From games to biology and beyond: 10 years of AlphaGo’s impact

DeepTrendLab's Take on From games to biology and beyond: 10 years of AlphaGo’s impact

A decade after AlphaGo toppled Lee Sedol, Google DeepMind is consolidating a far more consequential victory: the transformation of AI search methods from game-playing parlor tricks into the engine of scientific discovery. The anniversary retrospective isn't nostalgia—it's a progress report on how a single architectural insight has metastasized across biology, mathematics, and algorithm design. AlphaFold 2 solved protein folding, a problem that had resisted biological assault for fifty years. AlphaProof and AlphaGeometry 2 achieved medalist-tier reasoning at the International Mathematical Olympiad. An updated Gemini variant then surpassed even that, reaching gold-medal performance in 2025. Separately, an AI system called AlphaEvolve discovered novel matrix multiplication algorithms that are now foundational to modern neural networks. Each breakthrough represents not an isolated success but a proof that the same underlying logic—combining search with reinforcement learning—scales across wildly different domains.

The conceptual innovation of AlphaGo was never truly about Go. The game was a testing ground for a hypothesis: that you could train an AI to evaluate positions not through explicit human-curated rules but through a learned intuition, then use tree search to explore billions of possibilities and converge on optimal moves. In 2020, that same approach cracked protein folding by reconceiving the problem as a search through conformational space rather than a brute-force enumeration. The protein database that emerged went immediately open-source, and three million researchers now use it for everything from vaccine design to enzymatic engineering. The appearance of a Nobel Prize in Chemistry in 2024 for this work wasn't decoration—it was institutional validation that the line between AI and natural science had blurred past recovery. What followed was systematic application: the same search-and-reasoning pipeline adapted to formal mathematical proof, then to algorithmic discovery, then to multi-agent scientific collaboration.

The real implication is that DeepMind has cracked a meta-problem: how to transplant problem-solving patterns across domains. This is the inverse of the usual narrow-AI complaint. Rather than locking capability into a single task, DeepMind has shown that the *process* of combining neural intuition with exhaustive search is portable. Protein folding and IMO math problems have almost nothing in common structurally, yet the same toolbox applied to both produced breakthrough results. The emergence of Deep Think as a reasoning mode in Gemini suggests this isn't luck—it's a genuine advance in how to architect systems that can tackle open-ended, multi-step problems. For the AI industry, this shifts the competition from "whose model has the highest benchmark scores" to "whose methods can be adapted to unsolved real-world problems." DeepMind is signaling that they own the architectural framework.

The immediate beneficiaries are obvious: molecular biologists, pharmaceutical researchers, and mathematicians now have a research tool that didn't exist five years ago. But the reach extends deeper. Universities and small research labs can now access protein structure predictions that once required decades of lab work or crystal-lography expertise. Enterprises building on neural networks benefit from AlphaEvolve's algorithmic discoveries, potentially reducing compute requirements across the industry. The subtext is that AI research infrastructure is being consolidated into the hands of organizations with the computational resources to run these systems. A researcher in Lagos or Santiago can access the database, but the ability to *invent the next AlphaFold* remains a capability moat for institutions like DeepMind. This creates a technological dependency where global science becomes increasingly reliant on a few companies' research agendas.

Competitively, this is DeepMind reasserting dominance after years of chatbot-driven market noise. OpenAI captured public mindshare with scale and commodity capabilities; DeepMind is reclaiming the narrative by showing that their methods drive *discovery*, not just fluency. The mention of Gemini's Deep Think achieving gold-standard IMO performance is a signal to enterprises and academia that reasoning capability has advanced beyond what public benchmarks suggested. It also creates a soft lock-in: if your research pipeline now depends on AlphaFold and AlphaProof, you're invested in an ecosystem where DeepMind controls the frontier. Elsewhere, the AI co-scientist system being deployed with imperial researchers suggests a vision where AI doesn't replace scientific thinking but amplifies it—a more durable positioning than claiming AI can *do* science.

The trajectory to watch is whether this reasoning-plus-search pattern continues to generalize. If Deep Think can scale from IMO problems to open-ended scientific challenges, the next question is whether it can tackle industrial problems: drug discovery pipelines, materials science, complex systems optimization. Another critical signal is whether competitors can replicate the architectural approach or whether DeepMind's Moat is structural (better algorithms) or resource-based (more compute). Finally, there's the question of governance: as these systems become essential infrastructure for scientific research, who decides what gets solved first? DeepMind is claiming the role of scientific collaborator, but that role comes with enormous power over which human problems become computationally tractable.