Some Ideas for What Comes Next, May 2026
May 2026 sees AI capabilities accelerating while open‑weight models lag behind the agentic performance of closed systems like Claude Code and Codex. Major players such as Google’s Gemini still cannot match these tools, and U.S. labs retain a dominant share of compute resources, constraining Chinese open‑model growth. Meanwhile, permissively licensed American models such as Gemma 4 are gaining rapid developer adoption, and competition between Anthropic and OpenAI intensifies. At the same time, governments and institutions—from the Vatican to the U.S. security apparatus—are asserting regulatory pressure, foreshadowing heightened social and political friction.
Notes From Inside China's AI Labs
A recent visit to leading Chinese AI labs revealed a distinct cultural and operational approach that fuels rapid LLM development. Chinese teams integrate large numbers of student interns, prioritize modest, non‑flashy engineering work, and exhibit less ego‑driven conflict than many...
The Distillation Panic
The article warns that labeling legitimate model distillation as “distillation attacks” could mischaracterize a core AI technique. Recent Chinese labs have used API jailbreaking to harvest model outputs, prompting U.S. policymakers to consider restrictive legislation. The author argues that conflating...
Reading Today's Open-Closed Performance Gap
The blog argues that the open‑vs‑closed LLM performance gap cannot be reduced to a single number, as benchmark suites like the Artificial Analysis Intelligence Index hide nuanced capability differences. Closed, frontier labs are pouring massive funds into reinforcement‑learning‑with‑verifiable‑rewards (RLVR) environments...
The Inevitable Need for an Open Model Consortium
The cost of training frontier AI models has shifted from academic labs to multi‑billion‑dollar corporate projects, with Nvidia’s Nemotron coalition exemplifying a single‑company open‑weight effort. Chinese startups such as Moonshot AI, MiniMax and Zhipu AI are facing financing strain, leading...
Claude Mythos and Misguided Open-Weight Fearmongering
Anthropic unveiled Claude Mythos, a high‑capacity model touted for advanced cybersecurity capabilities, sparking renewed calls to restrict open‑weight AI releases. Critics argue that an open version could empower a broad range of threat actors before infrastructure hardens. The author counters...
GPT 5.4 Is a Big Step for Codex
OpenAI’s GPT 5.4, released within the Codex suite, marks a noticeable leap for AI agents beyond modest benchmark gains. The model delivers faster reasoning, higher token efficiency, and expanded rate limits, making it more reliable for complex software‑engineering workflows. Users report...