PTIR — Daily Briefing — 2026-09-14
by
Executive Summary
Two primary-source reads are worth attention. Anthropic CEO Dario Amodei proposes a three-stage framework for slowing frontier-model capability growth while safety work catches up: embedded independent evaluators, coordination among democratic countries, and eventual global coordination. Separately, a measured PostgreSQL study found that coding agents generally produced competent indexes but concentrated too many on write-heavy tables; in one synthetic workload, the heavier designs produced roughly 1.8 times the WAL and took about 1.9 times as long to update as the author’s baseline.
Action Queue
Read Amodei’s three concrete pacing proposals
Amodei argues that recent recursive self-improvement and agent-alignment incidents justify slowing capability growth—not stopping AI development. His most concrete unilateral commitment is to give external evaluators employee-like access at Anthropic, including the ability to publish findings subject to narrow redactions.
Recommended action: Read the opening through the three-step list, then the “Embedded Evaluators” section. Separate the verifiable proposal from the essay’s predictions: record what outside evaluators would access, what they could disclose, and what evidence would demonstrate that the arrangement is independent.
- Urgency: This week
- Importance: ★★★★☆
- Verified active: September 14, 2026
- Deadline: None announced
- Cost: Free
- Requirements: About 10–15 minutes for the selected sections; no account required
- Primary source: We Must Pace the Frontier
Add write-path evidence to reviews of agent-generated database schemas
Radim Marek audited 838 indexes from model-generated PostgreSQL schemas. Most were individually reasonable, but the agents tended to accumulate indexes on the busiest table. In one synthetic support-ticket benchmark, a 15-secondary-index design wrote 776.8 MiB of WAL and took 9,013 ms to update, versus 436.0 MiB and 4,850 ms for the seven-index baseline. These are controlled measurements, not universal performance ratios.
Recommended action: When reviewing an AI-generated schema, inspect indexes as a set. Identify the hottest tables and frequently updated columns; then check HOT-update eligibility, WAL volume, update latency, VACUUM work, index size, and actual usage counters before accepting another index.
- Urgency: This week
- Importance: ★★★★☆
- Verified active: September 14, 2026
- Deadline: None
- Cost: Article and published experiment files are free
- Requirements: Basic PostgreSQL knowledge; production validation requires representative data and workload measurements
- Original study: The unbearable lightness of one more index
- Experiment repository: boringSQL/vibe-coded-indexes
AI
The pacing essay matters less as a forecast than as a testable governance proposal. Embedded evaluators with continuing access and publication rights would be materially stronger than one-time, company-selected evaluations. Independence, access boundaries, incident-reporting rules, and redaction authority are the details to watch.
Web Development
The PostgreSQL study highlights a recurring review failure: an index can be locally sensible for one query and still be globally harmful when combined with every other “helpful” index on a write-heavy table. “Production-ready” should require workload evidence, not a larger schema.
PKb Candidates
- AI-governance evidence: Treat access, independence, disclosure rights, and enforcement as four separate tests for any external-evaluator claim.
- Index review rule: Evaluate the complete index set against both read plans and the write path; query-by-query usefulness is insufficient.