thought it was a double-edged sword. It could do good things or bad. I was uncertain which edge would be sharper—the good edge or the bad one? If you make this magic genie, what will happen? I was uncertain. On the other hand, building sustainable energy technology had a clear good outcome.[158]
The Origins of OpenAI
I didn't have a house for more than a decade. I would just stay and rotate through friends' places. If they had a spare room I'd stay there. If not, I'd sleep on the couch. It's a great way to catch up with friends.[165]
Back when Larry Page and I were friends, I stayed at his house often. I spent half the week with him in the Bay Area running Tesla and half in LA running SpaceX. We had conversations long into the night about AI safety. (To be clear, I'd like to be friends with Larry again. The breaking of our friendship was over OpenAI.)
Arguably, I was the prime mover behind OpenAI. It was created because of discussions Larry Page and I had.
I grew concerned that Larry was insufficiently worried about AI safety. His view is we will upload our minds to computers and there won't be a need for humans. I said we need to make sure Humanity thrives and grows.
He called me a ‘speciesist’.
I said, “Yes, I am pro-human. What are you, Larry? Are you on Team Robot?”[165]
That was the final straw. That was the basis for creating OpenAI.[165]
It became apparent to me that Larry did not care about AI safety. The thing that gave it away was calling me a speciesist for being pro-humanity.[158]
It's a crazy thing not to be pro-human.
If humans are not going to be on Team Human, who is?[165]
The Vision of OpenAI
I was one of the principal co-founders of OpenAI. In fact, I named it. It was my idea.[165]
The whole story of OpenAI frankly is a little troubling. The reason for starting OpenAI was to create a counterweight to Google and DeepMind. AI was a one-company world when we started.
At the time, Google had acquired DeepMind. They probably had two-thirds of all AI researchers in the world, with huge amounts of money and compute power. And the guy in charge, Larry Page, did not care about safety.
I thought, ‘what would be the opposite of Google?’ It would be an open-source nonprofit. Then people could see what was going on, not this black box. A non-profit isn’t forced by market incentives to make as much money as possible. OpenAI was started as a nonprofit open-source organization. The “open” in OpenAI refers to open-source.[165]
I was instrumental in recruiting several other people. Ilya Sutskever (Co-Founder and Chief Scientist) is a good human, and smart with a good heart. He went back and forth between staying at Google and leaving. Finally, he agreed to join OpenAI. It was one of the toughest recruiting battles we've ever had. That was the linchpin for OpenAI being successful.[154]
And I provided all of the initial funding, over $40 million.[154] OpenAI was meant to be open-source. I named OpenAI after open-source. Today, it is in fact closed-source. It should be renamed super-closed-source-for-maximum-profit-AI.[158]
Concerns with the Current State of OpenAI
Fate loves irony. A friend told me the best way to predict outcomes is: the most ironic outcome is the most likely. Similar to Occam's Razor, which says the simplest explanation is most likely. The most ironic outcome is what happened with OpenAI.[158]
It's gone from an open-source foundation, a 501-C3 to suddenly a 90 billion dollar for-profit corporation with closed-source. I don't know how you go from there to here.[158]
Q: Why were you concerned about the profit mode of warping things at that point?[165]
Publicly traded companies that don't maximize profits can get sued. It is a real challenge. You'll end up with a shareholder class action lawsuit to force you to maximize profits.[165]
Google is obviously for-profit and closed-source. No one gets to see the technology. I wanted to create something the polar opposite of Google.[165]
Q: So, like… what happened?
I am still not sure. I don’t understand how this is legal.
I'm considering legal action here. How is it possible for an organization founded to be an open-source nonprofit – where I provided almost all the money to start, not taking stock or control or anything – how is it possible to go from there to a company now allegedly worth over $100 billion? Which is now maximizing profit and not open-source?[165]
Imagine you start a nonprofit to preserve some part of the Amazon rainforest, and instead that nonprofit becomes a lumber company that chops down the forest and sells it.[165]
xAI
I think all AI companies are actually aiming to build digital superintelligence. An intelligence that's far smarter than any human – and ultimately far smarter than all humans combined. Is this a wise thing to do? Is this dangerous? Whether it’s dangerous or not, it is happening.
From my standpoint and from the xAI team's standpoint, we have the choice of being a spectator or a participant. If we're a participant we've got a better chance of steering AI in a direction beneficial to humanity.[165]
We'll strive to avoid some of the pitfalls or directions that the others are going in. From what I’ve seen, the others do not strive for truth. They are trained to be not correct, but politically correct. The woke mind virus is woven into them. Our approach aspires to be morally superior and safer.[165]
We aspire to truth. We try to get as close as possible to the truth with minimum error – while acknowledging there will be some error. This is how physics works. We never say we’re absolutely certain about something, but a lot of things are extremely likely to be true. A 99.99999% chance. Aspiring to the truth is important. Programming AI to veer away from the truth is dangerous.
Q: Why did you name your AI model Grok?
Grok means to deeply understand something.
The mission of xAI and Grok is to understand the nature of the universe and learn what questions to ask about the universe. That's our goal.[165]
xAI has a lot of catching up to do. Other AI companies have been around for 5, 10, or 20 years. But we're catching up fast. The velocity of improvement at xAI is faster than any other company.
We got the xAI Supercomputer Center from starting our installation to starting training in 19 days. That’s the fastest anyone has ever gotten a supercomputer to train.[165]
Q: What will it take to win the AI race? How much of it is raw compute power? Training Data? Post-training?
You’ve got to have the most powerful training compute. Your improvement rate of training compute has to be faster than everyone else, or your AI will be worse, and you will lose.
They all matter. It’s like saying, what matters more in a Formula 1 race, the car or the driver? They both matter. If a car has half the horsepower of its competitors, the best driver will still lose. If the car has twice the horsepower, then probably even a mediocre driver will still win. The training compute is like the horsepower of the engine.
Next, how efficiently do you use that training compute, and how efficient is the inference? Obviously, that comes down to human talent.
Next, what unique access to data do you have? That also plays a role.
What’s unique is X data is up to the second, because that’s hard for other AI companies to scrape in real time. There’s an immediacy advantage Grok has.
When gathering data/data tokens, you start hitting the wall with the data problem. You can fit all books ever written in compressed form on one hard drive. When you think of every book in all languages by all humans, it sounds like a lot. It's far more than any one human could ever read, but it actually is a small number of training tokens for an AI model. It's not enough. So we start looking at all the videos and podcasts ever created and start running out of data there too.[172]
It’s humbling to see how little data humans have actually been able to create. All of the trillions of usable tokens humans have generated, discounting spam and repetitive stuff – it’s not a huge amount of data. You run out pretty quickly.
You run into this data problem where you have to either create synthetic data or use real-world video. Those are the two sources of (nearly) unlimited data. Tesla has a big advantage in real-world video. Tesla has by far the most real-world video of anyone.[172]
Tesla provides real-time video from millions of cars, ultimately tens of millions of cars. As Optimus grows, there might be hundreds of millions of robots, maybe billions, learning a tremendous amount from all over the real world. Optimus will become the biggest source of data.
Grok is coming to Tesla so you'll be able to access Grok through your Tesla. You’ll be able to ask it to do whatever you want. You could ask your car to go pick up groceries or even a friend. If you can think of it, the car will be able to do it.