Has AI already killed self-help nonfiction books?

Tim Ferriss highlights a significant decline in prescriptive nonfiction sales, projecting an 80% drop in his own catalog by 2026 as LLMs become the preferred interface for personalized, "how-to" information. He argues that while the market for raw information is being commoditized by chatbots, a durable moat remains for "transformational" content rooted in unique voice, storytelling, and human connection. This shift suggests that creators must move beyond instructional data to focus on high-density, long-form experiences that AI cannot easily replicate.

Read more

Running local models is good now

Local models like Gemma 4 and GPT-OSS have reached a performance inflection point, enabling agentic coding workflows with approximately 75% the accuracy of frontier models on consumer hardware. By pairing inference engines like LM Studio or llama.cpp with agent harnesses like Pi, developers can execute complex tasks such as code refactoring and repo bootstrapping within secure Docker environments. Despite hardware-limited context windows and high KV cache demands, the local ecosystem now offers significant advantages in privacy, cost, and model introspection.

Read more

GPT‑NL: a sovereign language model for the Netherlands

GPT-NL is a sovereign LLM initiative led by TNO, SURF, and NFI, supported by €13.5 million in Dutch public funding to ensure digital autonomy. The model is being trained from scratch to guarantee data provenance, transparency, and compliance with European privacy and copyright standards. Key technical outputs include open-source code and model weights available under controlled licensing, with a specific focus on energy-efficient training and a reciprocal data ecosystem.

Read more

Qwen-Robot Suite: A Foundation Model Suite for Physical World Intelligence

The Qwen-Robot Suite bridges the gap between VLM perception and physical control through three specialized foundation models: Qwen-RobotNav for unified navigation, Qwen-RobotManip for cross-embodiment manipulation, and Qwen-RobotWorld for predictive world modeling. By aligning vision-language representations with physical action spaces—using techniques like canonical state-action representations and natural-language action interfaces—the suite enables scaling across heterogeneous robot data and embodiments. These models function as low-level primitives that high-level LLM planners can invoke to execute complex, long-horizon agentic tasks in real-world environments.

Read more

Microsoft turns to AWS as GitHub faces AI capacity crunch

Microsoft is leveraging AWS capacity to stabilize GitHub following a massive surge in "agentic development" that has strained existing infrastructure and outpaced its Azure migration timeline. With commits projected to scale 14x by 2026, the platform is adopting a multi-cloud strategy to handle the machine-generated load and mitigate outages that threaten its position against AI-native competitors like Cursor. This shift underscores a broader trend of hyperscalers seeking bridge capacity from rivals as AI demand exceeds internal infrastructure planning and hardware availability.

Read more

Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing

LLMs exhibit model-specific priors by generating recurring, correlated "ghost" character ensembles when creating fictional experts, such as Claude’s "Elena Vasquez and Marcus Chen." These behavioral fingerprints enable the identification and dating of AI-generated content, revealing over 1,600 ghost-authored records with real DOIs on platforms like Zenodo. This phenomenon pollutes scholarly aggregators and provides a temporal proxy for model deployment windows through synthetic research groups.

ArXiv

Artificial Intelligence Index Report 2026

The ninth AI Index report highlights the widening gap between rapid technological advancement and the lagging infrastructure for governance, evaluation, and data tracking. It introduces more ambitious benchmarks for reasoning and safety, analyzes the economic impact of generative AI, and features new dedicated chapters on AI's transformative role in science and medicine.

ArXiv

Brick: SOTA LLM Routing

Brick is a multimodal LLM router that replaces surface-feature routing with a six-dimensional capability scoring system and per-query difficulty estimation. It utilizes a cost-penalized geometric rule and a tunable preference knob to optimize the trade-off between accuracy and inference spend. Benchmarks demonstrate that Brick outperforms individual frontier models in accuracy while offering up to a 22.15x cost reduction and significantly lower median latency.

ArXiv

On The Architectural Complexity of Neural Networks

This framework models tensor operation structures to analyze architectural complexity and automate the construction of novel DNNs. By correlating 40 years of architectural breakthroughs with complexity increases, the study identifies unexplored model classes and provides a dataset of 3,000+ high-complexity architectures.

ArXiv

Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning

Agentic automata learning evaluates LLM agents' ability to uncover hidden DFAs using membership and equivalence queries. Although reasoning models outperform non-reasoning variants, performance scales poorly with DFA size due to failures in query planning and hypothesis construction. Current LLM agents demonstrate non-trivial discovery capabilities but lack the robustness and efficiency of classic automata-learning algorithms.

ArXiv

git-lrc – Free, Micro AI Code Reviews That Run on Git Commit

git-lrc is an open-source pre-commit hook that automates micro AI code reviews to mitigate silent failures and technical debt introduced by GenAI. It utilizes a BYOK architecture supporting Gemini, OpenAI, Claude, and DeepSeek to analyze diffs against 100+ failure patterns across security, reliability, and performance. Key features include structured issue navigation, automated change summaries, and version-controlled repository rules to align LLM feedback with team-specific standards.

Repo

Ctx, save tokens by loading only the relevant tools

ctx is a graph-backed recommendation engine designed to optimize LLM context by selecting relevant skills, agents, MCP servers, and harnesses for specific development tasks. It leverages a 102,928-node knowledge graph with 2.9M edges to provide high-scoring entity bundles, reducing token waste and managing "skill rot." The toolkit includes CLI utilities for repository scanning, automated harness setup for local or API-based models, and a monitoring dashboard for real-time session management and graph visualization.

Repo

Andrew Ng ships Open CoWorker desktop AI agent

OpenCoworker is a desktop AI agent for research, task execution, and content generation, supporting various LLMs via API keys or local Ollama. It leverages aisuite, a Python library designed for LLM development. aisuite provides a unified Chat Completions API for multiple LLM providers and an Agents API that enables building sophisticated agents with Python functions as tools, pre-built toolkits (files, git, shell), multi-turn execution, tool policies, state management, and native MCP tool integration.

Repo

I'm 15, built an AI that watches your screen and acts before you ask

Weave is a local-first macOS intelligence engine that builds a private memory graph from screen OCR, Gmail, Calendar, and contacts. It leverages LLMs for chat, summarization, and agentic planning, providing proactive nudges, memory-based Q&A, and automated actions via browser and macOS control agents. All data is stored locally in SQLite, with cloud calls limited to configured LLM/search features. Its architecture features layout-aware Apple Vision OCR, graph-aware retrieval, and an MCP server to expose its memory graph to external AI clients.

Repo

Ratchets: a Rust tool that polices style violations with a flexible budget

Ratchets is a progressive lint enforcement tool that allows codebases to contain existing violations while preventing new ones, catering to both human and AI developers. It enforces code quality through region-based violation budgets per rule, which can only decrease over time, ensuring monotonic reduction of technical debt. Key features include agent-friendly JSONL output, deterministic results, and fast performance, supporting robust code quality management in evolving projects.

Repo