Option 5 · The Web — the shape of the question
Six dimensions, one figure. No single date survives contact with a multidimensional problem — so this option refuses to give one. It opens with the evidence; the argument arrives by clicks.
Find the six timelines. The readings pre-filled below are one worked example, set from published evidence for this demo — scaffolding, not answers. The project’s real readings are found: each dimension goes to a student (or a pair), who brings back a measure, a source, and the readings the evidence supports. Type your found readings in and the web above redraws itself — the shape becomes whatever your evidence says it is. Then defend it: every reading ships with its source and its reasons, the same discipline as the green line.
| Dimension | 2023 | Today |
|---|---|---|
| Capability find it in: METR task-horizon measurements | ||
| Compute find it in: Epoch-style training-compute records | ||
| Deployment find it in: adoption & workplace-use surveys | ||
| Oversight & control find it in: agent identity & accountability practice | ||
| Safety find it in: evaluation & alignment research capacity | ||
| Autonomy find it in: agent run-length / human-in-the-loop reports |
The strips further down remain the worked example. Yours will differ — that difference is the discussion.
The chart opens with the data alone — the web a few years ago against today’s. The two argument layers are one click away, in the order you’d present them: first the shape worth defending, then the box worth fearing.
Six timelines. The web above is a photograph. Each axis is really a film, running at its own speed — solid to today, dashed where the pace merely holds. The dotted lines are that axis’s yellow (defended) level and its red (box) level.
Read the bottom two strips against the top four. The machine dimensions crossed their yellow levels years ago and press on toward the box. The human dimensions crawl — on their present pace they do not reach their yellow levels at all. There is no one clock. There are six, and they disagree.
What each axis reads:
Capability — tasks frontier agents complete at 50% reliability (METR): ~2–5 hours in early 2026, doubling roughly every 7 months. Compute — the scale of training, still climbing era over era. Deployment — how far AI has spread into real work and daily life. Autonomy — how long agents run without a human in the loop. Safety — the maturity of evaluation and alignment science. Oversight & control — whether agents are identity-bound and answerable: the yellow line, drawn as an axis.
Honesty note: the readings are illustrative, set for this demo from the published evidence. In the project they are set by the group, argued in the open, and published with their reasons — the same discipline as the green line.
The box and the yellow.
The yellow (dashed) is the defended shape — the same growth as today’s web, with oversight and safety grown out to meet it: AI under human control, drawn as a contour. It is the yellow line from Option 3, made visible.
The AGI box (red) is drawn at the rim on the machine axes — capability, compute, deployment, autonomy all but maxed — and pinched to little or no control or safety: its edge passes almost through the center on those two axes. That is the shape of arrival with nobody holding it.
Read today’s web against both: on compute it already presses toward the red box, while on oversight it sits far inside the yellow. The web is drifting from the yellow toward the box. The drift is the story.
A clock gives you a date. A web shows you a shape. The shape is lopsided — capability stretches toward the rim while oversight stays near the center. That lopsidedness is the finding.