The "sharp left turn" is a failure scenario described by Nate Soares of MIRI in a LessWrong post from June 2022 (Soares, 2022). At some point, a system's capabilities start working in domains far removed from its training, but its alignment does not generalize with them; in Soares's formula, capabilities generalize better than alignment. The techniques that had kept the model in check stop working exactly when they are most needed, and all at once. The image is of a curve bending abruptly on a chart.
The favorite analogy is evolutionary. Natural selection optimized hominids for reproductive fitness; when human intelligence took off, our capabilities generalized to agriculture, science and rockets, but the evolutionary "objective" did not: we invented contraception. It is the same intuition as mesa-optimizers, with the emphasis on discontinuity. The song uses it to cap its most technical verse: "Sharp left turn and there you are" (osmarks, n.d.).
For doomers, this is why today's good results (obedient models, passed evaluations) are not reassuring: the test that matters happens on the other side of the turn. Critics reply that, until someone specifies what would trigger it, the scenario is hard to refute, and that observed progress looks more like the continuous curves of scaling laws than like breaks. Pressman read it as a symptom of a shifting narrative: first the problem was that there was no way to point at human values; when language models made that less plausible, it became inner optimizers and the sharp left turn (Pressman, 2023). Phenomena such as grokking or emergent abilities are cited by both sides: as a sign that discontinuities exist, or as artifacts of the chosen metric (Schaeffer et al., 2023).