2026-04-09
tom sawyer
Dean Ball with some comments on Claude Mythos, and Kai Williams with specific highlights on its cybersecurity capabilities. It actually feels to me like many people are somewhat overupdating based off the reported performance here, because insofar as Opus 4.6 already demonstrated strong capabilities in coding, security analysis, and code use, it would actually be underwhelming if the Mythos failed to follow the usual trend of a second model rapidly climbing the ladder after a first model showed that it could grip the first rung. The cobra is the same size whether its hood is flared or not; what underpins accurate predictions is not how big something looks or how seriously people are taking things, but mechanistic models for what is or is not possible1.
Joshua Saxe also has a moderating piece on Mythos, noting that there is demand for cyberattacks is limited, which means that worry about cyberattacks should remain bounded. One could respond with the claim that demand is not fixed, but actually a function of supply, but to me this isn’t an argument against the existence of open models, just that frontier models should not be open2. (Edit: Nathan Lambert on this topic).
Alexandr Wang announcement of Muse Spark. Quite interesting given recent rumors about Meta Applied AI.
Adam Kroetsch on the history of attempts to reverse the bureaucratization of the US clinical trials process, which reads to me very much like a description of a software monolith that continued to accumulate features without any refactoring. Insofar as “large simple trial” are essentially analogous to processing bulk modular inputs via something like MapReduce, it seems obvious that simply changing the inputs will not do anything for efficiency, given a system which was designed around handling inputs that are bespoke and piecewise. (Edit: possibly related, Jenn Pahlka on the refactoring of the Louisiana Department of Education).
Luis Garicano on institutional factors affecting differential adoption of new technologies, including whether trust or merit is considered the more important factor for the purposes of hiring and promotion.
Aaron Kaufman with an interesting theory as to why social anxiety is often accompanied by self-loathing, from simulating potential negative reactions from others and mistaking that as baseline reality. I mostly agree with this, but it seems to me that much of the discourse around things like social anxiety and people-pleasing ignores the fact that expectations of social behavior are actually just another set of norms more or less arbitrarily chosen for the purpose of social coordination. If everyone also operated according to a protocol of always considering other peoples thoughts and feelings first, then that would also produce a coherent set of societal expectations and norms. But in the US, there are higher expectations for extroversion, openness, and neuroticism in one’s public persona, and lower ones for agreeableness and conscientiousness3, which means that anyone with very high levels of agreeableness will basically never have their instincts towards social harmony be reciprocated; of course this leads to self-loathing, because this is being constantly implicitly told that your social status is extremely low4. That being said, moving to Japan is not really a solution, because the normative personality of a society is only one aspect to deriving a sense of belonging from within a culture. But it seems to me that this understanding, that people-pleasing is not a moral failure but one of social coordination, is probably a much more robust way to get highly agreeable people to change their behavior. Actually, it’s very easy to get an agreeable person to behave disagreeably; just convince them that’s what other people want.
Chris Arnade interview with Colin Marshall on Korean self-perceptions.
Lost Futures on the history of the Langfang Republic.
Hill Citizen on how Archipelago would handle the existence of the nuclear bomb.
My current best theory (~10% certainty), based on various circumstantial evidence, is that RLM-style recursive chain-of-thought, both as a means for improved reasoning and as a means for generating training data, is why specifically coding, tool-use, and reasoning capabilities have improved so much. If so, this means performance in these areas will continue to improve, but notably there are also many tasks for which this approach is less well suited.
With the caveat that many vulnerabilities exist in the interaction between multiple pieces of software, and it will not be possible to preemptively scan every single possible combination in advance. This is why I foresee a period of software consolidation incoming, since there could be significant advantages to having every piece of software you run exist within the Google monorepo, particularly if you live in a country without privileged access to frontier models. I saw a joke on Twitter (cannot find) that IBM’s claim that the world will only need five computers wasn’t wrong, just early.
Source: I made it up. Basically you can search for Big 5 personality traits around the world, but then you get nonsensical results like Brazil being the most conscientious country in the world, exactly because people are reporting their personality according to what they understand as the social expectations from within their own country. When I say America is high neuroticism, I mean that in one’s public persona, America is relatively more welcoming of volatile expressions of emotion, based on the norm of authentic expression, which in other countries might be perceived negatively as a loss of self-control.
On that note, Alexander Wales has a post on the totalizing nature of status discourse. Personally, the reason I will often use status as an explanation for things is exactly because theories of social interaction are more useful the less you naturally identify with them. If you already intuitively understand status and use it to model social dynamics, then that’s just telling on yourself. But otherwise, learning about these sorts of concepts is like unlocking an entirely new sense, from which many previously nonsensical interactions become much more understandable.