Anthropic published its first update on Project Glasswing, the interpretability research effort it announced last year. The update doesn’t introduce a new model or product. It’s a progress report on mapping how Claude’s internals work, specifically how the model represents and processes information.

The team focused on what they call “features,” the internal activation patterns that correspond to concepts. They’ve identified features for things like code syntax, emotional sentiment, and specific knowledge domains. The interesting bit is scale: they’re now analyzing features across multiple layers simultaneously, which matters because concepts don’t live in just one part of the network.

The practical application is safety. If you can identify which features fire when a model generates harmful content, you can potentially intervene before the output happens. Anthropic says they’ve used this to reduce certain types of undesirable outputs in internal testing, though they don’t specify success rates or which outputs.

This isn’t a tool you can use yet. It’s research infrastructure. But if you’re building systems that need explainability for compliance or safety reasons, this is the direction the field is heading. The full update is on Anthropic’s research blog with technical details and examples.

Diffusion models for text generation

Nvidia released Nemotron-Labs Diffusion Language Models, and the pitch is speed. Traditional autoregressive models generate one token at a time. Diffusion models generate entire sequences in parallel by iteratively refining noise into text.

The benchmark numbers: 16x faster than comparable autoregressive models for similar quality output. That’s measured on standard generation tasks, not cherry-picked examples. The tradeoff is quality at the highest end. For tasks where GPT-4 level performance isn’t required, this matters.

The models are on Hugging Face. You can run them now. Nvidia published three sizes: 340M, 1.3B, and 7B parameters. The 7B model hits roughly LLaMA-2 7B quality at a fraction of the inference cost.

Who should care: anyone running high-volume generation workloads where response time matters more than perfect quality. Customer service responses, content summarization, data extraction. The cost savings compound quickly at scale.

Who can skip it: if you need reasoning or complex multi-turn conversations, stick with frontier models. Diffusion models aren’t there yet for that work.

Kanbots: agents on every card

Someone built a Kanban board where every card runs autonomous agents. It’s called Kanbots, it’s open source, and it’s a desktop app.

The concept: you create cards for tasks, and agents work on them in parallel. Each agent has access to tools like web search, code execution, and file operations. You can configure which models to use per board or per card. It supports OpenAI, Anthropic, and local models through Ollama.

This isn’t a web app. It’s Electron, runs locally, stores everything on disk. That matters if you’re working with proprietary code or data you don’t want leaving your machine.

The UI shows agent activity in real time. You can see which cards agents are working on, what tools they’re calling, and pause or modify their work mid-stream. It’s at kanbots.dev with full source on GitHub.

Early days, rough edges. The agent orchestration is basic compared to dedicated frameworks. But for teams that think in Kanban and want to experiment with parallel agent workflows without building infrastructure, it’s a starting point.

Quick hits

GitHub got named a Leader in Gartner’s Magic Quadrant for Enterprise AI Coding Agents for the third consecutive year. Copilot now supports multiple models, workspace indexing, and pull request reviews. If you’re already in the GitHub ecosystem, the integration depth is the selling point.

Google Search broke its own interface. Searching for “disregard” now returns an error or no results, apparently related to how the AI overview feature processes the term. It’s the kind of bug that makes you wonder what other common words trigger edge cases.

The memory shortage is real and getting worse. David Oks explains why: only three major manufacturers, fixed wafer capacity, and AI training eating more DRAM than projected. Consumer electronics that use memory are going to cost more. Plan accordingly.