In the second episode of Tokenizados, Jorge and Antonio dive straight into the tools redefining AI-assisted development — and along the way they explain the podcast’s name: tokenizados (tokenized), like the basic elements in the vector spaces that models think in.
Composer 2: speed or quality?
Cursor has launched Composer 2, a model positioned as the fastest on the market and boasting benchmarks against Claude Opus and Sonnet. The underlying question is the eternal trade-off: would you rather have an ultra-fast coding assistant that iterates with you in seconds, or a slower one with better judgment? The episode’s answer: speed matters more than it seems, because it enables a 10x iteration capacity in real development workflows.
Looking over the artificialanalysis.ai leaderboard, GPT still leads, followed by Gemini and the Claude models, but the quiet headline is the rise of the Chinese models (Kimi 2.5, GLM), climbing the rankings at full speed.
Garry Tan’s G-Stack: spec-driven development
The central section analyzes Garry Tan’s (Y Combinator) G-Stack: a spec-driven development methodology where roles like CEO, Engineer and Designer are modeled as skills. The Office Hours skill, for example, refines a business idea by challenging it with questions before writing a single line of code. The result reported by those who use it: 10x-20x improvements in execution quality.
The practical lesson is clear: the more context and structure you give the model, the better. Basic mode — opening Cursor and typing what you want — falls short compared to workflows with defined specs, roles and processes.
Positive sum: don’t optimize, aim higher
Drawing on an essay by Garry Tan, the episode lays out two ways of adopting AI: the conservative one (doing the same as always, cheaper) and the ambitious one (attempting what was previously impossible). The history of technology favors the second: positive-sum approaches grow the economy instead of dividing a fixed pie. As Jorge and Antonio sum it up: “you shouldn’t try to do the same thing with AI, but be more ambitious about what you want to achieve.”
Agents, security and NemoClaw
The episode also covers agent operating systems and MCP, the difference between personal and impersonated agent tokens, and the security considerations that skill marketplaces like Skills.sh bring: prompt injection is a real risk when you install third-party skills. NVIDIA, for its part, makes a move in the enterprise space with Nemo Claw, backed by Jensen Huang himself.
OpenAI doubles its headcount
A closing headline: OpenAI plans to go from 4,500 to 8,000 employees. Growth strategy or a contradiction with the promise of small teams supercharged by AI? The debate is open: maybe the future isn’t choosing between tiny teams or large organizations, but large organizations where each person performs like a team.
Artificial intelligence, they say, “is a superpower available to everyone — and if you’re not starting to use it, you’d better get on it.” Subscribe and don’t miss next week.