Newsletter Archive

A filmmaker exploring Sora discovered the text-to-video model frequently generated racist and sexist imagery, then found AI communities dismissing these problems rather than addressing them—raising uncomfortable questions about whose values shape these systems.

The Verge AI · 7 min read Policy

FBI director Kash Patel revealed the bureau can conduct mass surveillance at scale without AI by purchasing Americans' data directly, undermining Anthropic's refusal to let the DoD use its models for domestic surveillance. The standoff exposes a policy gap: AI firms can block weaponized use of their tech, but law enforcement sidesteps the constraint entirely through commercial data brokers.

The Guardian Tech · 1 min read Policy

Qatar's Ras Laffan facility, which produces 30% of global helium supply, halted operations after Iranian strikes on its natural gas infrastructure. A 14% cut in helium exports will hit chip manufacturers within weeks, threatening the semiconductor supply chains that fuel AI model training and deployment.

Fortune · 6 min read Hardware

OpenAI plans to nearly double its workforce from 4,500 to 8,000 employees by end of 2026 across engineering, research, product, and sales—while Anthropic now captures 70% of first-time enterprise AI purchases. The hiring push follows new DoD contracts and advanced talks with Brookfield Asset Management to deploy OpenAI tools across corporate portfolios.

Engadget · 1 min read Industry

Tutorial implementing a three-stage LLM pipeline that estimates confidence in answers, self-evaluates responses through meta-cognitive checks, and automatically triggers web research when confidence is low. Combines uncertainty quantification with autonomous information retrieval to build more reliable and transparent AI systems.

MarkTechPost · 8 min read Tools

Coding agents are fluent in Git, so humans can delegate version control complexity while staying aware of what's possible. Simon Willison explains how Git's full feature set—branches, history diving, remotes—becomes a force multiplier when agents handle the grunt work.

Simon Willison · 7 min read Tools

Supermicro insiders allegedly used counterfeit server chassis and forged export documentation to redirect Nvidia's most advanced AI accelerators to China, bypassing U.S. export controls designed to restrict AI chip access.

Ibtimes · 1 min read Policy

Four proven strategies for safely rolling out ML models to production: A/B testing splits traffic between old and new models; canary testing gradually increases traffic to the candidate; interleaved testing runs both models on identical requests to compare; shadow testing runs the new model offline to detect failures before real-world impact.

MarkTechPost · 7 min read Tools