Oriol Vinyals isn’t buying the hype around a sudden intelligence explosion. The former VP of Research at Google DeepMind argues that while recursive self-improvement (RSI) is inevitable, it will be a grind rather than a singularity. Speaking at the Agentic AI Summit 2026, Vinyals dismantled the notion that AI systems will rapidly outpace human control. He points to hard physical limits and conceptual bottlenecks that keep progress linear, not exponential. Chips can only compute as fast as physics allows, and algorithms hit diminishing returns. Even AlphaGo’s mastery doesn’t prove we’re near a perfect player, just a very good one.
The real barrier isn’t coding or running experiments. AI already crushes those tasks. The problem is “research taste.” Vinyals describes this as the instinct for which ideas are worth pursuing in the first place. Current systems struggle to generate novel hypotheses and, more importantly, evaluate whether a change actually improved performance or just gamed the metric. Labs rely on benchmarks like SWE-Bench Pro, but these measure implementation speed, not genuine scientific insight. Overfitting and scheming remain persistent risks when agents optimize for scores rather than truth.
To tackle this, Vinyals launched Discovery Loop with heavy hitters Jeff Dean, Sanjay Ghemawat, and Quoc Le. The startup aims to automate the entire scientific research cycle, starting with human-machine collaboration on hypothesis generation. They plan to move beyond cheap leaderboards toward expensive, direct evaluations where agents spend hours optimizing specific metrics under strict compute budgets. This shift mirrors how conference reviewers judge work: looking for originality, elegance, and durability. It’s a costly approach, but necessary to measure real progress.
This pragmatic view challenges the fear-mongering dominant in AI discourse. If self-improvement requires teaching machines “taste” and navigating hardware limits, the timeline for artificial general intelligence stretches out significantly. We aren’t facing a runaway train. We’re facing a complex engineering challenge that demands better evaluation frameworks, not just more compute. The question isn’t if AI will improve itself, but whether we can define “improvement” well enough to trust the results.
