I used Claude Code to discover connections between 100 books

Trails is a knowledge synthesis project by Pieter Maes that utilizes Claude Code to automate syntopic reading across a diverse library of books. The system identifies latent thematic clusters—such as "Proxy Traps," "Multiple Discovery," and "Conway’s Law"—to map non-obvious conceptual links across different texts. This implementation demonstrates the utility of LLMs for automated cross-document analysis and semantic discovery.

Read more

AI is a business model stress test

AI is stress-testing open-source business models by commoditizing "specifications" like documentation and UI components, as evidenced by Tailwind Labs' recent 75% workforce reduction. While LLMs can generate code and bypass traditional marketing funnels, they cannot automate operational requirements like 99.95% uptime, security, and CI/CD. Consequently, sustainable value is shifting from static assets to managed operations and hosting services where the open-source project serves as a conduit for the product.

Read more

LLMs have burned Billions but couldn't build another Tailwind

Tailwind CSS's recent 75% workforce reduction highlights a disconnect between massive LLM investment and actual software innovation. Despite billions spent on AI coding agents and token generation, no small AI-augmented team has replicated the sophisticated, lean, and foundational impact of human-crafted frameworks like Tailwind. The industry risks losing high-quality, efficient tools while failing to produce equivalent breakthroughs through current LLM-driven development paradigms.

Read more

Datadog, thank you for blocking us

After Datadog revoked access, Deductive migrated to an open-source Grafana stack in 48 hours, demonstrating that AI-assisted development tools like Cursor and Claude Code have effectively eliminated vendor lock-in by collapsing the cost of code changes. By leveraging OpenTelemetry and MCP, the team integrated live telemetry directly into the development loop, allowing AI agents to validate instrumentation and code behavior in real-time. This shift signals a move toward AI-native observability where agents, rather than human-centric dashboards, become the primary consumers of system signals.

Read more

Yuanzai World – LLM RPGs with branching world-lines

Yuanzai World is an AI-driven interactive simulation platform developed by Hangzhou Yuanzai Artificial Intelligence Application Software Co., Ltd. It enables users to explore generative environments and manipulate narrative worldlines through time-space anchoring features. The application is available on iOS and Android, focusing on dynamic, user-influenced world-building and exploration.

Read more

One pixel attack for fooling deep neural networks

Researchers developed a black-box adversarial attack using Differential Evolution (DE) that compromises DNNs by modifying only a single pixel. The method achieved success rates of 67.97% on CIFAR-10 and 16.04% on ImageNet, demonstrating that models are highly vulnerable to extreme low-dimensional perturbations. This study highlights the effectiveness of evolutionary computation for generating low-cost attacks to evaluate model robustness.

ArXiv

Extracting books from production language models (2026)

This study investigates the feasibility of extracting copyrighted training data from production LLMs, a critical aspect of copyright debates. Researchers employed a two-phase procedure, sometimes using BoN jailbreaks, to extract text from Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3. While Gemini 2.5 Pro and Grok 3 allowed initial extraction without jailbreaking, a jailbroken Claude 3.7 Sonnet could output near-verbatim entire books. This work demonstrates that, despite model and system-level safeguards, the extraction of in-copyright training data remains a significant risk for production LLMs.

ArXiv

Extracting books from production language models

This research investigates the feasibility of extracting copyrighted training data from production LLMs, a central issue in copyright discussions. Using a two-phase procedure involving initial probes and iterative continuation, sometimes with BoN jailbreaks, the study successfully extracted varying amounts of text from Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3. Notably, some models like jailbroken Claude 3.7 Sonnet output near-verbatim entire books, highlighting that copyrighted data extraction remains a risk for production LLMs despite implemented safeguards.

ArXiv

Synthetic Biology Meets Neuromorphic Computing

This research proposes a hybrid artificial olfaction system that integrates synthetic biology with neuromorphic electronics through a co-design approach. The architecture leverages synthetic sensory neurons and a bio-semiconductor interface to enable event-based encoding and computing via spiking networks. Validated by simulation, this pipeline aims to achieve energy-efficient, ultra-sensitive odor detection for environmental and medical diagnostics.

ArXiv

Deep sequence models tend to memorize geometrically; it is unclear why

Deep sequence models utilize "geometric memory" to synthesize global relationships between entities, enabling the reduction of multi-step reasoning into single-step navigation. This phenomenon, driven by a spectral bias analogous to Node2Vec, persists even when more complex than associative lookups. These insights reveal potential for enhancing Transformer memory and offer new perspectives on knowledge acquisition and unlearning.

ArXiv

Turn any topic into a 3Blue1Brown-style video

Topic2Manim is an automated pipeline that converts text prompts into educational videos by leveraging LLMs and the Manim animation engine. It uses models like GPT and Claude to generate scripts and corresponding mathematical visualizations, integrating TTS for narration and automatic scene concatenation. The tool supports multi-language detection and provides a streamlined workflow for producing concise, professional-grade technical animations.

Repo

Persistent Memory for Claude Code (MCP)

A-MEM is a self-evolving memory system for coding agents that organizes knowledge into a Zettelkasten-style graph using ChromaDB. It enhances standard RAG by using LLMs to automatically extract context, establish dynamic relationships, and evolve memory connections over time. Integrated via MCP, the system enables both semantic and structural search through graph traversal, allowing agents to maintain persistent, interconnected project knowledge.

Repo

Revibing nanochat's inference model in C++ with ggml

nanochagg.ml is a C++ inference engine for nanochat models built on GGML, serving as a drop-in replacement for the original GPT and KVCache classes. It supports CPU and GPU (Metal) backends and features automated PyTorch-to-GGUF conversion. While currently limited to float32, it provides a lightweight framework for experimenting with nanochat architectures in a C++ environment.

Repo

InfiniteGPU, An open-source AI compute network,now supporting training

InfiniteGPU is a distributed marketplace for AI compute that enables offloading inference and training tasks to a network of providers. The platform utilizes a WinUI 3 desktop client to execute ONNX models across NPU, GPU, and CPU hardware, orchestrated via an ASP.NET Core 10 backend and React 19 frontend. Key features include real-time task tracking through SignalR, automated Stripe payments, and a CQRS-based architecture for scalable resource sharing.

Repo

Build your own Atlas/Comet AI-browser (open source)

Chromium is an open-source browser project focused on building a safer, faster, and more stable web experience. Developers are directed to specific instructions for source code acquisition and documentation, with guidance on a product-centric directory structure for new top-level directories. Bugs can be reported via crbug.com.

Repo