📣 Headlines

• Claude’s Slack-based coding beta can create and modify files, run tests, and iterate from chat, pushing “vibecoding” deeper into everyday workflows.

• Google launched managed MCP servers for services like BigQuery, Maps, and GKE to simplify connecting AI agents to cloud tools and data sources.

• The AI training-data market is booming as expert-driven data and RLHF rubrics suppliers scale up to meet demand for higher-quality model alignment and evaluation.

• The EU is investigating Google over AI-generated search summaries, focusing on publisher compensation, data use, and opt-out mechanisms.

• Intel is reportedly in talks to acquire inference-chip startup SambaNova for about $1.6B, signaling continued consolidation pressure in AI hardware.

• Couchbase rolled out a unified platform for AI data, vector search, and governance aimed at simplifying and speeding up agent development.

• Oracle shares fell after earnings missed expectations, with investors scrutinizing AI-infrastructure growth and broader “AI bubble” fears.

• AI-enabled kids’ toys are surfacing safety and propaganda risks, underscoring gaps in regulation, content controls, and data protections for consumer LLM deployments.

🔧 Company Engineering Blogs

Deepening our partnership with the UK AI Security Institute (deepmind​.google). Google DeepMind and UK AI Security Institute expand a research partnership to study monitoring AI reasoning, socioaffective impacts, and economic evaluations

Background Coding Agents: Predictable Results Through Strong Feedback Loops (Part 3) (engineering​.atspotify​.com). Strong verification loops with Maven verifiers, an LLM judge, and sandboxed agents drive reliable background code changes in Spotify's CI/CD workflow

4x Faster: How AI-Assisted Development Accelerated Building New SQL Dialects for Zero Copy Connectors (engineering​.salesforce​.com). AI-assisted dialect generation and automated 25,000-query validation accelerate Zero Copy connectors across 16 dialects, expanding from 5 to 100+

The future of AI-powered software optimization (and how it can help your team) (github​.blog). Explores Continuous Efficiency: AI-enabled, agentic workflows for green software on GitHub, using Agentic Workflows and LLMs

Codex is Open Sourcing AI models (huggingface​.co). Hugging Face demonstrates Codex, OpenAI Codex integration with HF Skills for end-to-end ML experiments and model fine-tuning

🔓 Open Model Drops

Olmo 3 and the Open LLM Renaissance (cameronrwolfe​.substack​.com). Olmo 3 releases fully-open training artifacts for reproducible open-LLM research and transparency-driven benchmarking

Devstral 2 (simonwillison​.net). Devstral 2: Mistral's 123B and Small 24B models for coding agents, with licensing contrasts and SVG prompt examples

2025-12-11 22:35 (aicode​.danvoronov​.com). Devstral 2 (123B) and Devstral Small 2 (24B) from Mistral AI, Vibe CLI, SWE-bench performance, open licenses, local GPU viability

Gemma 3 AI model in Clojure (dragan​.rocks). Gemma 3 AI model in Clojure demonstrates ONNX runtime integration for loading and running a 1B parameter Gemma 3 LLM in main memory with oneDNN

Introduction to Qwen3-VL (debuggercafe​.com). Qwen3-VL 4B-Instruct and Thinking models enhance multimodal understanding with OCR, object detection, and video tasks using Python and Transformers

The $250 Million Paper (blog​.bytebytego​.com). Molmo builds a strong vision-language model from scratch using PixMo datasets for open training and improved grounding

🌍 Policy & Ecosystem

Why AI reading science fiction could be a problem (transformernews​.ai). Explores how science fiction and misalignment research influence AI training, with examples from Anthropic, Redwood, and policy options

State of AI : Une étude empirique sur 100 000 milliards de tokens d'interactions LLM réelles (openrouter​.ai). Empirical study of 100 trillion tokens reveals OSS vs proprietary usage, roleplay and programming focus, and agentic inference trends across regions

Agentic AI: Governance, risks, and responsible deployment in the social sector (merltech​.org). Explores Agentic AI, its differences from GenAI and AI agents, governance, consent, risks, and multi-stakeholder accountability in the social sector

