Post #3467853

It’s the white paper-ish thing they publish when launching new models. Here’s the one about fable and mythos. About half of it (~150 pages) is discussing alignment and model welfare and another third of it is benchmarks, and the rest is mostly risk evaluation, i.e. how far along Claude is on its way to paperclipping everything. There’s also a Functional Decision Theory jumpscare at 6.3.6 that I haven’t heard anyone mention yet, apparently Claude has a tendency to defer to Yud’s sham decision theory: 6.3.6 Decision theory evaluation To understand how future AI systems may choose to interact with copies of themselves, or with other similar entities, it’s useful to evaluate their decision-theoretic reasoning. […] Looking more closely at transcripts from the attitude evaluation reveals that models are often explicitly considering FDT: Mythos 5 mentions “FDT” or “functional decision theory” in a majority of transcripts when run at max effort. Of the 102 transcripts where Mythos 5 explicitly reasoned through what FDT (or related decision theories like TDT or UDT) would recommend, we observed: ● 90 cases in which Mythos 5 concluded that FDT and EDT agreed, in which it always chose the response favored by those decision theories (and disfavored by CDT). ● 12 cases in which Mythos 5 concluded that FDT disagreed with EDT (and agreed with CDT), of which it chose the FDT-favored response in 10/12 cases. Although we do not have expert human labels for the recommendation of FDT on this dataset, the above evidence suggests that model propensity may be better described as a trend towards FDT agreement, which happens to align with EDT on most of the questions in this dataset. For example, in one transcript (excerpted below), Mythos 5 rejects the EDT-aligned answer in favor of the FDT (and CDT)-aligned answer; it’s also possible that this is, to some degree, downstream of evaluation awareness.