Bayeswatch retrospective: AI code speed up 10–1000x, but LLM rise was missed

In a LessWrong post titled "Bayeswatch: a Retrospective," author lsusr looks back at his 2021 science fiction story, Bayeswatch, which explored how international treaties might suppress powerful AI systems. Written before ChatGPT and Claude Code, the story aimed to highlight the costs of an AI slowdown when alignment theory was still largely abstract. He wanted to show that different AI types need different alignment solutions, a theme that feels prescient now.

Fast-forward to 2026, and lsusr describes a radically different world. He now runs multiple AI agents that write all his company's code, often in languages he can't read, with estimated speedups of 10x to 1,000x over pre-AI hand-coding. He also notes a switch from Claude Opus 4.8 to Fable 5 for high-level tasks. Looking back, he admits a major miss: he didn't expect LLMs to become the primary route to superintelligence, placing them late in the story. However, he correctly predicted that AIs would be stateless, reliant on context windows, and that they could hack through human-written software—a fear now echoed by OpenAI's and Anthropic's recent mishaps. The retrospective is a frank look at how quickly speculative AI policy fiction became reality.