Back

Why this exists

Most AI-risk coverage puts three very different kinds of claim in the same confident tone. Something that happened this morning. A number from a dataset that was accurate six months ago. Somebody's guess about 2029. Read quickly, they all sound equally solid. This dashboard is built to keep them apart, and I didn't want to add one more risk score to the pile.

Every figure carries one of three tags, and each tag means something specific. Live means the figure is polled from the source RSS feeds every few minutes. Reference means it's real and cited, but dated: the data-center map comes from Epoch AI's published dataset (licensed CC BY 4.0), a snapshot rather than a live crawl. Simulated means it comes from a toy model I built by hand.

The map's Network view follows the same rule. Every line between a lab, a data center and its owner is a relationship Epoch lists. If Epoch rates a link as only likely or speculative, I draw it dashed. A site with no listed owner or user stays off the graph, because I'd rather leave it out than guess.

Simulated is the tag to be most careful with. The threat gauge, the capability-versus-safety curve and five of the seven risk-radar axes come out of a model I wrote by hand. None of it is a benchmark or an evaluation of any named lab or model, and none of it is a forecast. What the model does is draw one picture: roughly what happens if capability keeps compounding faster than the safety work meant to keep up with it. Drag the Time Machine forward and you can watch the gap open. Look at the shape, and ignore the digits after the decimal point.

Two axes are different. Energy & Resource Strain and Compute Monopolization are plain arithmetic on the real Epoch snapshot. The first compares total IT power across the tracked sites with Epoch's own estimate for 2030; the second is a Herfindahl–Hirschman index over who owns that power. Nothing in them is hand-set. Everything else tagged simulated is a toy, and I'd ask you to treat it as one.

Enter Dashboard