Artificial General Intelligence Is Already Here | NOEMA

agillmfrontier-modelsgeneral-intelligenceai-capabilities

Abstraction: Argument that current frontier LLMs have already achieved meaningful artificial general intelligence

Key points:

Authors Blaise Agüera y Arcas (Google VP/fellow) and Peter Norvig (Stanford HAI) argue frontier models (ChatGPT, Bard, Claude, LLaMA) are the first true AGI, analogous to ENIAC as the first general-purpose computer
Five dimensions of generality achieved: topics (vast training data), tasks (Q&A, summarization, translation, coding, etc.), modalities, languages (dozens), and instructability (zero-shot and few-shot in-context learning)
Four reasons for reluctance to acknowledge AGI: skepticism of metrics, ideological commitment to symbolic AI (Chomsky, Gary Marcus), human exceptionalism, and economic threat anxiety
Goodhart's law problem: models tuned to pass exams (bar, medical licensing) may not generalize — better tests like Stanford HELM are needed
Emergent abilities may be an artifact of nonlinear metrics: gradual capability growth looks like a sharp threshold when measured with binary pass/fail tests
The "ought" debate (who benefits, who is harmed, how to distribute gains) should be conducted openly rather than by denying AGI has arrived

Connections: Google · Openai · Chatgpt · Artificial General Intelligence · Large Language Models · AI Capabilities