Blog posts in English
I write my posts in Portuguese first, for a Brazilian audience, and then translate them. These are all of them in English, newest first. The original Portuguese blog has the same posts.

PMs up, devs stagnant and design down: the 2026 tech job market
A Lenny’s Newsletter report shows PM openings at their highest in almost four years, flat engineering, falling design and AI becoming a basic requirement in the tech job market.
![Cover of the post [TUTORIAL] How to train your own decision model (Jev) using an LLM](/images/capas/decision-models-en_hu_e7ec8023bfb85fb1.webp)
[TUTORIAL] How to train your own decision model (Jev) using an LLM
What a decision model is, how most companies built theirs and how to train one: the logits trick, the data, the training, the evaluation with its rules fixed before each round, and the server, with CatDecide, BSC's Catalan model, as the example, plus a step-by-step guide with code.

Revolut cuts AI costs by 90% by swapping Claude for a local Chinese model
Revolut started running GLM-5.3 locally for 2,000 engineers and says it cut costs on AI-based coding tools by 90%.
![Cover of the post [TUTORIAL] How to fine-tune and teach a new language to a voice model (TTS)](/images/capas/vermut-tts-en_hu_acc2499a6b0064a2.webp)
[TUTORIAL] How to fine-tune and teach a new language to a voice model (TTS)
How I taught Catalan to an open voice model, vermut-tts, with 1 hour of GPU: the preliminary test, the data and the edges, the conversation with laughter, the voice that comes from a 3-to-10-second recording, DPO and how to measure without fooling yourself, with code and audio to listen to.

Anthropic says you're building AI agents the WRONG way
Anthropic compared enterprise architectures and concluded that a single agent with tools, skills and progressive disclosure often beats multi-agent systems.
![Cover of the post [TUTORIAL] Pruning and distillation of LLMs: how a 40B model teaches a 20B one, token by token](/images/capas/pruning-distillation-en_hu_770244f5e8ad99a5.webp)
[TUTORIAL] Pruning and distillation of LLMs: how a 40B model teaches a 20B one, token by token
How to prune an LLM in half and fix the damage with distillation (KD), with a real training step of ALIA-20B in Portuguese and Spanish and a step-by-step guide with code.

Anthropic Wants to Go Public at a $1.5 Trillion Valuation, Even With $80 Billion in Losses
Anthropic filed its S-1 to go public on the Nasdaq, revealing explosive growth, billion-dollar losses, and a huge bet on AI infrastructure.

A Dangerous New Technology
A reflection on memory, knowledge, and the risks of a new technology that promised to make people smarter.

Muse Wonder: Meta at the AI frontier
After the big Llama 4 fiasco, Meta FINALLY got back in the fight with cutting-edge AIs, and this time with numbers that make sense on paper.

OpenAI Publishes 722 Math Solutions Generated by a New Secret AI
OpenAI published 722 solutions to classic math problems, with proofs formalized in Lean and the math community impressed.