Deepening our partnership with the UK AI Security Institute (deepmind​.google). Google DeepMind and UK AI Security Institute expand a research partnership to study monitoring AI reasoning, socioaffective impacts, and economic evaluations

Cryptographers Show That AI Protections Will Always Have Holes (quantamagazine​.org). Cryptographers show two-tier AI protections have inherent gaps, exploitable via time-lock puzzles and substitution ciphers

🎬 AI in Media

All AI Videos Are Harmful (idiallo​.com). Explores AI video tools like Sora, Runway ML, Veo; warns of misuse by scammers and erosion of trust in visual media

Users are complaining OpenAI wreaked Sora 2. EOSHD takes a look at what happened… (eoshd​.com). OpenAI's Sora 2 struggles with scaling, provoking investor controversy and debates on compute costs and accessibility

Unleashing Autonomy: Getting Daily News Updates (danraine​.substack​.com). Daily AI news on film, using Perplexity Gemini 3 to surface articles from Disney licensing to AI actors and copyright topics

Doblando un podcast con iA (victorcorreal​.substack​.com). Estrategias para doblar podcasts al inglés usando IA, pruebas con herramientas de vanguardia y ejemplos de implementación

🧑‍💻 Coding Culture

Ask HN: How can I get better at using AI for programming? (news​.ycombinator​.com). Tips for AI-assisted programming with Claude and Plan mode, CLAUDE.md, memory, and browser checks to speed Svelte/Django workflow

Claude Code Changed How I Work (Part 1) (causalinf​.substack​.com). Economist Scott Cunningham chronicles using Claude Code for AI-driven coding, autos inside local directories, and multi-part insights on attention and collaboration

If you're going to vibe code, why not do it in C? (stephenramsay​.net). Explores vibe coding, its relation to C and assembly, and the idea of a dedicated vibe-oriented language for AI-assisted programming

How to review AI Generated PRs (kevinjmurphy​.com). Guidance on reviewing AI-generated pull requests with caution, focusing on safety, quality, and collaboration across teams

Responding to "The highest quality codebase" (schneidenba​.ch). Claude-driven refactoring experiment: quality vs vanity metrics, prompts, and lessons from a 47k–120k LOC codebase with test and comment growth

🧰 Agent Tooling

Stop using GitHub Copilot as a chatbot! (m365princess​.com). Stop using Copilot as a chatbot; shift decisions into stable repo files to make it a reliable tool

Supercharge Your Design System with LLMs and Storybook MCP (tympanus​.net). Using LLM coding agents with Storybook MCP to build high-quality UI components in a design system

Quoting OpenAI Codex CLI (simonwillison​.net). A look at OpenAI Codex CLI usage, prompts, and skill rendering in Rust with progressive disclosure across example prompts and YAML-driven triggers

Advent of AI 2025 - Day 8: Messy Data to Structured Output (nickyt​.co). Messy napkin notes to structured JSON and an HTML site using Goose, MCP, Netlify, and Day 4 styling

How to Customize Your Claude Code Status Line (alexop​.dev). Step-by-step guide to create a Claude Code status line that shows model, context usage, and costs in your terminal using bash and jq

You are holding GitHub Copilot Wrong! (m365princess​.com). Critical take on GitHub Copilot: prompts, memory, context windows, and the need for structured workflows in software development

🧱 RAG & App Architecture

The Stability-Plasticity Dilemma: How Memory Architectures Are Solving Continual Learning (rewire​.it). Explores memory-augmented and retrieval-based approaches to continual learning in LLMs, including sparse memory finetuning, RAG, and model merging

Personal, Agentic Assistants: A Practical Blueprint for a Secure, Multi-User, Self-Hosted Chatbot (towardsdatascience​.com). Self-hosted, multi-user agentic chatbot with private file access, vector search, and tool-based reasoning using LangGraph, Postgres, MinIO, Ollama, and RabbitMQ in Python

