The Intelligence Index

The human brain contains roughly 86 billion neurons and an estimated 100 trillion synapses, organized into a six-layered neocortex, subcortical structures, cerebellum, and brainstem. It runs on about 20 watts of metabolic power — less than a dim lightbulb — yet supports language, abstract reasoning, embodied motor control, and continuous lifelong learning.

Cognition emerges from specialized regions working together: prefrontal cortex for planning and working memory, hippocampus for episodic consolidation, basal ganglia for action selection, and a default mode network active during self-referential thought. Synaptic plasticity lets the brain learn new tasks from a handful of examples — a sample efficiency no current artificial system matches.

Humans remain the only known system with confirmed subjective experience, theory of mind across arbitrary contexts, and the ability to construct and revise their own goals across decades of life.

Chimpanzees (Pan troglodytes) share roughly 98.8% of their DNA with humans and possess brains of about 350 grams containing an estimated 28 billion neurons — roughly a third of the human count, in similar cortical and limbic structures. They use and modify tools in the wild (termite-fishing sticks, hammer-and-anvil nut cracking), pass cultural variants between troops, and recognize themselves in mirrors.

Studies including Ai, Kanzi, and Washoe demonstrate symbol use, short-term memory tasks where chimps outperform adult humans, and limited comprehension of human grammar, though they do not produce recursive syntactic language. Their social cognition supports coalition politics, deception, and reconciliation.

Chimpanzees set a useful biological baseline for what intelligence looks like without human-scale language and without artificial compute.

Frontier large language models such as OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet / Opus, and Google Gemini 1.5/2.0 Ultra are transformer networks with parameter counts in the hundreds of billions to low trillions, trained on tens of trillions of tokens of text, code, and multimodal data. Training runs cost tens to hundreds of millions of dollars and consume megawatt-scale electricity.

After pretraining they are aligned with supervised fine-tuning and reinforcement learning from human or AI feedback (RLHF / RLAIF). Capabilities include fluent multilingual generation, summarization, translation, code synthesis, and vision-language understanding. They score in the 80–90% range on MMLU, pass bar and medical licensing exams, and solve a majority of HumanEval coding problems.

Their cognition is bounded by a fixed context window (typically 128k–2M tokens), frozen weights at inference, and a tendency to hallucinate plausible but false statements when extrapolating beyond training distribution.

Reasoning models — OpenAI o1 and o3, DeepSeek-R1, Google Gemini 2.0 Flash Thinking, Anthropic's extended-thinking modes — extend the frontier LLM architecture with reinforcement learning over long chain-of-thought traces. At inference they spend additional compute generating, verifying, and revising internal reasoning steps before producing a final answer.

This test-time compute scaling produces large gains on tasks with verifiable solutions: o3 reaches 87.5% on the ARC-AGI semi-private set, 96.7% on AIME 2024 mathematics, and competitive programming Elo above 2700. Costs per query rise by one to two orders of magnitude compared to a single-pass LLM.

They remain frozen-weight at inference and share the multimodal and memory limits of base LLMs, trading latency for accuracy — useful for math, code, scientific reasoning, and formal verification.

AI agents wrap a frontier or reasoning model in a control loop that perceives an environment, plans, calls external tools (web browsers, code interpreters, APIs, file systems), and writes to long-term memory. Reference implementations include Anthropic Claude with computer use, OpenAI Operator, Google Project Mariner, Devin, and open-source frameworks such as LangGraph and AutoGPT.

Persistent memory is typically implemented with vector databases, scratchpads, and structured stores rather than weight updates, so learning is online at the memory layer but the underlying model remains frozen. On the SWE-bench Verified software-engineering benchmark, leading agents resolve 60%+ of real GitHub issues; on WebArena and OSWorld they complete a meaningful but minority share of multi-step computer tasks.

Agents extend LLM cognition into action, but inherit hallucination, brittle long-horizon planning, and difficulty recovering from compounding errors in open environments.

Artificial General Intelligence (AGI) refers to a hypothetical system that matches or exceeds human performance across the full breadth of cognitive tasks — language, reasoning, learning from few examples, transfer to new domains, embodied control, and social cognition — rather than excelling in a narrow band. OpenAI's charter frames AGI as 'highly autonomous systems that outperform humans at most economically valuable work,' while DeepMind's 2023 levels framework defines it as performance at or above the 50th percentile of skilled adults across non-physical tasks.

No system today meets these criteria. Public timelines from researchers at OpenAI, Anthropic, DeepMind, and METR cluster between 2027 and the late 2030s, with substantial disagreement and tail risk in both directions. Required advances likely include continual learning without catastrophic forgetting, robust long-horizon planning, and grounded world models.

Subjective experience in such a system is unresolved and is scored 0 in this index to reflect the absence of accepted evidence, not a claim of impossibility.