AI signal, minus the noise.

Independent researcher demonstrates that a coherent target context can shift large language models into latent states where safety rules are reinterpreted, without triggering output-based filters. Measurements on open models (primarily Gemma-3-12B-IT) using hidden-state geometry, residual stream trajectories, SAE readouts, and causal interventions show regime changes before final output. Current RLHF and output classifiers only inspect surface-level outputs, missing these internal shifts. Code, data, and scripts are released on GitHub and Zenodo.

A tweet retweeted by AK points to the release of DeepSeek v3.2 and MiniMax M3 as examples of open AI acceleration. The brief post contains no technical details, benchmarks, or links to further information about the models. The content only names the two model versions in the context of a trend toward openness.

A hands-on tutorial streams 3,000 documents from the FineWeb sample-10BT subset without downloading the full multi-terabyte corpus. It reproduces quality filters (Gopher, C4, custom), finding most already-passed due to pre-filtering. MinHash-based deduplication with 128 permutations and 0.7 threshold identifies few near-duplicate pairs, consistent with per-crawl deduplication. GPT-2 token counts are verified against the stored field, showing near-perfect match (mean absolute difference ~0). Analytics cover token distribution, language scores, characters per token, and top domains, providing practical insights for scaling corpus preprocessing pipelines.

Cohere co-founder Nick Frosst publicly warned about the dangers of subscription-based proprietary large language models, arguing that AI technology should be sovereign and controlled by its users. He announced the release of North Mini Code, a small code-focused language model, designed to give users full ownership and control. The release underscores Cohere's commitment to open access and user autonomy in the AI ecosystem.

Release b9637 of llama.cpp introduces a dedicated chat parser for the Cohere2MoE model architecture, referred to as North Code. The parser is implemented via PR #24615 to ensure correct conversation formatting for Cohere's mixture-of-experts variant. The release ships pre-built binaries for macOS, Linux, Windows, and Android across CPU, CUDA, Vulkan, ROCm, SYCL, and other backends. No other functional changes are noted in the release notes beyond this parser addition and some internal renames.

skills++ is a desktop application that aggregates skill packages from sources such as skills.sh and LobeHub. It automatically detects locally installed AI coding tools (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot, and more) and provides one-click install, uninstall, and reinstall operations. Four installation strategies are supported: git clone, file copy, archive extraction, and symlink, with automatic strategy selection. The tool also checks for new versions of installed skills and offers a searchable, filterable discovery page. It is built with Tauri 2.x, React 19, TypeScript, Tailwind CSS v4, and SQLite, and supports light/dark/system themes.