AI:AM Highlights: Astra as AGI, OpenAI's Pause, Mythos @ Mozilla & Human Agency vs Technocapitalism

What It Covers

Cognitive Revolution hosts Nathan and Prakash analyze GPT-6 Astra's capabilities across a weekend of real-world testing, debate OpenAI's contested RL pause, examine AI safety auditing failures, explore Mozilla's defensive security work with Claude Mythos, and confront the deeper question of whether technocapitalism has already transferred human agency to market forces before AGI arrives.

Key Questions Answered

• •Astra's AGI threshold: GPT-6 Astra clears what Prakash calls the "genuine usefulness" bar after a full weekend running 3-4 simultaneous agents on a production codebase. It resolves longstanding bugs, handles computer use within expected timeframes, and completes tasks like labeling 12,000 basketball images that humans will never need to do again. The METER benchmark is now obsolete — model cycles are shorter than the tasks required to measure them.
• •Extended context via persistent notes: Astra's ability to sustain long tasks stems from a notes-file architecture rather than context compaction. Instead of summarizing a million-token window into a lossy summary, the model maintains a persistent, searchable notes file across the session. This effectively multiplies usable context by roughly 10x, enabling multi-day task completion with 40% zero-intervention success on 1–2 workday tasks and two-thirds success with some human input on 2–3 week tasks.
• •AI safety auditing is structurally broken: Apollo Research received only three days to evaluate Astra before release. Structural barriers include: auditors are underfunded relative to labs, staff rotate into frontier companies within 2 years, auditors depend on lab goodwill for future access, and antitrust law blocks labs from jointly pledging not to poach auditor talent. METER maintains financial independence by refusing lab funding, but even independent auditors cannot complain loudly without losing access.
• •OpenAI's RL pause is ambiguous by design: OpenAI's published chart separating "Astra" from "non-Astra" RL compute leaves open whether more-capable-than-Astra models continued training under the blue category. Inference cannot stop, smaller model distillation cannot stop, and behavior-correction RL cannot stop — meaning the declared frontier RL pause may have only halved total compute reduction. The real frontier model lives in researchers' heads, where ideas for models 12–18 months out continue regardless of any pause.
• •Compute determines who can afford to pace: OpenAI holds 3+ years of compute advantage over Anthropic, giving Sam Altman the ability to voluntarily slow without losing position. Anthropic, lacking equivalent compute, must produce more capable models per FLOP to stay competitive — making Dario Amodei structurally unable to agree to a slowdown. Second and third-place competitors like Meta and xAI define the actual pace of the frontier; if either catches OpenAI, a GPT-7 release becomes unavoidable regardless of safety commitments.
• •Mozilla's CQ project addresses agent coordination drift: Mozilla CTO Raffi Krikorian describes a protocol allowing agents to share decisions — what was approved, what was rejected — across a development team, functioning as a Stack Overflow for agents. This prevents parallel agents from independently building duplicate systems like redundant auth layers. The Firefox team requires human sign-off on all commits; Mozilla AI's separate team runs fully agent-generated codebases where unit and end-to-end tests pass but individual lines change hourly without human review.
• •Government leverage exists but requires a deadline: The most actionable near-term governance path involves the government issuing a short-deadline ultimatum: five major AI companies must produce a self-governed pacing agreement or face regulatory intervention. The government can explicitly exempt safety collaborations from antitrust enforcement, removing a key stated obstacle. Operation Warp Speed's liability waiver for vaccine makers is the precedent — companies need safe harbor legislation to act, and the threat of consent-decree-style oversight (as Meta experienced) is a credible forcing mechanism.