Walk for two hours and you cover twice the distance of one hour. That's linear progress - predictable, proportional, and completely wrong as a model for what's happening in AI.
Mustafa Suleiman, CEO of Microsoft AI and co-founder of DeepMind, makes this argument in a piece for MIT Technology Review. The "AI is running out of steam" narrative keeps regenerating itself, he writes, because human intuition is calibrated for linear change. AI doesn't work that way.
The Prediction Track Record
The shape of failed AI predictions has been consistent. Two years ago, GPT-4 was widely described as approaching a hard ceiling. Then o1 shipped and redefined what was possible on complex reasoning tasks. Then o3. Then the current generation of frontier models. At each step, analysts who predicted stagnation turned out to be wrong, usually by wide margins.
This reflects a structural pattern. Training large language models - feeding vast amounts of text through enormous computing networks to teach them to reason and generate language - kept producing returns that experts repeatedly underestimated. When pre-training data became a constraint, the industry shifted to synthetic data: examples generated by existing AI models to train their successors. When standard benchmarks (the tests used to measure model capability on specific tasks) hit saturation, harder ones appeared within months.
Each apparent ceiling turned out to be a transition point, not a terminus.
Compounding vs. Ceiling
The practical difference between exponential and linear growth is how quickly they diverge. A technology compounding at 30% annually doubles in less than three years. Predictions based on linear extrapolation fall further behind with every doubling cycle.
Suleiman's argument isn't that AI progress faces no constraints. Energy costs, chip availability, and the genuine difficulty of certain reasoning problems are real. His point is more specific: the particular ceilings that critics point to have historically been the wrong ones. The research community keeps finding routes around them.
For people building work habits around AI tools - content creators, marketers, analysts, developers - this shapes what kinds of investments are worth making. Tools that seem capable but limited today are likely to improve more than most forecasts suggest. Building genuine fluency with AI assistance now, rather than waiting for the technology to stabilize, pays compounding returns as the underlying capability keeps improving.
Suleiman co-founded DeepMind in 2010, years before large language models existed in their current form. He has watched multiple generations of AI limitation predictions dissolve in practice. His perspective on where the actual ceiling lies is informed by having been inside the industry during every previous cycle where stagnation was predicted and didn't materialize.