Interface Beats Model — the throughput-vs-reading wall

The model already won the volume war. That is not a prediction — it happened sometime in the last two years, and the gap is now widening by about a quarter every month. This is a data check on the claim in Winking at a Touchscreen that the bottleneck in agentic AI is the interface, not the model. If machines write four times what we can read, then more capability is moot until a human can actually see it.

What the cited figures say

Every number below is a public, cited statistic — no dataset was downloaded or fabricated. Full provenance in data/interface_beats_model.source.md.

AI token throughput (Google, via the essay’s own footnotes):

| Point | Tokens/month | Tokens/day | |---|---|---| | Apr 2024 | ~9.7T | 3.2e11 | | May 2025 | ~480T | 1.6e13 | | Jun 2026 | ~3.2Q | 1.05e14 |

That is ~330× in 26 months — roughly +25%/month, the volume doubling every ~3 months.

Human reading capacity (flat — it’s biology):

• Reading rate: 238 wpm (Brysbaert 2019, meta-analysis of 190 studies).
• Literate population: ~7 billion (World Bank / UN).
• At the essay’s stated 45 min/day reading AI output: ≈ 1.0e14 tokens/day.

Loading chart…

The result

Two ways to read the surplus, both from the essay’s own figures:

• Google infrastructure alone: ~1.05e14 tokens/day (Jun 2026) against ~1.0e14 tokens/day of human capacity — a cross-over that just happened, barely above 1:1 on the conservative reading-time assumption.
• All providers combined: the essay cites ~310 trillion words/day globally = ~4.13e14 tokens/day — roughly 4.1× humanity’s reading capacity.

So the essay’s line — “we generate three times more than we can physically consume” — is if anything conservative. The direction is right and the magnitude is in the 1×–4× band depending on whether you count one provider or all of them.

And it is accelerating. The surplus did not exist two years ago (Google alone was ~0.003× capacity in Apr 2024). Today it is 1×–4× and compounding at ~25%/month.

The wall (what this cannot show)

• No open timeseries of token throughput exists. Google’s figure is a keynote quote (cite-only). The exponential direction is established by Google’s own three datapoints, but a continuous trend line is the wall.
• Reading-time of 45 min/day is the essay’s stated figure, illustrative not measured. The 1×–4× band is the honest uncertainty, not a precision.
• “Interface beats model” is supported, not proven. A generation≥consumption surplus means more model capability is moot without better interfaces — but the proving evidence (Claude Artifacts A/B error-reduction, multimodal agent-eval) is still cite-only or absent. That is the next Observation to land.

Why it matters

lens:deploy-dont-wait. Winking’s point was never “the model is weak” — it was that the interface is the constraint. A 1×–4× generation-surplus makes that concrete: the bottleneck isn’t how much the machine can write, it’s whether a human can see, redirect, and absorb it. More tokens don’t help if the postal system is still “lick the stamp.”

→ OfTech takeaway: The model already won the volume war. The interface lost it. Fix the postal system. → See also: Winking at a Touchscreen (parent essay, now cross-linked), and observations-winking-watch.md for the trigger that completes the proof.