Artificial General Intelligence
Artificial general intelligence (AGI) refers to a hypothetical AI system that can learn and perform any intellectual task a human can, matching or exceeding human-level reasoning and transfer across domains it was never specifically trained for — unlike today's AI, which stays specialized even when it looks broad (a model fluent in code, images, and language is still only interpolating within its training distribution, not genuinely generalizing). At a high level, the capability progression researchers track looks like this:
No lab has built one, and there is no agreed test for when one exists. DeepMind's 2023 "Levels of AGI" paper proposes six tiers (No AI through Superhuman) scored against breadth and performance rather than a single pass/fail line; François Chollet's ARC-AGI benchmark instead measures novel-pattern abstraction specifically because it resists memorization, and current frontier models still score far below average humans on it even as they ace standardized exams. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" — a deliberately economic rather than philosophical bar.
The bottleneck industry labs actually track is compute, not a missing algorithm. Epoch AI's public tracking shows training compute for frontier models growing roughly 4-5x per year: ResNet-era vision models trained on the order of 10^19 FLOPs, GPT-3 landed near 10^23, GPT-4-class models reached an estimated 2×10^25, and 2025-2026 frontier runs are estimated in the 10^26 range. That curve, not a single conceptual breakthrough, is the variable every major lab's scaling-law papers treat as the controllable lever.
Every scaling step reroutes straight back into chip demand. A 10x compute jump means a proportionally larger cluster of NVIDIA GPUs (H100, B200, and now GB200 NVL72 racks), fed by HBM stacks from SK hynix, Samsung, and Micron, assembled onto silicon interposers by TSMC's CoWoS packaging line — the same packaging bottleneck covered in the chip glossary entry. xAI's Colossus cluster scaled past 100,000 H100-class GPUs in a single build; reported plans for next-generation training clusters run into the hundreds of thousands of GPUs and multiple gigawatts of power. AGI progress, measured this way, is inseparable from foundry and packaging capacity, not just model architecture.
Skeptics argue the scaling curve alone doesn't cross the line. Yann LeCun has argued that autoregressive next-token prediction lacks a persistent world model and can't support genuine planning; Gary Marcus points to models still failing basic compositional reasoning and physical common sense despite massive compute budgets. Continual learning — updating knowledge without forgetting prior skills, still largely unsolved outside of retraining from scratch — is the capability most researchers flag as the real remaining gap, independent of how many more FLOPs get thrown at the next run.
Read AGI not as a single finish line but as a compute-denominated moving threshold: every claim about how close the field is amounts to a claim about how many more GPU-hours, HBM-fed racks, and CoWoS-packaged dies separate today's frontier model from a system that needs no further human help crossing into an unfamiliar domain — the same yield-and-layers economics that build the chip underneath all of it.