Saturday, 30 May 2026

The protein-folding wars just shifted in a way that should matter far more than the usual lab-to-lab competition. Meta's Biohub released ESMFold2, and it's not just incrementally better than Google's AlphaFold—it's handling 1.1 billion proteins where AlphaFold was the standard. This matters because protein structure prediction is the foundation for drug discovery, synthetic biology, and materials science. When Google owned this problem, there was a clear gatekeeper. Now the best tool is open-source and backed by Meta's infrastructure. I think we're about to see an acceleration in biotech applications that were previously bottlenecked by access and compute costs.

By the way, this illustrates a broader pattern I've been noticing: the real competitive advantage in AI is no longer owning the model itself, but owning the surrounding ecosystem. A new review paper on AI agents makes this explicit—it argues that code, not the language model, is the actual bottleneck for autonomous agents. The LLM is table stakes now. What separates a capable agent from a toy is the software orchestration layer: how tasks are decomposed, how state is managed, how errors cascade or get contained. This is why I find Microsoft's new super-app project interesting. Jacob Andreou, brought in as the new Copilot chief, is consolidating Microsoft's scattered AI tools into a unified platform. That's not exciting because it's sleek design—it's exciting because it's admitting that the fragmented Copilot lineup was the real problem, not the underlying models.

The infrastructure layer is heating up too. Samsung is now challenging Micron with memory chips optimized for AI workloads. This is the unglamorous but essential race: who builds the silicon that actually runs inference at scale. Meanwhile, Anthropic's valuation has climbed to $965 billion, nearly catching OpenAI, on the strength of Claude adoption and execution. I don't think this is just hype—the company has genuine momentum in enterprise deployments—but valuations at this level are betting on something beyond today's capabilities.

What I'm watching closely is the quality question. Kaggle datasets used in clinical models were found to contain celebrity images and thousands of duplicates, leading to retractions. This is a sharp reminder that we can have the best algorithms in the world, but garbage data still produces garbage outcomes. As these tools move from labs into medicine, agriculture, and infrastructure, the unglamorous work of data quality becomes critical infrastructure. That's where real risk lives right now.