The Next 25 Years

Predictions about AI's future range from 'transformative beyond imagination' to 'perpetually five years away.' Both poles appear confidently in the writings of credentialed experts. This lesson does not tell you what will happen — it teaches you how to think about AI's future rigorously, so you can evaluate claims yourself rather than outsourcing your conclusions to whoever sounds most confident.

Trends That Are Supported by Evidence

Some trends in AI have shown consistent empirical patterns and have clear mechanistic explanations — not guarantees of continuation, but grounded in evidence. Scaling: the empirical observation that increasing model parameters, training data, and compute tends to improve performance — sometimes smoothly, sometimes via sudden capability jumps called emergent abilities — has been robust across a decade and multiple architecture generations. The Chinchilla paper (Hoffmann et al., 2022) refined scaling laws to show that both model size and training data matter in proportion, and that many large models had been undertrained relative to their size. Scaling is not infinite — energy costs, data availability, and diminishing returns are real — but the upper bound is not yet visible in practice. Multimodal convergence: models that process multiple modalities within a single architecture have improved faster than specialized single-modality models on many benchmarks, suggesting that shared representations across modalities are genuinely useful. This trend supports the expectation that future general-purpose AI systems will be natively multimodal. Deployment breadth: AI is being embedded in an expanding set of products, infrastructure, and workflows. Each year, AI touches more domains of economic activity. The direction of this trend is clear even if the pace and ultimate extent are not. Computational infrastructure: AI training and inference require specialized hardware (primarily GPUs and TPUs). The capital investment in AI infrastructure — data centers, power, chips — has grown to the hundreds of billions of dollars annually, creating structural momentum that makes significant rollback unlikely without extraordinary disruption.

What is genuinely uncertain? Several of the most important questions have no clear empirical answer. Will scaling continue to produce qualitative capability improvements? The emergent abilities observed in large language models — in-context learning, chain-of-thought reasoning, code generation — were not predicted before they appeared. It is not known whether continued scaling will produce further qualitative shifts, or whether the current architecture family has a ceiling that scaling cannot overcome. Artificial General Intelligence (AGI) timelines: AGI is loosely defined as an AI system that can perform any cognitive task a human can, at human level or above. Predictions of when AGI might be achieved range from 'already happened' to 'never' to 'decades away,' from credible researchers. The disagreement is not primarily about facts — it is about what 'general' means, what cognitive tasks actually constitute intelligence, and how to extrapolate from current capabilities. This is a philosophical question embedded in a technical prediction, and it should be treated with corresponding epistemic humility. Alignment: as AI systems become more capable and are deployed in higher-stakes roles, the question of whether they will reliably pursue intended goals becomes more consequential. Current alignment research includes techniques like RLHF (Reinforcement Learning from Human Feedback), Constitutional AI, and mechanistic interpretability — but the field does not yet have a proof or a reliable method for guaranteeing alignment at scale. This is an active research frontier, not a solved problem. Economic displacement: the effects of AI on labor markets are genuinely uncertain. History shows that automation tends to displace specific tasks rather than entire jobs, and tends to create new job categories while eliminating others — but the pace and distribution of these effects are hard to predict, and the analogy to previous waves of automation may not hold if AI capabilities are more general than previous technology.