@LiamFedus: GPT-4 + ChatGPT looked like the start of a Bostrom-style “decisive advantage → singleton” story. In...
GPT-4 + ChatGPT looked like the start of a Bostrom-style “decisive advantage → singleton” story. In 2022, GPT-4 felt ~18 months ahead of its time, and then it was soon followed by a viral product hit, providing a rich source of feedback for model development. But instead, a few years later, we're in a multipolar world with several competitors vying for the lead, shipping weeks (or days) apart. And, quite happily, this competition led to likely one of the largest consumer surpluses of all time. It's interesting to reconsider this now with the rise of coding agents (Claude, Codex, etc.). These are a more direct contributor to the next generation of models — a recursive self-improvement loop. But with intense market competition, no company seems to withhold its best tech for internal use. Instead, the latest and greatest is rapidly made available to consumers (and their competitors—see Anthropic banning xAI from using its coding models), implying no asymmetric advantage since everyone benefits. In 2026, do coding agents push us towards Bostrom’s singleton story, or does market competition keep the current status quo?