Automatic transcript made with Whisper. Expect mistakes, mostly in names and technical terms.
Chapters
= how much the chapter is worth if you only want the ideas (0 to 3 gems). Click a card to jump.
Summary
A Madrid.rb roundtable recorded on the stage at El Singular and streamed on Twitch. Josep, who runs the group, wanted to talk about what people really use LLMs for today, leaving trends and science fiction aside. With him are Rafa, a staff engineer doing a master's degree in AI, German Del Zotto, a senior software engineer at Workato, and Israel, an entrepreneur who works on running models locally 11:52.
Josep opens with the context: a Google Slides AI cover with a typo in it, CEOs who think their own executives aren't ready, and a CEO who built one of the biggest Rails apps and made AI use mandatory at his company 9:03. Then the round. Rafa uses it instead of Google, to learn with the Socratic method, and spent a month without writing a single line of code himself 14:01. For German, Cursor changed things six or seven months before, because it was the first tool that seemed to understand the whole context, and he likes it for exploring ideas cheaply 15:26. Israel wants to run as much as possible locally 16:54.
At Rafa's company about 70 % of people picked Cursor and the rest GitHub Copilot, and he thinks the tools now differ in UX and in the context they get 19:21. German's employer pays for whatever he needs, with strict rules about data leaving the company. He pays his personal Cursor account himself because speed changes how he plans his day, and his local models can't match it anymore 21:31. At Workato the data question is mostly regulatory, and as an orchestration product they add AI on top of connections they already had 26:00.
Rafa compares the moment with installing every Vim plugin or declaring Emacs bankruptcy for the third time. Nobody keeps up, and no company has a moat yet 31:22. Israel says LLMs have raised his capability but not yet his productivity, and the panel mostly agrees. Rafa still finds them a 10× accelerator for learning 34:20.
The part about teams is the most concrete. German's side project doesn't need the graphic designer anymore. He ships pages good enough to test with users, on his own schedule, without waiting for anybody 39:06. Rafa's designers started sending animated mockups built with bolt.new, and that raises the bar for everyone 43:59. German expects more generalists who cover 70 % of a field with the LLM as a crutch, and he's worried about juniors, who already learn more isolated since the pandemic 45:47. Rafa partly disagrees. For him it's a bad time to be a junior in the market and a good time to be one who wants to learn 49:23.
A Twitch viewer asks about vibe coding. Rafa watched a self-described vibe coder show his setup and found cursor rules, templates and patterns, which is programming 57:05. German compares the moment with PCs arriving in offices in the early nineties, when anybody with a spreadsheet, Visual Basic or dBase could automate their own work, something we lost in the last decade 59:20. He's happy vibe coding throwaway tools, like a word cloud for his own admin panel 1:01:29.
Someone in the audience asks what a 12-year-old should learn. The panel answers with curiosity, and with stories about an Amstrad CPC sold next to the washing machines in a Gijón store and a ZX Spectrum 1:06:54. The last round is about uses beyond code. At Workato, an agent with access to the org chart now decides on requests to enable USB storage and calls the right APIs 1:18:00.