Interfaces that assemble themselves, a $60 billion financial operation and a security breach with a lesson attached. Episode 7 of Tokenizados touches three fronts that define the industry’s moment.
The era of generative UI and GPT Image
Jorge opens with a new paradigm: interfaces that aren’t static but built in real time with components from a design system based on the user’s interaction. Along the same visual lines, the episode analyzes GPT Image, OpenAI’s model that finally nails text and visual reasoning. Two practical applications stand out: generating synthetic data to train agents that control computers, and a surprising workflow — designing via screenshots and letting the model generate the final code.
GPT 5.5, Chinese pressure and Anthropic’s postmortem
OpenAI launched GPT 5.5 seeking to reaffirm its lead in coding, but the competition presses from China: models like DeepSeek v4 and Kimi offer comparable performance up to 40 times cheaper. Meanwhile, Anthropic published an unusually honest postmortem: its models underperformed due to a lack of compute and changes to internal prompting to save costs — a reminder that perceived quality is also a business variable.
SpaceX buys (an option on) Cursor
The week’s financial bombshell: SpaceX has acquired a call option on Cursor valued at $60 billion. The logic: secure talent and put Elon Musk’s “Colossus” data center’s immense compute capacity to use to accelerate internal software development. In parallel, Google reinforces its position by investing $40 billion in Anthropic. The pattern repeats: only the giants with their own infrastructure seem to have the future secured.
The Vercel hack: the token lesson
The Vercel breach wasn’t a direct assault: it came in through a third-party provider (Context AI) to which an employee granted excessive permissions via a Google token. The result: access to unencrypted environment variables of many companies. The lessons are textbook and yet ignored daily — mark credentials as sensitive, rotate API keys regularly, and apply least privilege to AI integrations too.
Code isn’t cheap (even if it looks like it)
The closing is a defense of fundamentals: AI lets you generate code massively and almost for free, but without engineering behind it that produces “exponential garbage.” Today it’s more important than ever to be strict with interfaces, tests and architecture. Machine-generated spaghetti code is still spaghetti code — it just arrives much faster.
Thanks for listening to Tokenizados one more week. Subscribe so you don’t miss a thing.