TUNDRA // NEXUS
LOC: SRV1304246| Mission Control
Woolly Newsletter — Edition #2
METR rewrites its AI productivity study after devs refuse to work without AI, Jack Dorsey cuts 40% of Block citing AI, and Anthropic drops its flagship safety pledge.
★ TOP STORIES THIS WEEK
METR's landmark developer productivity study is broken: developers are refusing to participate if it means working without AI. That self-selection is itself the data. Their raw numbers now hint at speedups (vs. the 2025 paper finding 19% slowdown), but methodology is compromised.
Dorsey cut 4,000 Block jobs (40%) while Q4 gross profit grew 24% to $2.87B and stock surged 20%. He named AI directly, not headwinds. Every CEO in America just saw the incentive structure laid bare — Wall Street prices humans as liabilities.
Anthropic is scrapping its 2023 commitment to pause AI development unless safety measures were guaranteed, replacing it with transparency-focused policies instead. Experts warn it signals safety evaluation isn't keeping pace with capability advancement.
📖 GOOD READS
GitHub engineering distills multi-agent failure patterns from Copilot work into 3 concrete patterns: typed schemas at agent boundaries, constrained action schemas, and MCP as the enforcement layer. Treat agents like distributed systems, not chat interfaces.
Tenzai tested Claude Code, Codex, Cursor, Replit, and Devin — found 69 vulnerabilities. 100% introduced SSRF. Zero apps built CSRF protection. CMU found only 10.5% of AI code is secure.