How AI Will Unleash a New Kind of Intelligence | Ken Ono, Axiom Math | EO

[gentle music] [keyboard clacking] The very first time I heard the term artificial intelligence was in 1993, and I met a faculty member who said, "I work in artificial intelligence." My very first words were, "Oh, that's interesting.

Um, I specialize in natural intelligence." And I was cocky.

I would have my butt handed to me many times for using words like that, but that's where I came from.

Exactly one year ago, I was a happy-go-lucky university professor writing my papers, enjoying life at the University of Virginia, and then there was a dramatic change.

I came face to face with large language models at work.

The Frontier Math Program, where this company based in Berkeley called Epoch AI hired professional mathematicians from around the world to assemble very difficult math problems, and their goal was to assess the capabilities of large language models as, as they improve.

For the first time, I struggled to assemble questions that ChatGPT would get wrong.

I was devastated.

At the time, I was one of the few scientists who had been given access to these state-of-the-art models, and I had to be patient for a few months to go by before the world did recognize that, yeah, these models know so much.

These models know more facts than any human you would ever find.

So for a few months, I was devastated.

I was thinking, "How am I going to stay ahead of AI?" But I actually think that's the wrong question.

If our goal is to always stay a- ahead of AI, then, uh, I think we're gonna lose.

Nobody would be interested in watching Usain Bolt race against a motorcycle in the one-mile run.

It's not a fair race.

But we still watch the Olympics.

We, as a society, now know how to accept that machines can outperform humans in every physical way, but we're still now coming to grips with the fact that the brain, deep inquiry, computers have caught up.

The large language models should be thought of as the most extraordinary librarian the world has ever seen.

If it has been written down, the large language model has probably seen it.

If it's on YouTube, large language model has probably been trained on it.

If you read a newspaper article, by the afternoon, the large language model has probably seen it.

Good luck with competing with that ability to collect information.

It- information, knowledge is now cheap.

But how you use it and how you verify it has become more expensive.

Do you want your librarian to be your neurosurgeon? Do you want your librarian to be your air traffic controller, somehow keeping an eye on the hundreds of planes that are flying over North America or Korea? No way, because that human judgment is important.

[gentle music] [keyboard clacking] The very first time I heard the term artificial intelligence was in 1993, and I met a faculty member who said, "I work in artificial intelligence." My very first words were, "Oh, that's interesting.

Um, I specialize in natural intelligence." And I was cocky.

I would have my butt handed to me many times for using words like that, but that's where I came from.

Exactly one year ago, I was a happy-go-lucky university professor writing my papers, enjoying life at the University of Virginia, and then there was a dramatic change.

I came face to face with large language models at work.

The Frontier Math Program, where this company based in Berkeley called Epoch AI hired professional mathematicians from around the world to assemble very difficult math problems, and their goal was to assess the capabilities of large language models as, as they improve.

For the first time, I struggled to assemble questions that ChatGPT would get wrong.

I was devastated.

At the time, I was one of the few scientists who had been given access to these state-of-the-art models, and I had to be patient for a few months to go by before the world did recognize that, yeah, these models know so much.

These models know more facts than any human you would ever find.

So for a few months, I was devastated.

I was thinking, "How am I going to stay ahead of AI?" But I actually think that's the wrong question.

If our goal is to always stay a- ahead of AI, then, uh, I think we're gonna lose.

Nobody would be interested in watching Usain Bolt race against a motorcycle in the one-mile run.

It's not a fair race.

But we still watch the Olympics.

We, as a society, now know how to accept that machines can outperform humans in every physical way, but we're still now coming to grips with the fact that the brain, deep inquiry, computers have caught up.

The large language models should be thought of as the most extraordinary librarian the world has ever seen.

If it has been written down, the large language model has probably seen it.

If it's on YouTube, large language model has probably been trained on it.

If you read a newspaper article, by the afternoon, the large language model has probably seen it.

Good luck with competing with that ability to collect information.

It- information, knowledge is now cheap.

But how you use it and how you verify it has become more expensive.

Do you want your librarian to be your neurosurgeon? Do you want your librarian to be your air traffic controller, somehow keeping an eye on the hundreds of planes that are flying over North America or Korea? No way, because that human judgment is important.