Amodei said that "we" were able to provide a strong theoretical basis for having conviction that superposition was the right hypothesis to pursue; the surrounding text does not say who "we" refers to.

Judge as of 2025-02-18, reading 'we' as Anthropic and/or its interpretability researchers (including Amodei's collaborators). True: published work by those researchers by that date (e.g. 'Toy Models of Superposition', 2022) gives a theoretical or toy-model basis for the superposition hypothesis that motivated pursuing it (e.g. through dictionary learning or sparse autoencoders). Mostly true: such work exists, but it is mainly empirical or toy-model rather than a 'strong theoretical basis', or it was written partly by people not then at Anthropic. Mixed: the superposition idea was central to the work, but the theoretical basis came mainly from others. Mostly false: the group's role was minor, or the hypothesis came almost entirely from others. False: no such work by the group exists.

Set before the research began.

1. E1 Toy models of superposition Anthropic · 14 September 2022 Own record

   “In this paper, we use toy models — small ReLU networks trained on synthetic data with sparse input features — to investigate how and when models represent more features than they have dimensions. We call this phenomenon superposition.”

2. E2 Towards monosemanticity Anthropic · 5 October 2023 Own record

   “we outline evidence that there are better units of analysis than individual neurons, and we have built machinery that lets us find these units in small transformer models.”

E1 is Anthropic's 2022 'Toy Models of Superposition' paper. It used small ReLU networks trained on sparse synthetic features to study how models represent more features than they have dimensions, a phenomenon it names superposition. E2, Anthropic's 2023 'Towards Monosemanticity' paper, reports building tools to find units of analysis better than individual neurons, which is consistent with pursuing that hypothesis. Both items are the group's own published work, and the basis they show comes from toy models rather than formal theory. That fits the 'mostly true' tier better than a 'strong theoretical basis'.