Mario Gabriele

“By leveraging frontier models, we've been able to build an extensive signal-gathering apparatus that would not have been possible to conduct without a very large team, spanning news, industry outlets, social media, academic research, government and regulatory filings, code repositories, model-adoption aggregators, fundraising data, prediction markets, website monitoring, open-source intelligence sources, and talent flows.”

“The system hunts ceaselessly, bringing up a furious exhaust of data, noise, and signal, all in one. It is up to the human to think, judge, analyze, and write.”

“I've been reading internal versions of this as we tuned the format over the past couple of months, and have found myself looking forward to it every week.”

“Reading Henrik Karlsson's excellent piece, 'Childhoods of exceptional people,' in which the author outlines the patterns he observed by studying the early years of remarkable writers, mathematicians, philosophers, and composers... prompted me to think more deeply about the value of studying childhoods specifically.”

“There is no way I would have been able to take on a project of this scope without the latest models. If I monastically devoted every waking second to reading founder biographies, I calculate it would have taken me nearly a full year to read the full library, and several more to encode them properly.”

“If you tell an agent to read and then write, it optimizes for the output... Splitting those jobs up helped. It forced the reading agent to pay close attention and transcribe its findings into notes, and the writing agent to focus on crafting a great profile. Eventually, we realized you could have Sonnet conduct the reading work without loss of detail and save Opus for writing.”

“None of this would be useful without a human at the heart of it, not just in structuring the machine and stoking its fires, but in actually doing the work of *studying* these people... That is not something Claude, which tends towards the glib and simplistic at times, excels at.”

“Once I felt we had a strong process that reliably yielded good results, I started to rely on Modal (a service to run your code across multiple machines in the cloud) for parts of the process. This reduced the time on our batches considerably, likely shifting our timeline from weeks to days on some tasks.”