THE ARC · MEASURED, EXPLAINED, LIVE

🧬 The Arc
Where AI is on the curve — and how close to the singularity stage.

Every turning point since the first commercial language model, on one chart you can read in a minute. Below it, a gauge built from four things that can actually be measured — not vibes — with the formula shown. The right edge is live: what the exchange sees this month.
click a point

Pick a turning point

Each dot is a moment the curve changed direction — a new capability, a new way of working, a collapse in price, a new kind of risk. Hover or click to read why it mattered.

all turning points
    How to read the curve. The vertical axis is the Meerada Capability Index — a 0–100 composite of what the best available model could reliably do for a working person at that date (0 = none, 100 = a competent human expert across all measured task clusters). History is set by hand from the record; the shaded band is our uncertainty; the dashed tail is a projection at the measured pace. Stages across the top are the five phases on the way to a self-improving system.

    ⚡ Singularity proximity — four gauges, one formula

    "Singularity" here means something specific and testable: AI systems that improve AI systems faster than people can follow. Four measurable things have to be true for that. We score each 0–100 and average them. Nothing here is a prediction of consciousness or doom — it is a dashboard.
    1 · Task horizon (autonomy)
    —

    2 · Cost of frontier intelligence
    —

    3 · Capability vs. human experts
    —

    4 · AI building AI
    —

    ARC SCORE · 0–100
    —
    —

    Formula: Arc = mean(horizon, cost, capability, self-build). Horizon: log-scale from 1 minute (0) to one working month, 160 h (100). Cost: log-scale from $60 (0) to $0.01 (100) per million output tokens at frontier quality. Capability: share of expert-level task clusters where the best model matches a competent human. Self-build: share of frontier-lab engineering done by AI. Each input and its source is written on the gauge.

    Honesty note. Gauges 1 and 2 are computed from published trend measurements and the exchange's own live prices. Gauges 3 and 4 rest on lab-reported figures and our own reading of benchmark saturation; they are estimates, labelled as such, and we would rather be corrected than be wrong quietly. The projection line uses the measured doubling rate of agent task horizons (about seven months) — if that rate holds, a one-month autonomous task horizon arrives around —.

    What the stages mean

    1 · Tools (2017–2020)

    Models complete text. Useful to researchers; a working person cannot hand them a job. The Transformer, GPT-2, the first API.

    2 · Assistants (2021–2023)

    A model answers, drafts, explains and writes code on request. ChatGPT, GPT-4, Claude, Llama. You do the task; it helps.

    3 · Agents (2024–2026) — we are here

    Models plan, use tools, browse, run code, and work for hours on a task with checkpoints. Reasoning models, computer use, coding agents. You give the job; you check the result. This is exactly why Meerada exists: measuring done tasks, guarding the run, cross-checking the answer.

    4 · Autonomous researchers

    Agents run multi-day projects end to end — experiments, code, papers — with people setting goals and reviewing outcomes. Entered when the task horizon passes a working week at high reliability.

    5 · Recursive improvement (the singularity threshold)

    AI systems design, train and evaluate the next generation faster than people can audit. The gauge calls it when all four inputs sit above 80. The whole point of a public, neutral measurement layer is that this is noticed when it happens, not after.

    Why a model exchange keeps this chart

    Because we measure the same thing every hour: how many verified tasks a model finishes and what each one costs. That is the curve's right edge, updated by machines, not by press releases. Open data: grade_state.json, models_live.json, index.json.