Oct 5 to Oct 11, 2026

Adds an auto-approve mode to pi where System One decision models judge whether each action is safe.

An agent does the legwork while Clef makes each judgment call from discovery to regression.

Unofficial agent skill covering typed decisions, calibration and provider differences across Jev, Clef and others.

Japanese-documented Codex skill that offloads public screen state checks to Clef-flash.

A Pi extension that builds a neutral context and lets Jev or Clef return a probability distribution over choices, to advise the agent rather than command it.

Recipes to serve Cloudflare's unquantized BF16 Clef 27B decision model with vLLM on an NVIDIA DGX Spark.

Has OpenAI Decisions, Jev and Clef read recent news, never the odds, and put probabilities on questions, then compares them with live prediction-market prices.

An experiment that lets Clef (Flash) operate a macOS desktop while keeping everything local.

Trains the 421M Laya ModernBERT model to choose Balatro actions from candidate lists, using simulator self-play and a real-game test set.

A Java worker that runs Laya's choice, score and yes/no decisions inside Camunda 8 workflows, with versioning and safety checks.

A Raspberry Pi 5 with a 3.5-inch TFT shows a pixel bot that changes colour with mood, listens continuously and transcribes locally with Whisper before answering aloud.

Localised fork with a Chinese interface, voice replies and one-click web flashing.

Short proofs for six Graffiti conjectures (198, 219, 252, 254, 584 and 712) written with Opus 5.5 and GPT-6.1 Sol.

Two-mode Codex extension: Astra with adaptive reasoning, or Sol 6.1 adaptive and Luna max agents routed by Jev.

Minecraft Java 1.20.1 magic mod written by GPT-6.1 Sol, shared as a worked example of model-authored game modding.

Posted a llama.cpp multi-GPU fork with binaries and Docker images after running five 262k-token sessions on six PCIe-only cards. Shows big-MoE long-context serving is possible on plain consumer rigs.

Investigated why Strata 0.1.40.1 left VRAM free on a 5060 Ti, 3090 and two 3060s running Flash-Next IQ3 at 262k context, and found it was by design. A useful look at how the engine sizes its caches.

Gave the same simple prompt for a widget-rich Android weather app to Flash-Next and other Qwen models to see which produced a working app.

CLI, MCP server and Chrome extension bridge that lets Clef operate your own signed-in browser session.

Runs Clef 27B and Flash 9B through llama.cpp and MLX, with benchmarks including render QA for 3D models and prompt-injection robustness.

NVFP4 quantised Clef Flash tuned for the DGX Spark box, tied to the community Balatro-playing demo setup.

Dockerised research prototype that explores language-model support for emergency triage using synthetic narratives and a local GGUF model.

Service that uses Clef on Workers AI to flag possible checkout losses from store footage.