Max Welling's upcoming AMA is the kind of event that matters precisely because it refuses to treat AI as a parlor trick. This is someone who has spent decades moving between the theoretical foundations of machine learning and the messy, physical reality of scientific discovery. For our readers, that means an opportunity to engage with a researcher who is not just asking "what can models do?" but "what should models do next?" That distinction is where the real value sits, and it's why we think this conversation deserves your attention.
The practical takeaway here is about the shift in where machine learning is heading. Welling's trajectory from graph neural networks and Bayesian deep learning into AI for materials and earth systems is not a pivot away from core ML; it's an expansion of its scope. When he talks about treating the physical world as a live data generator for frontier model training, he's pointing to a future where the lab loop and the computation loop are the same loop. For anyone who has struggled with the gap between a well-curated dataset and the noise of real-world deployment, this framing is not academic. It's a direct acknowledgment that the hardest problems are not always the ones with the cleanest loss functions.
What stands out in the proposed topics is the emphasis on reliability and human involvement. Welling isn't asking for blind trust in foundation models; he's asking how we build systems that remain useful when data is sparse, noisy, or only partially observable. That's a question that should resonate beyond the research community. Whether you're building tools for carbon capture or just trying to make sense of your own spreadsheet workflows, the underlying principle is the same: the model is only as good as its ability to function where certainty is a luxury. His focus on synthesizability and data quality is a reminder that the bottleneck is often not the architecture but the interface between the model and the physical constraints it must obey.
So here is the concrete point we want you to take with you: when you join the AMA, don't just ask about the next big architecture. Ask about the failure modes. Ask about the gaps between what models predict and what labs can actually produce. Ask about the decisions that were made when the data did not cooperate. Welling has spent a career navigating those edges, and his answers will give you a clearer sense of what is genuinely solvable today and what remains an open problem. That is the kind of insight you cannot get from a benchmark leaderboard. It's the kind of insight that comes from talking to someone who has been in the arena long enough to know what the arena actually demands.