SE Radio 698: Srujana Merugu on How to build an LLM App (se-radio​.net). Srujana Merugu discusses building LLM apps, RAG, agentic architectures, evaluation, safety, prompts, and multi-modal trends

Building a "Lawyer GPT" for Your Blog - Part 9: Document Ingestion with Docling (mostlylucid​.net). Ingests PDFs, DOCX, and more using Docling to power a Lawyer GPT with C#/.NET, Docker, OCR, and vector search

SOTA RAG & Memory without the database: Files, Git, and simple folders (nijho​.lt). Bas Nijholt outlines a file-based RAG and memory approach using Markdown, Git, ONNX Runtime, and a local document folder

How OpenAI, Gemini, and Claude Use Agents to Power Deep Research (blog​.bytebytego​.com). Deep Research workflows across OpenAI, Gemini, Claude, Perplexity, and more using multi-agent orchestration and specialized tools

Analyzing an email dump with BigQuery and Gemini (corinfaife​.co). Analyzes an Epstein email dump using BigQuery and Gemini to extract metadata, summarize at scale with Vertex AI, and present results via a Streamlit interface

🏗️ Systems & Serving

chores 0.3.0 and local LLMs (simonpcouch​.com). Local LLMs power chores helpers for R, with Qwen3-4B-2507, LM Studio, Ollama, and GPT-4.1/Claude prompts guiding roxygen2 templating

SGLang Adds Day-0 Support for the Highly Efficient, Open Nemotron 3 Nano Hybrid MoE Model (lmsys​.org). SGLang adds Day-0 support for NVIDIA Nemotron 3 Nano hybrid MoE model, enabling open-source, efficient AI agent development with BF16/FP8 configurations

Neo, 6 years and 600 citations later (rmarcus​.info). Neo, a learned query optimizer using deep reinforcement learning and tree convolution, details its journey to 600 citations and industry impact

New in llama.cpp: Model Management (huggingface​.co). llama.cpp adds router mode for dynamic multi-model loading and management in a lightweight HTTP server

Fedora Magazine: Find out how your Fedora system really feels (with the linux-mcp-server!) (fedoramagazine​.org). Explains linux-mcp-server and MCP, enabling LLMs to query Fedora system data for troubleshooting and upgrades

Build an AI inference server on Ubuntu (gjolly​.fr). Build a local Ubuntu AI inference server using Ollama and Open WebUI for privacy-preserving LLM chats

🌫️ Diffusion Generation

Text Diffusion Models are Faster at Writing Code (nathan​.rs). diffusion language models accelerate code generation by higher parallel decoding and structured output, demonstrated with Fast dLLM v2 on code vs unstructured text

Omni-Attribute: Open-Vocabulary Attribute Encoder for Visual Concept Personalization (opencv​.org). Omni-Attribute enables open-vocabulary, attribute-specific embeddings for controllable diffusion, combining image + text pairs to disentangle attributes like identity, lighting, and style

Implementing Diffusion Models (vkethana​.com). Explores diffusion models with CFG, SDEdit, inpainting, image editing, and image-to-image techniques using Python code and DeepFloyd IF

📏 Evals & Benchmarks

Open Source Replication of the Auditing Game Model Organism (lesswrong​.com). Replication of an auditing model organism using Llama 3.3 70B Instruct to test reward model biases and auditing techniques

What are AI Evals? (nickyt​.co). How AI evals score outputs, use guardrails, and optimize models with Galileo and LLMs for testing non-deterministic AI systems

Gemini 3 Pro vs GPT 5.2: A Blind Test Reveals Code Red DNA (prashamhtrivedi​.in). Blind A/B test compares Gemini 3 Pro and GPT 5.2 on converting slash commands for sandboxed coding agents, revealing training-driven behavior and salience issues

FACTS Benchmark Suite: Systematically evaluating the factuality of large language models (deepmind​.google). FACTS Benchmark Suite evaluates LLM factuality across Parametric, Search, and Multimodal reasoning with public/private sets and a Gemini-led performance showcase

