One line of calculus sinks the simplest way a model could hide a goal from training
The worry Evan Hubinger named in 2019 survives the collapse of its own first example, and hangs on how far real training departs from the idealized gradient descent that collapse assumes.
Segment Anything returns a mask and never a name
Kirillov and his Meta coauthors let the 2023 model draw 99.1% of its own billion-mask training set, then measured it losing to a task-specific model on the standard segmentation score.
Two AIs reported the same 35/42 gold at the 2025 IMO, and the IMO graded only one
An IMO medal is a proof scored by hand against a rubric, which makes the grader part of the number and turns one gold medal into two different measurements.
Quantization rounds a model's weights to a coarser grid, and a few oversized values decide what that costs
A careful 4-bit scheme costs a 7-billion-parameter model under 1% of its perplexity, but round the same weights naively, or below four bits, and its answers can collapse.
Meta's Galactica invented citations for real researchers, and its public demo lasted about two days
The demo warned on every output that the text might be invented, and Meta's own paper had already measured the model fabricating most citations to rarely-cited work.