English translation of the Portuguese original. The audio is in Portuguese.
In the /root episode with Guga Stocco, published in 2023, I recalled the early days of the internet in the 1990s. Websites were slow, we connected through modems and had to build highly optimized pages. Back then, we used to say we were crawling, that everything was just beginning. Looking at artificial intelligence and the topics of the moment, I asked Guga:
Are we still crawling, or have we reached a stage where we are crawling faster, perhaps?
Guga’s answer (summary)
Guga answered that it depends on the topic. In many, we were still crawling; in some, things were getting serious. The metaverse, which had been a trend, was an example of the first group: it still required putting on a big headset just to see something.
Artificial intelligence was in the second group. According to him, neural networks, large networks and model training had existed for a long time. What was missing was the processing power to handle such a large volume of data, and the availability of that data. When those pieces came together, there was a breakthrough, and the image he used was going from crawling straight to jogging, skipping the walking stage.
That leap, he said, has consequences, because not everyone can go from crawling to running at the same time. Those who can collect data and train models quickly gain a very large competitive advantage. He also pointed to structural limits in Brazil and gave a striking example: in his assessment, if ChatGPT had been created here, Brazil’s data protection law (LGPD) would not have allowed it. For Guga, in this move from crawling to running, not everything keeps the same pace.
The remark about the LGPD is his view in that 2023 conversation, not a legal analysis.
In the video of this question, the 2023 remark cites Brazil’s Real Digital (later called Drex) among the topics of the moment; the project later left the Central Bank’s priorities and should not be read as an ongoing agenda.
A reflection for today
In 2023, Guga saw the competitive advantage in whoever could bring together processing power, data and model training quickly. That reading still holds for those who build models. For those who build companies on top of them, the competition gained another layer once models started acting inside systems that already exist: code repositories, tests, review queues, conversations with the team. Access to a good model has become an increasingly common starting point. What now sets one operation apart from another is the ability to hand it an entire process, end to end, with the necessary context, clear decision criteria and evidence that the result came out as it should.
Karri Saarinen, CEO of Linear, described on November 6, 2026 how the company organizes that work with a cloud coding agent. By his account, most of their code (5 million lines last quarter, a self-reported figure) is now written with that agent. Fixes come straight in Slack (“@linear fix this”); a Loop investigates with Sentry and Datadog and writes the patch; designers, PMs and non-engineers contribute code; every pull request arrives with screenshots, video and a preview build; and every project starts with 10 to 500 customer notes. His recommendations close on letting agents build, making contribution easier and raising the ambition of what teams attempt.
What strikes me in this design is where the value ended up. The agents execute, and whoever designs the operation defines who can trigger the work, what evidence comes with each delivery and which criteria separate what gets approved from what goes back for adjustment. The request surface, the automated investigation and the mandatory proof on the PR turn volume into reviewable throughput. It is the CEO’s account of his own company’s operation, and the post brings the line count as a self-reported metric, without external productivity multipliers, so it is worth studying as a design. For anyone leading a company, the work that remains is to organize processes so that an agent can carry them out from start to finish: write down the decision criteria, define what counts as proof that it worked and choose where human approval comes in. In 2023, the differentiator was processing power, data and training. Today, more and more, it lies in the operational organization that lets AI work inside the company.
The full episode is available on RSS.com.
Credits: /root, O Podstech da Celcoin. Hosts: Adriano Meirinho and Thiago Zaninotti. Guest: Guga Stocco. Episode 6: Tecnologia, pra quê?, published in 2023. Question at 10:52 to 11:45; summarized answer at 11:45 to 13:52.