Making Intelligence Ubiquitous
Leopold Aschenbrenner, a former AI alignment researcher at OpenAI, proposed a concept called Human Equivalence. It measures the capacity to produce tokens with unit time and cost. Though the training process demands significant computing power and consumes electricity, the inference process runs efficiently at a low marginal cost. Rather than embodying manufacturing know-how in robotic pipelines, knowledge work is prone to be replaced by AI agents driven by LLMs. Entry-level programmers, lawyers, teachers, and even doctors become vulnerable. The execution process is automated, while humans still need to nail the deal and take responsibility for it. The computational part would be compressed and processed in parallel, leaving us confronting the incompressible parts. We need to have the bravery to say "that's enough" and make the choice. Stuart Russell argued in his book Human Compatible that the judgment of AI shall be inherently uncertain to avoid catastrophe.