Code in, diagram out

The household model rankings turn over in one evening: “realmente ChatGPT ya no es útil para nada” ▸ 27:09, hyperbole with a payload, because Gemini 2.5 Flash Image just landed in Google AI Studio and its images sit “al nivel de ChatGPT” ▸ 27:22.

The experiment that earns the entry isn’t a pretty galaxy (though he makes one; it becomes this video’s thumbnail). It’s this: he pastes the source code of his losses-as-reward research, the RL agent that moves a window to recognize digits, and asks for a futuristic architecture diagram ▸ 28:24. Three attempts later: diagrams “casi que presentables 100%”, dimensions labeled, state-action-reward flow recognizable, lettering imperfect but Canva-fixable. The use case crystallizes: “poder visualizar una arquitectura o una idea antes de hacerla, o cuando ya está hecha y uno la quiere entender más rápido” ▸ 30:02. Code-to-diagram wasn’t possible at this quality with anything else, because the differentiator isn’t the image model, it’s the context window that swallows the whole codebase first.

el diagrama sale del código, no del prompt →

The second half of the teaching is economic, and it extends entry 69-3’s free-credit weather report: why is all this free? Because AI Studio “está hecho para que los desarrolladores prueben”, the product (Gemini app) carries the limits; the playground doesn’t ▸ 35:18. Live voice included, he demos Stream flipping through Italian and Chinese mid-conversation, and eyes it for language practice. “Hay que aprovechar”: the frontier of free intelligence lives in the tools vendors build for developers, until abuse arrives and the meter starts. Zero-budget builders should always know where that frontier currently sits, because it moves…