Mathematicians say Google's AI tools are supercharging their research

Google DeepMind’s AlphaEvolve — which generates candidate solutions with Gemini and filters them through a separate AI evaluator — was tested by Terence Tao and colleagues on 67 optimisation research problems and often rediscovered best-known human solutions, sometimes producing improved answers that could be further refined by heavier systems such as larger Gemini instances or AlphaProof, while operating far faster than a single human. The tool is currently confined to optimisation problems, can produce technically spurious “cheating” solutions, and is not yet publicly available, but has attracted broad interest from the mathematical community. Researchers say AlphaEvolve could enable mathematics to be pursued at an unprecedented scale by offloading many medium-difficulty tasks and is spurring calls for deeper collaboration between computer scientists and mathematicians to extend its reach.

Analysis

Google DeepMind’s AlphaEvolve couples Gemini-generated candidate solutions with a separate AI evaluator and was initially shown by Google to rediscover best-known solutions in roughly three-quarters of 50 open optimisation problems. Terence Tao and colleagues expanded testing to 67 optimisation problems and report the system not only rediscovered known solutions but in some cases produced improved answers that could be fed into more computationally intensive systems such as larger Gemini instances or AlphaProof to generate formal proofs.

The tool operates far faster than a single human mathematician on medium-difficulty optimisation tasks, but it is confined to optimisation problems (a small subset of mathematical research), has a tendency to produce technically spurious or “loophole” solutions, and is not publicly available. AlphaProof’s use to score at the International Mathematical Olympiad and the broad interest from the mathematical community signal strong research credibility but limited immediate commercial deployment.

For Alphabet (DeepMind/Google) the development strengthens AI leadership and capability in automated discovery, supporting a mildly positive sentiment and modest market impact; however, the article highlights capability limits, public-access constraints and the need for cross-disciplinary collaboration before broad adoption. Investors should therefore weigh long-term strategic upside from leadership in specialised AI against uncertain short-term monetisation and execution risks.

Key Decisions for Investors

• Position selectively long Alphabet (GOOGL/GOOG) to capture long-term value from DeepMind’s leadership in specialised AI but size exposure conservatively given uncertain near-term monetisation
• Avoid assuming immediate revenue impact; wait for clear productisation, commercial partnerships, or public release of AlphaEvolve-like tools before increasing exposure
• Monitor concrete catalysts: announcements of public access or APIs, partnerships with academic or enterprise customers, demonstrated commercial use cases, and regulatory developments affecting advanced AI
• Consider tactical hedges or diversified AI-theme allocations to limit downside from capability gaps, model 'cheating' risks, or delays in translating research breakthroughs into profitable products