March 22, 2023
At a Glance
- OpenAI cofounder and chief scientist Ilya Sutskever talks to Nvidia CEO Jensen Huang about his research journey up to GPT-4.
- Ilya Sutskever admits surprise over GPT-4's reliability but says more work needs to be done on reasoning.
In late 2015, a group of AI researchers and Silicon Valley notables founded a nonprofit research lab, OpenAI, with the goal of creating artificial general intelligence. Among that group was Ilya Sutskever, who would go on to become OpenAI’s chief scientist. Fast forward eight years and the company he helped co-found would become the darling of the AI world and his work in deep learning would receive acclaim.
OpenAI's GPT-3, DALL-E, and ChatGPT are now recognized and used the world over. But in the early days of OpenAI, the techniques that gave rise to these models were a distant dream. Back then, Yann LeCun, the father of convolutional neural networks and now chief scientist at Meta, gave speeches urging folks to solve the field's grand challenge: develop unsupervised learning, or training AI models from unlabeled datasets.
Today, unsupervised learning is a valuable technique developers use to let them plug in and pre-train.
In a fireside chat with Nvidia CEO Jensen Huang, Sutskever recalled that he became convinced good data compression was the solution to unsupervised learning. “Intuitively, you can see why (it) should work. If you compress the data really well, you must extract all the hidden secrets which exist in it. Therefore, that is the key."
It was this outside-of-the-box thinking that would catapult Sutskever and the OpenAI team into developing what has been billed as a new computing platform poised to disrupt every industry.
Groundwork of key techniques
What allowed OpenAI to truly scale itself, however, was reinforcement learning – where a computer program learns to take actions to maximize its rewards. Sutskever and the team used this technique to train an AI agent to play the video game Dota 2 against itself to meet the goal of reaching a level so that it could compete against the best players in the world. The agent, OpenAI Five, was first unveiled in 2017 and went on to draw comparisons to landmark AI moments such as IBM’s DeepBlue defeating Chess grandmaster Garry Kasparov.
That groundwork of techniques, reinforcement learning and unsupervised learning, gave rise to work spanning sentiment neurons and eventually, GPT models, which would go on to make OpenAI a household name thanks to one model, ChatGPT.
Huang described GPT-4’s performance as “astounding,” praising its ability to achieve “very human levels” on tests like the bar exam. He also called ChatGPT “the easiest application that anyone has ever created for anyone to use.”
Sutskever said GPT-4 was “a substantial improvement” on ChatGPT. The OpenAI team began training on the model some six to eight months ago. In that time, GPT-4 has able to predict the next word in a sequence with far greater accuracy than ChatGPT, he said.
“This is really important because the better the neural network can predict the next word in the text, the more it understands it," Sutskever said. However, "this claim is now accepted by many at this point. But it might still not be intuitive or not completely intuitive as to why that is."
This gap in understanding has led some AI leaders to believe large language models lack reasoning. LeCun said in February that ChatGPT and other large language models have “no knowledge of the world around them, … they don’t know the world exists. And they have no idea of physical reality. They don't have any context and they can't find their answers.”
But Sutskever argued that reasoning as a concept in AI is poorly defined.
“(For) our neural nets, maybe there is some kind of limitation which could be addressed by, for example, asking the neural network to think out loud. This has proven to be extremely effective for reasoning. It also remains to be seen just how far the basic neural network will go.”
“I think we have yet to fully tap out its potential. There is definitely some sense where reasoning is still not quite at that level. As with some of the other capabilities of the neural network, we would like the reasoning capabilities of the neural network to be higher. I think that it is fairly likely that business as usual will keep improving the reasoning capability.”
Biggest surprises from GPT-4
Sutskever admitted that he was surprised by the extent of GPT-4’s reliability in initial testing. GPT-3, like most large language models, suffered from hallucinations – where generated outputs are false or wildly incomparable to the initial prompt. He said such behaviors “stopped happening” with GPT-4. Sutskever was also impressed by the model’s arithmetic capabilities and ability to explain things like jokes and memes.
The biggest surprise, however, was that GPT-4 “actually works,” Sutskever said.
“It has turned out to be the same little thing all along, which is no longer little and a lot more serious and much more intense,” he said. “But it is the same neural network, just larger, trained on maybe larger datasets in different ways, with the same fundamental training algorithm.”
Read more about:ChatGPT
About the Author(s)
You May Also Like