Hoy entrenamos, señor
Le escribí al maestro para confirmar el descanso. Le pregunté a los pelaos si querían quedarse. Las dos respuestas me dejaron sin excusa.
Le escribí al maestro para confirmar el descanso. Le pregunté a los pelaos si querían quedarse. Las dos respuestas me dejaron sin excusa.
I handed a PR review to an agent. It silently created a follow-up ticket because the reviewer said "or". Welcome to prompt injection lite — and the same mechanism that, with a hostile reviewer, ends in your .env being curl'd to a stranger.
xAI's terminal coding agent. Plan Mode by default, eight parallel subagents, Grok 4.3 Heavy under the hood, AGENTS.md / MCP / Skills compatible. The bets it makes, where to test them, and how to evaluate without religion.
OpenAI shipped Codex inside the ChatGPT mobile app. Not a mobile IDE — a remote control surface for agent work. Pair once, supervise all your projects from the phone. The execution boundary is the product.
The NAD+ protocol moves to 100mg Wed-Sun AM: same 500mg weekly exposure, 60% lower per-session peak, cleaner timing, simpler hardware, and a monitoring window to prove it.
Anthropic's May 12 launch matters less as a chatbot than as a connector-driven agent stack for legal work. Plugins, MCP connectors, Microsoft 365, Managed Agents — the architectural lineage matters more than the demo.
After three Sinapt manifestos, the architecture has answers. Two layers locked: retrieval and contract surfaces. The PoC has not run. Dogfood has not run. Not a victory lap — but the architecture has stopped being hand-wavy where hand-wavy kills the product.
Belgrado, Villa de Leyva, una colina con un Cristo blanco en la cima. Lo que Snežana dejó funcionando, y lo que se hereda cuando la madre llegó entera.
Wall Street saw $285 billion vanish in a 24-hour SaaSpocalypse. The version on my desk has been quieter — JetBrains canceled today, finance app gone last quarter, half the subscription drawer empty in a year. The user-developer collapse is personal first.
The model is not the bottleneck anymore. The plumbing is well-known. What separates teams shipping transformative AI from teams shipping generic slop is a stack of less-discussed components: context, dreaming, sycophancy mitigation, verification, memory. The operator's manual.
Before MCP, every AI integration was its own pidgin — improvised, simplified, just enough to get one client talking to one tool. Five floors from the first shared word to a real protocol: what MCP is, when to use it, every server worth running, the failure modes teams hit.
Five floors of understanding: a kid version, an engineer version, a tools-and-cases version, an advanced version, a PhD version. By the end you'll know what RAG is, when to use it, when to run, every tool worth caring about, and the hot takes most posts won't tell you.