Nemo

first ever blog post from me: “straight lines on graphs” as a researcher at @METR_Evals i’ve had hammered into me again and again that AI progress is (not only rapid) but remarkably _regular_. this intuition then suggests a bunch of useful mental models which aren’t obvious without the initially counterintuitive starting place. i focus on “jagged capabilities” a la @emollick, acceleration from RL a la METR/@EpochAIResearch, and compute slowdowns a la @whitfill_parker as case studies. (but there are many other phenomena i mentally organize in this way!) i don’t expect this post to convince people who are suspicious of this intuition about its value — i think this wouldn’t have worked for me when i was starting out skeptical! but i hope it gives a sense of some helpful basic tools that people focused on AI use to understand the past + future. post linked below.