Better than humans? AI barrels towards AGI

Artificial intelligence (AI) breakthroughs are coming ever faster. AI technology is already found across a multitude of uses, from addressing climate change to exploring space, developing cancer therapies and providing real-world navigation for robots. The number of research papers focused on AI in recent years has grown so rapidly that it seems almost exponential. While we are still some ways away from widespread AI adoption across all spheres of human endeavor, it is safe to say the technology has now crossed the chasm between early adopters of new and little-known products and mass adoption by mainstream users. 

The most buzz-worthy AI breakthrough of the year is the new category of generative AI, which is based on large language models. Almost overnight, a proliferation of image generation tools appeared, including DALL-E from OpenAI, Imagen from Google, Stable Diffusion from Stability.ai and Midjourney. I wrote a few months ago about the disruptive impact of these tools on creative occupations, ranging from digital artists to programmers. 

As dramatic as these developments are, possibly more significant is the new conversational text bot ChatGPT, also from OpenAI and based on GPT-3.5. It has been trained on a massive amount of text data from a variety of online sources. Among other things, it can chat, answer questions, create plays and articles, write and debug code, take tests, manipulate data, provide advice and tutor.

ChatGPT has already been widely discussed online, including by savvy reporters Kevin Roose in the New York Times and Derek Thompson at the Atlantic. Thompson calls this and other recent generative AI tools a “second mind for the creative class.” Roose wrote that ChatGPT “is already being compared to the iPhone in terms of its potential impact on society.” 

However, ChatGPT is in its early days. Nearly everyone, including OpenAI, acknowledges that the tech is far from a perfected product, as evidenced by the “occasionally incorrect information” it generates. Nevertheless, as Jack Clark states in his Import AI newsletter: “In a few years, these systems might be better than humans, which is going to have wild implications.”

AI: New master of strategy

While these notable generative AI highlights are hugely important, several other recent AI developments may ultimately have even greater impact on the world. One example is the recent AI defeat of human experts in Stratego, a strategic war game for two players that requires long-term thinking, bluffing and strategizing. Deep Mind’s DeepNash algorithm, a trained autonomous agent that can develop human-level expertise, underpins the AI playing Stratego. DeepNash is based on an entirely new approach to algorithms using game theory and model-free deep reinforcement learning. 

Unlike chess and Go, Stratego is a game of imperfect information: Players cannot directly observe the identities of their opponent’s pieces. It is thought to be among the most difficult games, due to its seemingly infinite number of possible moves (a staggering 10⁵³⁵), more than even the notoriously complex Go (10³⁶⁰). To win, DeepNash mixed both long-term strategy and short-term decision-making like bluffing and taking chances, a unique capability for an AI.

As reported by Singularity Hub, the researchers stated: “In creating a generalizable AI system that’s robust in the face of uncertainty, we hope to bring the problem-solving capabilities of AI further into our inherently unpredictable world.”

The art of diplomacy

Speaking of unpredictability, Meta recently unveiled “Cicero” — an AI system named after the classical statesman and scholar who witnessed the fall of the Roman Republic — that bested people in another strategic war game, Diplomacy. 

Unlike Stratego, chess or Go — which are all zero-sum, winner-take-all competitions – Diplomacy is collaborative and competitive at the same time. Up to seven players compete, negotiating using deception and collaboration, trust and betrayal, to form and break alliances in pursuit of total domination. In other words, Diplomacy is much like real-life strategic negotiations among multiple competing entities, be they game players, businesses or countries. As reported by Gizmodo, “to ‘win’ at Diplomacy [the AI] needs to both understand the rules of the game efficiently [and] fundamentally understand human interactions, deceptions, and cooperation.”

This rich capability gets to the heart of what Meta was seeking to develop: “Can we build more effective and flexible agents that can use language to negotiate, persuade and work with people to achieve strategic goals similar to the way humans do?” The company claims Cicero achieved more than double the average score of the humans playing on webDiplomacy.net and ranked in the top 10% of participants who played more than one game.

Meta positions Cicero as a research breakthrough that combines two different areas of AI: strategic reasoning and natural language processing. According to three-time Diplomacy world champion Andrew Goff: “Cicero is resilient, it’s ruthless, and it’s patient.” He adds: “It makes the best decision, not only for itself but for the people it’s working with.”  

Generalized AI

“Narrow AI” incorporates algorithms that do only one thing, albeit extremely well — such as making a recommendation for what book you might like based on books you’ve previously viewed on an ecommerce site. A narrow AI algorithm cannot effectively transfer anything it has learned to another algorithm designed to fulfill a different specific purpose. 

The other end of the AI spectrum is deemed “strong AI” or alternatively, artificial general intelligence (AGI). Probably every AI expert would agree this does not exist today and remains in the realm of science fiction. If and when AGI is achieved, it would be a single AI system — or possibly a group of linked systems — that could be applied to any task or problem because it can act and think much like humans.

Murray Shanahan, a professor of cognitive robotics at Imperial College in London, said on the Exponential View podcast that AGI is “in some sense as smart as humans, and capable of the same level of generalization as human beings are capable of, and possesses common sense that humans have.” This sounds much like the capabilities of this new wave of strategy algorithms.

However, there is not a single AGI definition. For example, Elon Musk does not think that ChatGPT qualifies, as it hasn’t invented anything amazing: 

Toward artificial general intelligence (AGI)

By these criteria, at least, ChatGPT is not AGI, and neither are DeepNash or Cicero. What they all have in common, however, is a clear advance in this direction. As Stuart Russell, professor of computer science at the University of California and a leading researcher in artificial intelligence, notes: “The actual date of arrival of general-purpose AI, you’re not going to be able to pinpoint; it isn’t just a single day. It’s also not the case that it’s all or nothing. The impact [of AI] is going to be increasing. So with every advance of AI, it significantly expands the range of tasks.”

With each passing year, we can expect to see much greater capabilities on the march to AGI as these models become more sophisticated and new systems appear.

Given the pace and scope of recent AI breakthroughs and the huge growth in the number of research papers, we can expect developments to come ever faster with profound implications for work and life. For example, within several years, ChatGPT or a similar system could become an app that resembles Samantha in the movie Her. ChatGPT already does some of what Samantha did: an AI that remembers prior conversations, develops insights based on those discussions, provides useful guidance and therapy and can do that simultaneously with thousands of users. Or imagine NATO using tools like DeepNash or Cicero with its members or in negotiations with rivals.

We are witnessing a gathering momentum towards AGI, though experts’ estimate of its time of appearance is 2045. AGI or not, AI technology is becoming much more sophisticated and becoming deeply engrained in the fabric of our lives. 

Gary Grossman is the senior VP of technology practice at Edelman and global lead of the Edelman AI Center of Excellence.