More the later than the former… they are better than purely marketing focused stuff pushed out by the LLM companies, and if you dig through them and read between the lines you can occasionally sift out useful details. Like here is a pretty solid sneer digging through Mythos’s ‘system card’ and pointing out all the ways it contradicts the hype and press headlines Anthropic was pushing.

But even so they have some big problems…

• the benchmarks the system cards reference are kind of useless and heavily gamed
• LLM companies want to keep lots of details secret from competitors, so fundamentals like number of model weights or parameters or size/quality of the training data set or other training specs are deliberately left out
• lots of the stuff they reference is booster garbage and/or doomer crit-hype
• they tend to be long, wastefully so, imitating the length of academic papers without having the corresponding amount of depth or information
• despite their length and wordiness they also neglect basic practical usage advice that isn’t even proprietary (or at least would be bound to leak if you poke around with the model at all and thus not worth keeping secret in the first place). Like not even big picture stuff I mentioned in my second bullet, but really simple stuff…

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 half baked sham of a 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.