The Closing of the Frontier

Essay by Tanya Verma arguing that restricted access to frontier AI models (specifically Anthropic’s Mythos/Glasswing) represents a “closing of the frontier” analogous to Frederick Jackson Turner’s 1893 thesis about the American West.

Key Claims

  • The internet was the last permissionless frontier — a poor person and the wealthiest had access to the same tools. AI is changing this.
  • Anthropic’s Glasswing announcement restricts Mythos to enterprise partners, creating a two-tier AI access system
  • State-scale capabilities (zero-day discovery, exploit generation) are being privatized to a handful of organizations without state-scale accountability
  • Anthropic is simultaneously manufacturer, regulator, and appeals court — violates separation of powers principle
  • MATS researchers (serious AI safety program) use Chinese open-source models because they can’t get frontier access
  • The “mainframe era” analogy: open-source models run 3-12 months behind frontier, like the Apple II vs mainframes

Key Arguments

  • For access: public access forces latent capabilities into the open (better for safety via “fail fast and fix”), API access doesn’t mean full ownership but allows innovation
  • Against restriction logic: security is always an arms race (cf. fuzzers → CI pipelines), restricting access stifles defensive innovation
  • Neofeudalism (George Hotz): those with capital when labor-replacing AI started have permanent advantage
  • Should have FOIA-style obligations, clear access criteria, appeals mechanisms

Takeaways

  • Deeply resonant essay — touches on the fundamental tension between AI safety and AI access
  • The comparison to nuclear non-proliferation is flawed: intelligence is economically valuable in a way nukes aren’t
  • Rudolf Laine’s point: capital now trivially converts to superhuman labor
  • The essay is a response to Glasswing specifically but represents a broader concern about AI centralization

Connections