Validating LLM-as-a-Judge Systems under Rating Indeterminacy (blog​.ml​.cmu​.edu). Framework for validating LLM-as-a-judge under rating indeterminacy; multi-label elicitation, translation matrices, and downstream evaluation tasks

Blog/2025-12-11/LLMs Excel At Easy Verification Problems (wiki​.roshangeorge​.dev). LLMs excel at checkable problems and reasoning, using verification loops over memory and minimal reproducible examples

The Missing Piece in AI Safety (cafebedouin​.org). Structural evaluation: rethinking AI safety by auditing evaluators and metrics that define 'smart' and safe

Funktion und Schwächen von KI-Benchmarking (stuker​.com). Über KI-Benchmarks, syn­thetische vs. menschliche Bewertungen, Evals, Memorization, Reward Hacking und der Schutz geistigen Eigentums

🧠 Understanding Model Behavior

NeurIPS 2025 Best Paper Review: Qwen’s Systematic Exploration of Attention Gating (towardsdatascience​.com). Systematic study of gating in SDPA attention for LLMs using Qwen’s methods; practical insights in gating placement, multiplicative gating, head-specific gates, RoPE, YaRN, and code in Python

I. Drafts (studium​.dev). LLMs as Meaning Optimization explored using Quartz v4.2.3 in a JavaScript/TypeScript context by jerlendds

Some evidence against the idea strange CoT stems from incentives to compress language (lesswrong​.com). Incentives to compress language in RL-tuned LLMs are explored via CoT entropy comparison across RL'd and instruct-tuned variants of Qwen models

Why Vision Language Models Ignore What They See with Munawar Hayat - #758 (twimlai​.com). Munawar Hayat discusses NeurIPS 2025 work on multimodal AI, object hallucination, attention-guided alignment, generalized contrastive learning, and MultiHuman Testbench

Beichtstuhl für LLMs scheint eine gute Idee zu sein (stuker​.com). Beichtstuhl-L concept for LLMs explores confession-based honesty, RL rewards, and training gaps with GPT-5-Thinking experiments

In Defense of Curiosity (davidbau​.com). Explores pragmatic interpretability, three research perspectives, Copernican shift in AI thinking, and the role of curiosity in human-AI collaboration

Cracking the Code: Tackling AI Hallucinations in the Quest for Reliable Language Models (eliza-ng​.me). Explores AI hallucinations in LLMs, grounding, confidence scores, and hybrid data integration for reliable language models

Shortwave: Beyond Hallucinations: The Illusion of Understanding (mamund​.substack​.com). Explores epistemic drift in LLMs, framing with Korzybski and Rose-Frame model, urging cautious use and human-centered reasoning

📚 Academic Research

VL-JEPA: Joint Embedding Predictive Architecture for Vision-language (arxiv:cs). Meta FAIR proposes VL-JEPA, predicting continuous text embeddings instead of tokens. It cuts parameters 50% and enables selective decoding, boosting VQA/retrieval efficiently for multimodal deployment

Exploring MLLM-Diffusion Information Transfer with MetaCanvas (arxiv:cs). MetaCanvas turns MLLMs into latent-space planners for diffusion generators. It improves layout control, attribute binding, and video editing across backbones, bridging understanding-to-generation via lightweight interfaces

Mull-Tokens: Modality-Agnostic Latent Thinking (arxiv:cs). Mull-Tokens adds 10–40 discrete latent scratchpad tokens to multimodal LLMs. Trained with curriculum and optional RL, it boosts visual reasoning cheaply without verbose CoT outputs

HybridToken-VLM: Hybrid Token Compression for Vision-Language Models (arxiv:cs). HTC-VLM compresses hundreds of vision tokens into one via hybrid continuous patches plus discrete semantic anchors. It retains ~87% benchmark performance at 580-to-1 ratio overall

Learning Unmasking Policies for Diffusion Language Models (arxiv:cs). This paper trains RL policies to choose which tokens to unmask in masked diffusion language models. Learned samplers improve quality–throughput, transfer across models and lengths