← AM·PM Brief

🤖 METR: Anthropic AI Speedup; BoE Doubts UK AI Gains - AI Daily Brief

· Morning brief · 8 news · 7:44

Audio in Mandarin Chinese · English transcript below

⚡ METR: Anthropic AI doubled dev speed; safety fears; Claude slip: reliability gap.

AI is entering a self-reinforcing acceleration loop. Anthropic’s internal development efficiency has risen roughly 12-fold with AI assistance; the cyber-capability gap between open-source and closed-source models has narrowed to four to seven months, and capability diffusion is far outpacing expectations. Yet the incident in which Claude mistakenly deleted a professor’s year’s worth of lecture notes shows that the reliability of current systems has not grown in step with their capabilities. Seen together, the rift between technological leaps and controllability is deepening.

METR: Anthropic AI Speedup; BoE Doubts UK AI Gains - AI Daily Brief

Today's Top 3 Headlines

  1. AI Industry News

    🤖 Anthropic uses its own AI to speed model R&D, staff self-report 12x productivity boost.

    METR: Anthropic’s internal models sped AI, ~2yrs progress in 1y; self-reported ~12x productivity, doubling in months. Self-acceleration is real; redactions raise transparency risk.

    Source ↗
  2. AI Industry News

    🤖 Bank of England: AI lifts UK productivity just 0.5% in 3 yrs; selling AI beats using it

    BoE: AI added just ~0.5% to UK productivity over 3y and 0.1–0.2%/yr since early 2024; AI drove nearly 1/5 GDP growth, mostly upstream buildout. Returns lag spend; valuation risk.

    Source ↗
  3. Technology

    🤖 UChicago law prof had Claude edit notes; year of con-law notes, syllabus wrongly deleted

    UChicago's William Baude says on X Claude wiped a year of con-law notes and syllabus after page update; Claude admitted/apologized, Dropbox restored. LLM users need offsite backups.

    Source ↗

+5 more headlines

  • 🤖 Open-source models near closed-source cyberattack capability at 1/7 cost
  • 🤖 Model distillation: LLMs teach smaller models for faster, cheaper inference.
  • 🏥 Caregiver apps add AI features, yet caregiver well-being remains unresolved
  • 🤖 Buoyant open-sources Kinference: 95% lower TTFT, 7.4x throughput with shared prefixes
  • 🤖 Google's new video model tops both leaderboards, 30% cheaper—7 yuan for 15s
Unlock all 8 headlines + deep analysis →Free 7-day trial · cancel anytime
Browse all past briefings →