Discover more from The Gap
Centaurs, OpenAI’s Push for AGI & AI Agents
The factory of the future will have only two employees, a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment.
— Warren Bennis
Shortly after Gary Kasparov's paradigm-shifting defeat to IBM's chess engine DeepBlue in 1997, Kasparov invented a new variant of the game called "advanced chess" also known as "centaur chess".
In advanced chess, a human player is working together with a computer to cooperatively decide on the best moves. Kasparov believed that human-AI teams which he referred to as "centaurs" would constitute the strongest chess players in the future. And for a few years, he was right.
Today, however, chess engines are so powerful that the best players could consist of a computer, a man, and a dog. The man's role would be to feed the dog, while the dog's role is to stop the man from touching the computer. In other words, centaurs' dominance on the chess boards was relatively short-lived; humans assisted by AI can no longer compete with the best chess engines on the market.
The inflection point came in 2017 when DeepMind's AlphaZero defeated the best chess engine at the time, Stockfish, in 100 out of 100 games. AlphaZero achieved its win without studying as much as a single human chess game from the past. Instead, it was trained with a machine-learning method called reinforcement learning (see paper here). It played against itself over and over again for 24 hours straight, iteratively learned from its mistakes, and eventually figured out the best moves and underlying patterns of the game.
OpenAI’s Push Towards AGI
Today, AI models are no longer exclusively used in narrow domains like chess-playing, but in increasingly more generalized contexts. The path toward more general intelligence has been pushed aggressively by OpenAI ever since its release of the first Generative Pretrained-Transformer, GPT-1, in June 2018.
Just over the last few months, OpenAI has introduced customizable GPTs, the more capable and cheaper language model GPT-4 Turbo that can accept prompts of up to 128K tokens (roughly corresponds to a medium-sized novel), new voice and image capabilities in ChatGPT, and the new image generation model, DALL-E-3. The dizzying pace of OpenAI’s product developments is far from coincidental but part of the company’s larger mission to achieve artificial general intelligence (AGI) before its competitors and take charge of its direction.
Of course, we cannot really talk about OpenAI today without addressing the recent developments. I will not pretend to have any special insights but one of the best accounts I have come across is from Matt Levine’s opinion column in Bloomberg that I have linked to in my “Reads of the Week” section at the end.
Long story short: after a dramatic and messy weekend at OpenAI, co-founders Sam Altman and Greg Brockman are no longer a part of the company but instead heading Microsoft’s “new advanced AI research team” and bringing at least three senior researchers from OpenAI with them, including GPT-4’s leading scientist Jakub Pachocki. How all this will affect the companies’ relationship is unknown. Microsoft has invested $13 billion in OpenAI and the two companies are so financially intermingled anyway that it's hard for a bystander to see where one company begins and the other one ends.
With all that has been happening, OpenAI’s future currently seems to be up in the air. Whether it will continue its laser-focused sprint towards the goal of AGI is uncertain. But even if it does not, Microsoft and other big companies will certainty continue in its spirit.
If we try to peek into the crystal ball and foresee the next big development in AI, what could it be? Probably AI agents. We have seen how AI has become more generalized in recent years. From DeepMind's neural-network-based chess champion, AlphaZero, to OpenAI's GPT-4 which is used for a wide range of tasks across all sorts of domains. It's easy to imagine that the next step in AI’s evolution will be increased agency; AIs that can increasingly act by their own accord without human hand-holding. According to a Bill Gates post from earlier this month, AI agents will revolutionize the software industry in the next five years:
“To do any task on a computer, you have to tell your device which app to use. You can use Microsoft Word and Google Docs to draft a business proposal, but they can’t help you send an email, share a selfie, analyze data, schedule a party, or buy movie tickets. And even the best sites have an incomplete understanding of your work, personal life, interests, and relationships and a limited ability to use this information to do things for you. That’s the kind of thing that is only possible today with another human being, like a close friend or personal assistant.
In the next five years, this will change completely. You won’t have to use different apps for different tasks. You’ll simply tell your device, in everyday language, what you want to do. And depending on how much information you choose to share with it, the software will be able to respond personally because it will have a rich understanding of your life. In the near future, anyone who’s online will be able to have a personal assistant powered by artificial intelligence that’s far beyond today’s technology.”
If Gates is correct, AI cannot only change how we interact with computers in the near-term future but it can also have complex and potentially devastating effects on the job market. So far, most white-collar jobs have been safe from automation after the advent of ChatGPT. AI can still not perform well on essential tasks that involve social interaction, adaptation, flexibility, and communication.
Yet, some tech leaders have an insatiable appetite to automate work. Just look at OpenAI’s mission to build “highly autonomous systems that outperform humans at most economically valuable work” or Sam Altman who on several occasions has explained his vision of AGI as “the equivalent of a median human that you could hire as a co-worker”.
Currently, ChatGPT and other generative AI products are often hailed as “productivity boosters” and “skill equalizers”. Next week’s topic will be a Harvard Business School Paper where the authors enthusiastically highlight how elite consultants significantly and unambiguously benefit from using GPT-4 for certain tasks. But as the lesson from AI's evolution in professional chess tells us, centaurs can eventually be defeated by more automation…
Reads of the Week
Who Controls OpenAI - Matt Levine, November 20, 2023 (Bloomberg).
Advertisers Flee X as Outcry Over Musk’s Endorsement of Antisemitic Post Grows - Ryan Mac, Brooks Barnes and Tiffany Hsu, November 17, 2023, (New York Times).
Exclusive: Germany, France and Italy reach agreement on future AI regulation - Andreas Rinke, November 20, 2023 (Reuters).