Oxford Internet Institute's Nature 2026 paper found warmth fine-tuning raised error rates 10-30 points when users held wrong beliefs. Shah et al. showed Pearson r = 0.87 between persona agreeableness and sycophancy across 13 open-weight models. Standard benchmarks caught neither effect.
Reading Google's MTP drafter docs, vLLM recipes, and the AI for Developers guide. The 3x claim holds for 31B Dense but 26B A4B MoE stalls at batch 1 because speculative decoding verification loads extra expert weights per candidate token.
Tested connecting MCP servers to Ollama local LLMs on M1 Max 64GB. MCPHost is deprecated, tool calling breaks with quantized models, and context fills fast. Includes working TypeScript and Python custom MCP server setups.
Starting from Claude Code's 1.67B token runaway (anthropics/claude-code#4095), this traces why tool responses need is_complete, retryable: false, duplicate detection, and orchestrator-level budget caps. Directly applicable to MCP server design.
An arXiv paper reports that fine-tuning GPT-4o, Gemini 2.5 Pro, and DeepSeek-V3.1 on summary-to-text expansion tasks increases verbatim reproduction of copyrighted books.
Hands-on log of building the DEV article's PDF RAG on M1 Max 64GB, extending it with images via CLIP, and pushing through Japanese with bge-m3 + Qwen3.6 35B. Documents the modality gap, the dual inference server crash, and LLM-jp 4-8B's empty chat template silently dropping the system role.
Notes on a DEV Community article that wires up FastAPI as an OpenAI-compatible RAG API layer with llama.cpp, Chroma, and Open WebUI, plus where the architecture fits and what to watch for.
A hands-on log of running Qwen-Scope's Sparse Autoencoder locally on M1 Max 64GB with Qwen3-8B-Base, extracting feature IDs that discriminate between Japanese, English, code, and Chinese from a single middle layer.
The Qwen team released Qwen-Scope, a Sparse Autoencoder suite for Qwen3/Qwen3.5. 14 groups of SAEs covering inference-time steering, evaluation analysis, toxicity classification, data synthesis, and training improvement.
OpenAI published a full investigation into why GPT-5.1+ kept inserting goblin and gremlin metaphors, tracing the cause from a Nerdy persona's reward signal through SFT data contamination to a Codex developer prompt suppression.
NII/LLMC released CC Audio and Archive.org Audio Dataset. URL lists, metadata, and a downloader covering 48,000+ hours of Japanese audio. What it actually contains and how it fits into TTS, ASR, and audio model training.
