prg․ai meetup: The Evolving Role of Games in AI by Murray Campbell

Games have been at the center of researchers’ attention since the first serious attempts to program a thinking machine in the 1950s. For a long time, researchers thought that games like chess were the key to uncovering the secrets of intelligence in order to replicate it. And while DeepBlue beat Kasparov more than a quarter of a century ago (and chess programs have achieved a significantly super-human performance since then), there are still many unknowns in what makes an entity intelligent.

What is a game, really?

Murray Campbell, from the AI Foundations group at IBM, was right at the center of the action, being the hand of the computer and sitting opposite Garry Kasparov, so he is well aware of the discrepancy. He starts with a more fundamental question: What is a game, and why are games useful? Games are quests in which we try to overcome burdens. They are an environment where we make interesting, informed decisions. But, most importantly, games have a well-defined set of rules, actions, and goals (unlike in real life), which makes them useful for AI researchers. They are an excellent sandbox for trying out our algorithms. However, now and then, researchers hit the (computational) limit of what kind of games are possible to study and how the results can be used elsewhere. Murray shows us how this is manifested in the interest of the AI community in games versus other techniques.

Chess and general intelligence

Going back to chess, one of the reasons that neither Deep Blue nor AlphaZero brought us any closer to general artificial intelligence is that, while chess is challenging to master for humans, for a computer, it is a very simple environment that can be solved without much complex reasoning. This made many scientists overly optimistic in the 1950s and 60s. However, Murray explains that “the only” aspect that is hard about chess is the vast size of the state space, which results in high hardware demands that were not practically satisfiable until the end of the last century. Still, with enough computational power, we are left with simple programs that neither look nor perform like human brains.

The realization that we are nowhere near human intelligence led to multiple shifts in the AI paradigm throughout the last century. From reasoning to expert systems to the current “learning era,” Murray thinks that while learning will stay vital in the future, we are still missing some critical (possibly symbolic) ingredients in our AI programs.

Human-AI collaboration and the future of games in AI

With the growing engagement of AI in our daily lives, games can become interesting benchmarks showing us the possibilities and limits of human-computer or human-AI collaboration. For a concrete game, Murray comes up with a division of time into four eras. In the first phase, computer programs are much weaker in the game and do not add much value to the human experts. In the second phase, programs become good at specific capabilities with which they can assist human experts. In the third era, AI has generally become better than humans; however, there are certain gaps in its performance, which experts can avoid. Finally, the fourth phase marks a stage where any human input is practically meaningless, as the AI program becomes significantly better in all aspects of the game and its decisions incomprehensible to the expert.

This is a fitting parallel to any human conduct that is tackled by AI algorithms and leads to a realistic yet sobering realization. The possibilities of human-AI collaboration have an expiration date, after which humans will be evicted from their respective fields by an overperforming AI. Murray finishes by advising to choose a career in a field where the time interval and possibility of collaboration with AI is as big as possible.

Apart from this parallel, Murray suggests that interactive games (such as text-based adventures) might become an important testing bed for general algorithms that perform very well even on unseen games. This way, games could also replace some of the well-established machine learning benchmarks that cause many issues today.

Q&A highlights

• On comparing computers playing chess to ChatGPT writing essay, Murray points out that while human players continue to enjoy chess as a fun activity, the necessary and often mundane activities (which, however, help us to develop ourselves), such as writing, are in danger.
• One of the reasons why the complexity of AI was underestimated in the past was a misunderstanding of the importance and complexity of system-1 thinking (as defined by Kahneman), claims Murray. Researchers concentrated on the system-2 reasoning and failed.
• Transferring knowledge between computer programs and humans is hard, and explainability is needed. Sometimes, even the best suggestions from AI are useless without understanding the reasons.
• Murray’s current work is focused on adding reasoning to AI algorithms and creating benchmarks that reliably detect whether an AI model (e.g., an LLM) is capable of reasoning.

The author of the text is AI researcher at FEE CTU Martin Krutský.