Google AI solves decades-old maths problems, but DeepMind CEO says AGI is still far away

Google DeepMind has said its artificial intelligence system AlphaProof Nexus has solved several long-standing mathematical problems, including some that remained unsolved for more than 50 years. The development has drawn attention across the AI community, especially as debates grow around whether such breakthroughs signal progress towards artificial general intelligence (AGI).

According to researchers, the system solved nine open problems related to mathematician Paul Erdős, along with 44 open conjectures and a 15-year-old problem in algebraic geometry, showing how AI is increasingly being used in advanced scientific research.

How Google’s AI solved complex maths problems

The system, called AlphaProof Nexus, works by combining AI reasoning with a formal verification tool known as Lean. Instead of just generating answers, the AI produces step-by-step proofs, which are then checked automatically by the system.

This approach helps ensure that:

Every logical step is verified
Incorrect assumptions are rejected
The final proof meets strict mathematical standards

Researchers said the cost of solving each problem was only a few hundred dollars in computing power, highlighting how efficient such systems can be.

Why ‘hallucinations’ remain a concern

One of the key challenges in AI-generated mathematics is the issue of hallucinations, where systems produce results that appear correct but contain hidden errors.

Experts explain that AI models may:

Invent new statements that are not proven
Skip difficult parts of a problem
Present incomplete proofs as finished

To address this, Google’s system uses Lean to verify every step, reducing the risk of such errors and making the results more reliable.

How this compares with other AI developments

The announcement comes shortly after OpenAI said one of its models solved a well-known mathematical problem. However, OpenAI relied on external experts to verify the result, while DeepMind used automated verification.

This difference highlights two approaches in the AI industry:

Human-reviewed proofs
Machine-verified proofs
Researchers believe combining both methods may improve future results.

Does this mean AI is reaching AGI?

Despite the progress, Demis Hassabis has said that current AI systems are still far from achieving AGI. He explained that solving specific problems does not mean AI can match human intelligence across all areas. True AGI would require:
Creativity across different fields
Independent reasoning
Ability to generate original ideas
According to Hassabis, today’s systems are powerful tools but do not yet show the level of understanding seen in human experts.

What this means for research and industry

The development shows how AI is becoming a useful assistant in scientific research. Instead of replacing mathematicians, these systems can help:

Explore complex problems faster
Check proofs more efficiently
Support researchers in difficult areas

Experts say this could change how mathematics and science are conducted in the future. As AI models continue to improve, their role in research is expected to expand. However, the debate around safety, accuracy and long-term capabilities will remain important. For now, while AI can solve complex mathematical problems, it is still seen as a specialised tool rather than a system with general intelligence