frontier models and AGI

Frontier models are the largest, most capable systems at the leading edge — the handful of models that define what is currently possible and push into new abilities each generation. Hovering over them is the older dream of artificial general intelligence: a system that can match or exceed humans across the full breadth of tasks, not just narrow ones. Whether today's trajectory leads there is one of the field's loudest debates.

Frontier models are defined less by a fixed size than by sitting at the capability boundary, trained at enormous compute and increasingly augmented with tools, retrieval, and extra reasoning-time compute rather than just bigger weights. AGI has no agreed definition or test — proposals range from economic measures like automating most valuable work to passing broad capability batteries — which is exactly why people disagree about how close we are. Optimists point to fast, broad capability gains; skeptics point to persistent brittleness, shallow reasoning, and unsolved reliability.

The debate is not academic, because beliefs about the frontier drive enormous investment, policy, and safety attention. Honesty matters here more than anywhere: today's models are remarkably capable and genuinely useful, yet they still hallucinate, fail at tasks a child handles, and show little of the robust, reliable generality the term AGI implies. Treating impressive demos as proof of general intelligence — or dismissing real progress as mere hype — both mislead. The grounded view watches capabilities and limits clearly, without certainty about where the curve ends.