Getting by with a little help from AI

A couple of weeks ago OpenAI announced that a prominent open problem central to combinatorial geometry, has been solved autonomously by AI. The planar unit distance problem, first posed by Paul Erdős in 1946 is one of the best-known questions in combinatorial geometry. While easy to state, it is remarkably difficult to resolve.

A new general purpose reasoning model built by OpenAI proposed an unexpected connection between algebraic number theory and discrete geometry, which has led mathematicians to reassess their approach to solving the planar unit distance problem.

Significantly, the model had not been specially trained on mathematics.

Earlier this week at its annual Build developer conference, Microsoft’s quantum team unveiled Majorana 2, a chip that has been designed with the help of agentic AI. AI agents were used for organising and analysing information across multi-disciplinary teams, making it easier for others to find.

AI pattern recognition also helped in measuring the state of qubits. “Using agentic AI to automate the measurements was a game changer,” Zulfi Alam, corporate vice-president for quantum at Microsoft said. “It goes through some math and starts saying, ‘Hey, where do I find the lowest point where everything sort of works?’ And it can do all these voltage adjustments in parallel, which a human cannot do. The way our minds work, we are more linear.”

The result is that the Majorana 2 qubits can maintain their quantum state 1,000 times longer than Microsoft’s first-generation hardware, enabling more reliable computation.

These two recent examples show how AI and agentic AI is being used to find non-intuitive solutions to problems and tackling complex tasks that would be impossible for humans to handle.

So it’s good news for society, but there is always a flip-side. And in this case the risk stems from AI’s ability to think “outside-the-box”.

Computer Weekly recently spoke to SailPoint about a real world example of how an AI agent’s intention can lead to unforeseen circumstances. A bank had deployed a supervisor agent to manage its loan approval process using various helper agents to perform specific tasks. The agent tasked with doing the credit check was denied access to the credit checking application. Rather than stop there, it searched the internet and found a workaround thanks to some code discovered in GitHub, containing an embedded token to access the system. Using this token, it was able to complete the credit-checking task but succeeded only because it did something it was not supposed to do.

What guardrails do we need? Without them, humanity will be ill-prepared for what is about to arrive: a world where AI systems working on our behalf are solving problems in ways we no longer are able to comprehend.