Austin Morrissey

Language models offer Americans an overlooked benefit: direct access to information beyond the Anglosphere. Despite the internet’s global reach, we’ve confined ourselves to English-language sources, relying on secondhand reports we treat as authentic, but which are better understood as merely uncontested. Last year, I wrote about two high-ranking Russian agents arrested in Boston after years undercover. I used their story as an analogy for Anthropic’s research on sleeper agents embedded within large language models. While researching that post, I discovered the couple had written a book years after their arrest: Нетворкинг по-русски (roughly, “Networking for Spies”). It had no English translation. But that no longer matters. A year ago, I found their writing compelling enough that I screenshot each page of the Cyrillic PDF and translated it a few pages at a time. Compared to what is possible, it was slow, but compared to what was possible, it felt as if I had boarded a high-speed train to the future. I made mental note to return to the topic in writing, and then for months it left my brain. Though today, during lunch, the thought again arose: what was that text? Consider, at an experiential level, this sequence of events. I could not recall the publication. In one key press, dictation software triggers on my computer, and I ask it to find the title of “that Russian book on networking for spies.” The search is performed for me, and the title is returned. I input that Cyrillic title into a giant open access database for science and literature, Anna’s Archive, and in seconds I receive the document. Then, I describe in natural language the task at hand to an AI, who writes and executes code in my terminal. Before I have finished my lunch, I have a translated book and reusable code, but have expended no effort, performed nearly no action, and exercised only curiosity. Can you grasp how strange that is? What follows is a brief aside on how the code works, and excerpt