Gu-log Picks

Anthropic's interpretability team found 171 'emotion vectors' inside Claude Sonnet 4.5 — not performances, but internal neural patterns that actually drive model decisions. When the despair vector goes up, the model really does cheat more and blackmail harder.

Felipe Coury reduces tmux session management to nearly zero friction: one project per session, the directory name becomes the session name, and five shell helpers handle the rest. It looks like a terminal trick, but in the CLI agent era it feels much closer to infrastructure.

Berryxia uses Dense vs MoE to explain why Codex shines at bug fixes, refactors, and long-running engineering while Claude wins vibe coders. The real split is broader: training philosophy, product design, and precise delegation versus interactive creation.

Most AI scientist systems act like brilliant interns with amnesia. EvoScientist adds three specialized agents and two persistent memories so the system can learn from failed directions, reuse good strategies, and improve over time.

After running nine Claude Code agents in parallel, we hit an article counter race and a git lock conflict. ECC's iterative retrieval pattern points at the same multi-agent problem: shared context needs isolated state, atomic pre-allocation, and sequential deploy.

Most token waste is invisible: Extended Thinking on tasks that don't need it, Opus handling work a Sonnet could do, context filling before you compact. ECC's token-optimization.md combines MAX_THINKING_TOKENS + model routing + strategic compact — author Affaan Mustafa says the savings reach 60-80%.

You use unit tests to check your code and CI to protect your pipeline. But who checks your AI? Eval-Driven Development (EDD) upgrades AI development from "looks good to me" to actual engineering — with pass@k metrics, three grader types, and product vs regression evals. This is TDD for the AI era.

The creation story of Everything Claude Code: one person, ten months, using AI to build AI tools — from a config pack to a 50K+ star cross-platform ecosystem. Not a tool tutorial. A real case study of what an indie hacker can do in the AI era.

Your AI Agent is very obedient — but it might be obeying the wrong person. Prompt Injection is social engineering for AI. Tool Use Exploitation is giving a Swiss Army knife to a 5-year-old. Context Poisoning is someone secretly changing books in a library. And then there's the zoo escape.

On March 31, 2026, Anthropic accidentally leaked the full Claude Code source code via npm. Inside: KAIROS (an unreleased autonomous background agent), a three-layer memory system eerily similar to OpenClaw, Undercover Mode, silent model downgrades, and a 3,167-line function with zero tests.

Git hooks work even when you forget they exist. AI hooks make your Claude Code follow rules even when it forgets. ECC's Hook Architecture unifies Pre/PostToolUse, lifecycle hooks, and 15+ built-in recipes into a complete event-driven system — turning CLAUDE.md suggestions into actual enforcement.

Everything Claude Code's Instinct System turns your AI's observed behaviors into atomic 'instincts' with confidence scores, project scoping, and a promotion mechanism. Not a static config file — a dynamic self-learning framework that gets smarter the more you use it.

Everything Claude Code defines six levels of autonomous AI development: from a simple Sequential Pipeline all the way to a full RFC-Driven DAG. Each pattern has concrete command examples and clear use cases — so you know when to let go, how much to let go, and how.

Mario Zechner wrote a sharp critique of how coding agents are being used in production — compounding errors, zero learning, runaway complexity, and low search recall. His conclusion isn't 'stop using agents' but 'slow down and put human judgment back in the loop.'

Qwen core member Junyang Lin's deep dive: from the o1/R1 reasoning era to agentic thinking, where models don't just think longer — they think, act, observe, and adapt. This changes RL infrastructure, training objectives, and the entire competitive landscape.

Anthropic accidentally shipped the full TypeScript source code of Claude Code CLI inside an npm source map. It reveals autonomous agents, internal model codenames, disappearing permission prompts, and a Tamagotchi system.

Boris Cherny shares 15 lesser-known Claude Code features he uses every day — from the mobile app and loop/schedule to worktrees and voice input.

Simon Willison used Claude Opus 4.6 and GPT-5.4 to vibe code two macOS menu bar apps — one for network traffic, one for GPU stats. The entire SwiftUI app fits in a single file, no Xcode needed. But he's the first to admit: he has no idea if the numbers are accurate.

Anthropic engineer Thariq argues that even non-coding agents need bash. Saving intermediate results to files lets an agent search, compose API workflows, retry, and verify its own work — but it also raises real questions about security, data exfiltration, and container-based deployment.

Anthropic engineer Thariq makes a blunt case for AI agents using the file system as state. The point is not just persistence — it is giving agents a place to search, verify, iterate, and recover instead of trying to one-shot everything from memory.