The Flat Curve Society

Source: Steve Yegge Date: 2026-06-19 URL: https://steve-yegge.medium.com/the-flat-curve-society-36c8b01eb33b?source=rss-c1ec701babb7------2

Summary

Steve Yegge’s long-overdue essay (voices.md had flagged it as overdue for weeks) argues against the “flat curve” position — the belief that AI model capability has plateaued. His frame: we’ve reached “the moment where model intelligence has become dangerous.” Today’s and this year’s models won’t one-shot a Fortune 100 codebase, but they are capable of remarkable work and still require “grown-up human supervision” — so engineers remain necessary, and AI literacy (training, patience, budget management) doesn’t come for free. The piece sits alongside his multi-agent orchestration framework (Gas Town / Gas City, tracked), which encodes that supervision discipline into tooling at stages 7–8 of his developer-evolution model.

Implications

• Capability bifurcation (the two clocks): Yegge is publicly contesting the “curve flattened” read on the same day the open-weight flood empirically falsifies it — GLM-5.2 quantized to GGUF/FP8, a swarm of near-frontier coding fine-tunes trending. The curve isn’t flat; it moved to a layer you have to list, not poll.
• Floor-as-supervision: “Intelligence has become dangerous → human supervision still required” is the same grain as the day’s floor releases — Claude Code fencing destructive commands in auto mode and blocking automated triggers from self-approving actions; Vibe moving credentials to the OS keyring. The orchestration author and the harness vendors converge on the same claim: the durable engineering is in bounding what the autonomous agent may do, not in the weights.
• Voices thread resolved: the essay voices.md called “overdue” landed; the author of a tracked orchestration stack (gastownhall) crystallizing a category is exactly the taxonomy-framing signal he’s tracked for.