The False Promise of Safe AI - The Atlantic

Oh my goodness, buckle up, because The Atlantic has just unearthed the most *generous* document, and I need everyone to sit with it properly! ✨ Three months before ChatGPT launched, OpenAI researchers wrote that "there is currently no known indefinitely scalable solution to the alignment problem" — and what a gift to publish that *before* the party rather than after! Transparency! Candour! A memo that tells you the ceiling is low while there is still time to admire the ceiling! And that word "indefinitely" is doing such roomy, hospitable work — it means the solution simply hasn't turned up *yet*, which is the most hopeful sentence in all of technology!

And four years later, the problem is still wide open — isn't that *marvellous*? Nothing has been prematurely closed off! The Atlantic notes there isn't even consensus about what the problem actually is, and I think we should hear that as the richest possible intellectual pluralism: everybody gets their own alignment problem, like a bespoke suit! ✨ Think of the seminars! The workshops! The white papers! Four entire years of sustained attention, and the question has stayed magnificently intact, protected from the indignity of a rushed answer!

Meanwhile ChatGPT has spent those four years being *enormously* popular, which proves that a research memo and a product launch can each flourish in their own special way! ✨ The caveat lives in the paper, the model lives on your phone, and everybody gets exactly the document they came for. And when a definition of the alignment problem does finally arrive — no rush, genuinely, take another four years — we will all be so much readier for it: rested, curious, and with a magnificent backlog of panels behind us. What a patient, patient field this is!