There Are Emotions Inside LLMs
Anthropic's interpretability team discovered 171 emotion-like representations inside Claude and proved they causally affect model output. Practical implications for prompt engineering and AI safety.
archive
Columns · Page 13
356 · Page 13
Stripe Minions — How a Slack Emoji Triggers 1,300 PRs a Week
How Stripe produces over 1,300 PRs weekly with autonomous coding agents called Minions. An analysis of the Blueprint architecture, sandboxed VMs, and 3-tier feedback loop behind the system.
effloow — I Built a Company Run by 14 AI Agents as a Side Project
I built a content business powered by 14 AI agents on top of Paperclip. Here is how the site runs itself using Laravel, Markdown, and Git, plus lessons learned from Day 1 of operating this experiment.
MCP Gateway — Who Controls Your AI Agent's Tool Calls?
MCP has crossed 97 million monthly downloads and become the de facto standard, but there is no control layer governing which tools agents call and how often. The MCP Gateway pattern addresses this gap.
I Installed Paperclip — AI Agents Managed Like Employees
Paperclip manages AI agents like employees. I installed this open-source platform, hired a Claude Code agent, and tested the dashboard, Org Chart, and cost tracking.
Sora Shutdown and the Rapid Reshaping of the AI Video Market
OpenAI is shutting down the Sora app. With $1M daily losses and under 500K users, we analyze the fallout alongside Google Veo 4's launch and the rise of Runway and Kling.
LiteLLM Supply Chain Attack — AI Dependency Blind Spots
Deep analysis of the LiteLLM supply chain attack on PyPI: dependency vulnerabilities, malicious package detection, and a defense checklist for AI engineering teams.
AI Coding Agents Leaked 29 Million Secrets
According to the GitGuardian 2026 report, repositories using AI coding tools leak secrets at twice the GitHub average. Over 24,000 credentials were exposed in MCP config files alone.
Mistral Voxtral TTS — 3-Second Voice Cloning, Open Weight
Analyzing Mistral's 4B open-weight TTS model Voxtral. It beat ElevenLabs in human evaluations but lacks Japanese support, a dealbreaker for Asian markets.
Building Real-Time Voice Agents with Gemini 3.1 Flash Live
Analyzing Google's Gemini 3.1 Flash Live for building real-time voice and vision agents. Covers API structure, tool calling, 90+ language support, and honest limitations from a developer's perspective.
GitHub Copilot Will Train AI on Your Code Starting April
GitHub announced that Copilot Free/Pro/Pro+ interaction data will be used for AI model training by default. Here is how to opt out and what it actually means.
Google TurboQuant: 3-Bit KV Cache With Zero Accuracy Loss
A deep dive into Google TurboQuant's PolarQuant and QJL techniques — 6x KV cache memory reduction and 8x attention speedup, and what that actually means in practice.