Mythos has been rolled out to trusted users for a decent amount of time now, and is apparently being accessed by unauthorized users as well so I’m curious if there is any testimony on it’s capabilities from non-Anthropic employees.
I think Anthropic’s own claims are probably legit, I’m mostly just curious because I haven’t encountered any which is surprising to me, and it would be interesting to see how it feels compared to public SOTA.
I agree with your conclusion; though additionally I think two factors for why we don’t see such repentance often is just the general human tendency of not wanting to admit you are wrong (magnified in the case where a dictator admitting they are wrong is extremely costly personally and morally taxing), plus the type of personality that gets into those positions selects against this kind of behavior.
This tendency also reminds me of a potentially similar dynamic of scientists not admitting they are wrong https://en.wikipedia.org/wiki/Planck%27s_principle
Part of the difficulty in pointing specifically to how and when Wikipedia was politically captured is that this sort of ideological takeover happens in a pretty abstract and diffuse way without much of a paper trail, which is part of what makes it so hard to defend against.
Tracing Woodgrains documents some interesting cases like a major admin/user that consistently edits in bad faith and generally gets away with it or how figures like Mao get treated rather differently from other big dictators.
A good recent example I encountered was the Olympics Boxing scandal around Imane Khelif where having an SRY gene was a very salient point that was reported on and later confirmed, but there was a huge effort to conceal that info from the page, and then later downplay it heavily.
That said, I still think Wikipedia is a pretty substantial net good for providing information.
The recent controversy with Anthropic and the DoW has shown pretty viscerally how US politics can impact AI, and interactions will almost certainly grow more frequent and significant as AI continues to advance.
Something it brought to mind is how the timing of political events (particularly after US midterms in 2026, and after US presidential inauguration in January 2029), would likely act as a significant step change to the distribution of AI outcomes, and how that relates to the timing of AI reaching particular thresholds in capability and political salience.
edit: More specifically, I think predictions and actions based on those predictions about the future should take this discontinuity into account, and I haven’t really seen that, although I guess because of the nature of the issue, it might be detrimental to publicly take account of that.
An interesting piece of potential evidence in favor of this is that METR time horizons measurements didn’t vary significantly for ChatGPT and Claude models when using a basic scaffold as compared to the specific Claude Code and Codex harnesses.
https://metr.org/notes/2026-02-13-measuring-time-horizon-using-claude-code-and-codex/
I am curious what aspects of sci-fi utopias seem unappealing to you. Something like ‘The Culture’ for instance doesn’t have any downsides I can think of.
I really wish someone tried out o3/gemini with a weaker harness (say equal to claude), which is where it would be more interesting and also it would make a cross-model comparison easier.
I just want to note another data point about reforming institutions which was postwar Iraq. De-Baathification was an explicit policy undertaken to explicitly remove and replace members of the government associated with the Saddam affiliated Ba’ath Party, and it’s generally considered a failure and having lead to a lot of sectarian violence, the rise of ISIS, and generally contributing to an ineffective government afterwards.
It’s a somewhat different situation since that was more of an ideological project, but is I think notable and relevant.
Meta is delaying their Behemoth model launch because of disappointing evals.
This is another major lab (both OpenAI, Anthropic have also experienced this) that has seen disappointing results in trying to scale their model via raw parameter size into the next generation, which suggests to me that there really is some sort of soft or hard wall at this size. It’s good news for people favoring a slow/pause, though of course there is now RL to pursue. I am genuinely curious what’s going on though; it seems like maybe it’s just getting enough high quality tokens that’s an issue, and synthetic data is too hard to get or it could be a qualitative shift like a reverse of the original change with LLMs.
I definitely think this should update the priors of RSI folks though, because if these sorts of barriers keep cropping up along different avenues of scaling, I would expect linear increases in intelligence rather than exponential ones.
This is also somewhat commingled by the lab specific issues with Meta that seem very hard to ignore now (i.e. bad execution), so maybe this particular instance shouldn’t update you too much, but it is still of note.
If we’re going full meta, these sorts of questions are always happening in the sense of being posed as hypotheticals, rather than literally happening to you, so the only real effect of your answer is leaking information about how you would behave. This doesn’t require alternate realities or retrocausality.
Also, I don’t know if you count this as alternate realities, but even if the hypothetical were truly appearing to happen, it could just be a simulation spun up by the predictor.