---
title: "In-Context Learning: The Front Door for Brazilian AI Engineering"
description: "I keep coming back to the GPT-3 paper. What stays with me isn't the 175-billion-parameter scale, but in-context learning — the idea that AI capability is something you compose, with context and care, rather than something you own. That's where the engineering we do in Brazil lives."
author: "Anderson Henrique"
date: "2026-07-13T22:25:21.302662Z"
updated: "2026-07-13T22:25:21.302662Z"
category: "philosophy"
tags: ["AI","In-Context Learning","GPT-3","Software Engineering","AI in Brazil","Multi-Agent Systems","NTLabs"]
canonical: "https://www.ntlabs.dev/en/blog/in-context-learning-engenharia-brasileira-ia"
locale: "en"
---

It's been a little over five years since I first read "Language Models are Few-Shot Learners" (Brown et al., NeurIPS 2020) — the paper that introduced GPT-3. I come back to it every now and then, and each rereading shifts, just slightly, how I think about AI engineering.

The first reading, the one everyone does, is about scale: 175 billion parameters, and capabilities that seem to appear on their own as the model grows. It's an important part of the story — and also the part a Brazilian lab is unlikely to reproduce, because it demands computing power that belongs to very few hands in the world.

But what keeps bringing me back isn't the scale. It's a quieter idea living inside the paper: *in-context learning*. The discovery that you can teach the model a task inside the prompt itself, with a few examples, without training anything new. It sounds like a technical detail, and it's really a change in nature: AI capability stopped being something you **own** — by force of compute — and became something you **compose**, with care, orchestration, and context.

That shift is, to me, a door that opens precisely for those far from the training frontier. And it's exactly the bet at Neural Thinkers Lab: to democratize access to AI in Brazil not by building the biggest model, but by building the most useful, transparent, and locally-rooted systems on top of what already exists.

That's where the engineering we do here finds its ground. We don't compete on raw compute; we build systems that explain themselves, that reason about Brazilian problems, grounded in our own data and auditable by the people who depend on them. We call it Organic AI: an intelligence that grows with use, integrates into its cultural context, speaks real Portuguese, and augments people rather than replacing them. In Cidadão.AI, for instance, the value doesn't come from the size of the model underneath, but from an agent architecture that investigates public contracts and can show, in plain language, why it raised a flag.

The paper itself has the honesty to name its limits — few-shot still fails on many tasks, and learning from the entire web carries the entire web's bias. Far from a footnote, that reads almost like a map: it shows that the frontier that's ours to claim isn't the biggest model, but applied, explainable AI, rooted in the place where it will actually live.

As we like to put it here: between calculation and dream, intelligence resides. GPT-3 handed us a generous share of the calculation. What we do with it — for whom, with what care, and to what purpose — is still, thankfully, a choice of our own.

*Paper: Brown et al., "Language Models are Few-Shot Learners", NeurIPS 2020.*