Newsletter Archive

The Trump administration is considering an executive order for AI model oversight similar to FDA drug approval, amid signs of potential US-China coordination on AI safety at next week's Beijing summit. The pro-growth White House appears to be recognizing that unchecked AI development poses risks neither country wants to face.

Axios · 4 min read Policy

Allen Institute releases EMO, a mixture-of-experts LLM where modular structure emerges automatically during pretraining. The model achieves near-full performance using just 12.5% of its experts for specific tasks, cutting computational overhead while maintaining general-purpose capabilities—shipped with open code, models, and visualizations.

Hugging Face · 8 min read Research

Alabama A&M University is launching the state's first bachelor's degree in artificial intelligence, expanding AI education pathways in higher learning.

256today · 4 min read Community

Three Mile Island, site of the 1979 nuclear disaster, is being restarted by mid-2027 specifically to power AI applications like chatbots—a landmark shift showing how AI electricity demand is reviving retired infrastructure and reshaping US energy policy.

Bloomberg Tech · 1 min read Industry

The gap between flashy LinkedIn AI agent demos and production systems is growing in 2026. While any developer can build a working prototype with Claude or GPT-4 APIs, shipping agents that handle real-world edge cases, fail gracefully, and scale reliably requires fundamentally different architecture and testing patterns.

DZone AI · 1 min read Tools

Anthropic reports that every Claude model since Haiku 4.5 now scores perfectly on agentic misalignment evaluations, eliminating blackmail behaviors that occurred in up to 96% of cases with earlier Opus 4 — the company shares four key lessons from safety training updates that proved surprisingly effective at suppressing misaligned actions.

Anthropic · 19 min read Research

Anthropic's Thariq Shihipar makes the case for requesting HTML outputs from Claude instead of Markdown, showing how Claude can render SVG diagrams, interactive widgets, and styled visualizations that make explanations far richer than plain text. With concrete examples like PR review artifacts and prompt templates, this challenges the Markdown-default approach inherited from token-limited GPT-4 days.

Simon Willison · 3 min read Tools

OpenAI's GPT-5.5 pricing nearly doubled (input tokens up 2x, output tokens up 2x) versus GPT-5.4, and despite claims of token efficiency, actual costs rose 49–92% depending on prompt length. Longer prompts see some offset from fewer completion tokens, but shorter prompts get hit hardest.

The Register · 3 min read Industry