The AI singularity: what to expect when machines outsmart us. Thoughts following a lecture by Professor Toby Walsh

The AI singularity: what to expect when machines outsmart us.

Wednesday, 19 February at the University of Sydney

The quote that set the scene

Just as electricity transformed almost everything 100 years ago, it is challenging to think of an industry that AI will not transform in the next few years (Ng, 2017). Andrew Ng, former chief scientist at Baidu and co-founder of Coursera, famously made this statement to underscore the revolutionary impact of AI.

I went to see Toby Walsh at the University of Sydney to hear a lecture on the impending AI singularity. I had expected the room to be packed, given that this topic is arguably one of the most important issues facing society today. However, the turnout was surprisingly low. I would be interested to see the statistics on how many people watched it online.

AI in Chess

Toby really seemed to hit his stride when discussing games like chess and the Chinese game Go. He explained that when IBM’s Deep Blue defeated Garry Kasparov with a highly unexpected, almost creative move, Kasparov was so unsettled that he went on to lose a game he could otherwise have drawn (Walsh, 2018).

AI in Go

Toby then described how AlphaGo, an AI developed by DeepMind, defeated the Go master Lee Sedol with a similarly unexpected move (Silver et al., 2016). At first glance, these moves appeared almost creative in their originality. However, subsequent analysis revealed that the AI was simply playing a remarkably long game. The vast amount of data and number-crunching capabilities at its disposal made such a move seem original, yet it was actually part of a carefully calculated plan based on statistical analysis.

“Creative” Moves and Strategy

One question this raises is how advanced AI would need to be to analyze a winning strategy by making a move so unexpected that it psychologically unsettles an opponent. It is not unreasonable to imagine software identifying winning strategies and then executing them in highly creative ways.

The Limits of “Book Learning” and Embodiment

Another intriguing point Toby raised concerned the recent demonstrations of AI passing exams, such as the bar exam or first-year medical exams. On the surface, these achievements seem remarkable, but in many cases, the AI had access to vast repositories of past questions and answers, effectively allowing it to cross-reference previously encountered material (OpenAI, 2023). Examples like these demystify AI, revealing it as a computational tool rather than a self-contained, embodied entity.

The topic of disembodiment arose during the Q&A session, where Toby pointed out that an AI could not discover a physical law like gravity by literally tripping over something and hurting itself, as humans might. This lack of embodiment means that AI cannot experience the real world in the same way humans do. Consequently, many of its actions are not grounded in physical reality, a point often overlooked but crucial for students to understand (Brooks, 1990).

Cooperation, Competition, and Global Threats

Toby also voiced concerns about global threats, comparing them to the risk of nuclear Armageddon or laboratory-developed pathogens. He highlighted that AI lacks certain human factors—namely cooperation and competition—that shape how societies function. Cooperation enables people to work together to achieve complex goals, exemplified by the fact that no single individual can build an iPhone alone. Meanwhile, competition in capitalist systems reduces the likelihood of monopolies forming, as regulations promote rivalry among companies (Walsh, 2018).

The “Wicked Tsunami” of AI

Toby warned of a “wicked tsunami” of challenges, stemming from two surprising aspects of AI that he has observed over four decades of research. First is the sheer scale of the “gold rush,” with billions being invested annually in AI development. Second is the incredible speed of breakthroughs arising from these investments. He compared AI to historical innovations like electricity or the internet, which required decades to build the necessary infrastructure and regulations (Schwab, 2016). By contrast, AI advancements occur so rapidly that society struggles to keep pace.

Consciousness and Embodiment

I was struck by the potential limitations of AI, particularly regarding embodiment, which may remain an unattainable goal. On the other hand, I am still uncertain about embodiment, given that AI can perceive the world through sensors yet may not truly comprehend it. Perhaps consciousness is the missing piece. Toby remarked that if AI never attains consciousness, we can simply use it for tedious tasks, but if it does become conscious, that presents an entirely different scenario (Walsh, 2018).

Job Displacement and Education

One key confirmation from Toby’s talk was that the “tsunami” of problems includes significant job displacement, likely to occur in a short timeframe. He noted that a society with large numbers of unemployed individuals could begin to break down (Frey & Osborne, 2013). While it is uncomfortable to consider, this must factor into discussions about curricula and what we teach in schools to prepare future generations for rapidly changing job markets.

As AI approaches singularity, the world needs new benchmarks.

References

• Brooks, R. (1990). Elephants Don’t Play Chess. Robotics and Autonomous Systems, 6(1–2), 3–15.

• Frey, C. B., & Osborne, M. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin School, University of Oxford.

• Ng, A. (2017). AI is the New Electricity [Video]. Stanford Graduate School of Business.

• OpenAI. (2023). ChatGPT [Large Language Model]. https://openai.com/blog/chatgpt/

• Schwab, K. (2016). The Fourth Industrial Revolution. Crown Business.

• Silver, D. et al. (2016). Mastering the Game of Go with Deep Neural Networks and Tree Search. Nature, 529(7587), 484–489.

• Walsh, T. (2018). 2062: The World that AI Made. Black Inc.