Newsletter Archive

Explores the hidden microarchitectural costs of running multiple LLM agent workloads concurrently on shared GPUs via Kubernetes time-slicing, with real-world performance implications for cost-conscious deployments.

Towards Data Science · 20 min read Tools

The US government asked Anthropic to geofence Claude access and block certain countries from using its advanced models, echoing similar pressures on other AI labs. The move reflects escalating national security concerns about AI proliferation and sets a precedent for how frontier model distribution will be controlled globally.

Aljazeera · 5 min read Policy

South Korea faces a critical power shortage threatening its AI expansion plans, as the country scales semiconductor manufacturing and AI data center buildouts. Energy constraints could slow deployment of domestically-developed models and competitiveness against U.S. and China's AI capabilities.

Koreaherald · 8 min read Hardware

Claude Opus 4.8 now completes 89% of WorkBench tasks vs. GPT-4's 43% in March 2024, while unintended harmful actions plummeted from 26% to 2.5%. The study reveals capability and safety improvements go hand-in-hand rather than trade off, though frontier models still botch basic actions like sending emails to wrong recipients.

arXiv AI · 3 min (abstract) Research

LLM-as-a-Judge evaluations flip their preference verdicts 13.6% of the time on identical tasks, with GPT-4o-mini showing 72% first-position bias and cross-judge agreement dropping to 76%. The paper reveals that judges often declare winners despite tiny score gaps (0.19–0.36 on 10-point scale) and that 11+ repeated trials are needed to reliably recover ground truth—undermining the validity of single-evaluation leaderboards and reward models.

arXiv NLP · 3 min (abstract) Research

The Trump administration has blocked foreign access to Anthropic's most advanced AI models—a striking assertion of government control over frontier AI that signals tighter export restrictions ahead for the entire industry.

Bloomberg Tech · 1 min read Policy

As AI companies conduct mass layoffs, a tiny circle of founders and early investors are accumulating extraordinary wealth—widening the industry's inequality gap to breaking point.

TechCrunch · 6 min read Community

Anthropic shut down access to its most advanced models following a Trump administration order citing national security. The decision could reshape how AI companies handle model deployment and trigger similar restrictions at OpenAI, Google, and Meta.

Bloomberg Tech · 1 min read Policy