Artificial Intelligence vs. an IMO Medalist?
Written By: Vidya Sinha
Artificial intelligence models such as Chat GPT are often criticized for their mediocre problem solving abilities. Although they can effortlessly churn out undergraduate-level essays and regurgitate vast amounts of information, they often fail in the face of tasks that demand creative thinking.
The International Math Olympiad (IMO) is one of the most cognitively demanding competitions for high-schoolers. Its problems often require ingenuity and sophisticated reasoning skills, presenting robust immunity against mindless brute-force solving tactics. This is precisely the sort of task that would evade mechanical algorithms. Or is it?
Recently, a team led by Google DeepMind and New York University introduced AlphaGeometry, an AI program designed to solve IMO geometry problems. What were the results? On a test-run, the program solved 25 out of the 30 geometry problems. A human IMO gold medalist solved 25.9, barely scraping a victory for humankind.
So what about the program permits it to vastly surpass the problem solving skills of existing language-processing models? The answer lies in the algorithms. Let’s open up the black box and peer inside.
Neural Language Models
Neural language models are the mechanism underlying Chat GPTs eerily human-like conversations. In short, they involve leveraging neural networks to assess the likelihood of arbitrary sentences.* This entails using a process of “self-attention” to predict the relationships between words in a sentence and converge on a plausible continuation of the sentence. By being trained extensively, GPTs can absorb patterns in human language and generate language efficiently.
The caveat of this model is that it primarily rests on pattern recognition rather than explicit cognitive algorithms. GPTs simply imitate the human mind via observation rather than direct encoding of reasoning algorithms. They are incredibly fast, but lack the nuanced “understanding” of ideas needed to solve difficult problems.
*If you’re unfamiliar with neural networks, go read our article about neural networks!
Symbolic Deduction Engines
In contrast, symbolic-deduction engines use formal logic to arrive at conclusions through rigorous reasoning. This involves translating sentences into logical symbols and then operating on these sentences in correspondence with the principles of logic. They are cumbersome compared to their language-predicting counterparts, but they generally produce accurate conclusions.
AlphaGeometry
AlphaGeometry combines the rapid-fire power of neural language models with the deliberate reasoning of symbolic deduction engines to tackle IMO problems. In this sense, it mimics the System 1 and System 2 systems of human cognition, where System 1 comprises our innate reflexes and System 2 refers to rational judgment.
Takeaway
Artificial intelligence is approaching human-level intelligence at breakneck speed. As researchers formulate more sophisticated models of Artificial Intelligence, it is not implausible that Artificial Intelligence may surpass human intelligence in the near future. Many of the current limitations of AI, such as “utilizing common sense” and “detecting context clues” may be attributed to a dearth of training data; if an artificial intelligence hypothetically had access to every piece of information known to mankind, it seems highly likely that this discrepancy would disappear.
However, one quality that artificial intelligence currently does lack is a fully functional notion of human ethics. Would you want a person with the intelligence level of an IMO Gold Medalist and a sense of disregard for human life to control any aspect of your society? Presumably not! Although artificial intelligence is marvelous, the superintelligent descendants of AlphaGeometry may wreak havoc if not perfectly aligned with the goals of mankind.
It is imperative that we take meticulous steps to wield artificial intelligence appropriately. The future of humanity depends on it.
Works Cited
AlphaGeometry: An Olympiad-level AI system for geometry. (2024, January 17). Google DeepMind. https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
Bischoff, M. (2024b, March 19). AI matches the abilities of the best math Olympians. Scientific American. https://www.scientificamerican.com/article/ai-matches-the-abilities-of-the-best-math-olympians/#:~:text=The%20researchers%20reported%20that%20the,escaped%20the%20attention%20of%20experts.
