Silent Sirens: AI Progress Becomes Invisible

Summary

Jack Clark (Anthropic co-founder, Import AI 438) reflects on how AI progress is becoming invisible to most people. He one-shot a predator-prey simulation with Claude Code Opus 4.5 in five minutes — something that would have taken weeks a decade ago — and uses this as a jumping-off point for a broader observation: the gap between people who work with frontier AI and those who don’t is widening into a parallel world. By summer 2026, these two populations will barely recognize each other’s economic reality.

The core metaphor: AI is an “excession” (borrowing from Iain M. Banks) — it exists in a dimension we can only see slices of. Most AI activity will be invisible to humans: agents trading with agents, token seas, websites built for AI consumption. The economy will move with counter-intuitive speed, like crypto but touching far more of regular economic reality.

Key Ideas

• The five-minute test: one-shotting a predator-prey simulation that would have been a multi-week project reveals the compression of capability — and how easy it is for non-practitioners to miss it entirely
• Parallel worlds by summer 2026: “many people who work with frontier AI systems will feel as though they live in a parallel world to people who don’t”
• Invisible AI economy: most AI activity won’t look like anything to humans — agents transacting with agents, infrastructure built for machine consumption, token flows with no human-legible surface
• The narrowing funnel: curiosity x access x ability to convert curiosity into tasks x time — each stage filters out more people, leaving a small population who actually see what’s happening
• Counter-intuitive speed: the AI economy will move like crypto markets but with real-world economic consequences, not just financial speculation
• The excession metaphor: borrowing from Banks, AI is a high-dimensional object we can only perceive in slices — and it’s incumbent on us to try to see it for what it is

The Narrowing Funnel (Mental Model)

This is the essay’s core contribution. At each stage, the population shrinks:

1. Curiosity — do you even notice something is happening?
2. Access — can you get to a frontier model?
3. Conversion — can you turn curiosity into concrete tasks for the AI?
4. Time — do you invest enough hours to build real intuition?

Most people stall at stage 1 or 2. The real filter is stage 3 — the ability to convert curiosity into agent tasks. This is a learnable skill, but it requires a mental model shift that most people never make.

Connections

This essay is a direct description of what 01-projects/phdata/index does for enterprises — helping organizations get through the narrow funnel by providing both access (stage 2) and conversion ability (stage 3). Every phData engagement is essentially widening the funnel for a client.

The “parallel worlds” thesis echoes 06-reference/2026-04-04-situational-awareness — Aschenbrenner’s argument about who has awareness of what’s coming. Clark is saying the same thing from the practitioner side rather than the policy side: awareness is filtering fast and the filtered-out population doesn’t know what it’s missing.

Stage 3 of the funnel — converting curiosity into tasks — is exactly the skill described in 06-reference/2026-04-04-talking-to-agents-is-all-you-need. Prompt engineering is the old framing; task decomposition for agents is the new one. Clark’s funnel explains why this skill is so differentiating.

The “parallel world” of agent practitioners is already forming, as described in 06-reference/2026-04-04-100x-business-with-ai. Clark gives the macro view; that piece gives the micro view of what it feels like to be inside the parallel world.

The compounding advantage of being through the funnel early connects directly to 06-reference/2026-04-04-compound-engineering. Each hour invested at stage 4 compounds — you build better mental models, delegate more effectively, and the gap widens.

SOUL.md defines Ray Data Co as an entity that’s already through the funnel. The always-on COO, the agent architecture, the vault-as-compounding-asset — this is what it looks like to have passed all four stages. Clark’s essay frames why this matters: the funnel is narrowing, and being through it early is a durable advantage.

Open Questions

• Is the funnel irreversible? Can someone who stalls at stage 2 for a year catch up, or does the compounding gap make it permanent?
• Clark says “by summer 2026” — we’re nearly there. How do we measure whether the parallel worlds have actually diverged?
• The “invisible AI economy” implies most economic value creation will be illegible to traditional metrics (GDP, employment stats). What replaces them?
• How does phData’s enterprise motion change when the client’s leadership is stuck at stage 1 versus stage 3?