Nolan Warns AI Is a "Trojan Horse"

Christopher Nolan, director of *The Odyssey*, has described AI as "a Trojan horse that everyone knows has Greek soldiers inside." It's a short TechCrunch piece, but it's drawing attention as a clear warning about the technology from a prominent creator.

Current AI Races to Build "the WWW for AI"

The nonprofit Current AI, aiming to build AI that "leaves no culture behind," is moving fast on a platform that anyone can use for free across multiple devices and chat interfaces. The move sharply frames a "decentralized, public" counterweight to centralized mega-APIs, putting the issue of public AI infrastructure front and center once again.

Pro-AI Super PAC Holds $31 Million for the Midterms

According to Axios, a super PAC friendly to the AI industry is holding $31 million in funds for the US midterm elections. Political money that foregrounds supporting innovation over regulation is starting to move, suggesting that the map of policy debate is shifting in parallel with the technology development itself.

YouTube's "S-CTS": The AI Slop Purge Has Finally Arrived

YouTube has launched a full-scale crackdown on low-quality AI-generated content (so-called "slop") on its platform through a new system called "S-CTS." Reported by Inside the Creator and gaining traction on Hacker News, the story signals that "quality judgment of AI content" in the creator economy is becoming visible as a platform-side standard.

1-bit LLM "Bonsai": An 8B Model That Runs in Just 1.15GB

"Bonsai," a 1-bit LLM family announced by Caltech-born startup PrismML, is making big waves in the local LLM scene. Rather than conventional quantization, the models are trained in 1-bit (only +1/-1) from the training stage onward, which reportedly lets an 8B-class model run in 1.15GB with roughly 1.5GB of required memory. It's released under an Apache license, and the claim is that ordinary PCs without a GPU can run inference at practical speeds. A detailed usage guide has been published on Zenn.

AI Agent Memory "Starts Rotting the Moment You Write It"

When you run the same project across multiple agents and multiple sessions, the accidents of "stepping on old records as if they were current facts" quietly take their toll. A separate article reporting on nine months of agent operations concluded that "the enemy of a memory system isn't retrieval accuracy but freshness," and setups that manage freshness with plain Markdown rather than RAG or vector DBs are being reevaluated. Shared memory alone isn't enough — explicit invalidation and update mechanisms are what's needed.

"katalint" Arrives for Linting AGENTS.md / CLAUDE.md

As Claude Code and Codex operations mature, the quality of instruction files on repositories (AGENTS.md, CLAUDE.md, subagent definitions, and the like) is starting to shape agent behavior. A linter called "katalint" has been released to tackle "bloat," "ambiguous references," and "going stale without updates." Along the same lines, Qiita has posted a practical guide covering the 30-line rule, imports, monorepo support, and more — a sign that the design of agent instruction files is finally being recognized as an engineering discipline.

---

Other Briefs

Webflow's security team** looks back on how "AI amplified rather than replaced" their work

A price comparison of 145 LLM/STT/TTS models** aggregated on hail.so is drawing attention on Hacker News

CallBro**: A meeting notes tool that lets you choose Codex / Claude Code / a local LLM has arrived

Visuali**: An AI agent that generates and edits images on an infinite canvas

An attempt to peer into **PLaMo's intermediate layers** with a Jacobian Lens (research preview)

Findings that **Reasoning Effort is not an "IQ knob"** (GLM-5.2 / Codex)