https://www.quarter--mile.com/The-Locals-Dont-Know fair enough; living in a place is not at all like being a tourist. locals have a daily routine. they probably aren't exploring all the most fun things to do. when tourists come to my city, i'm usually embarrassed about how little I can tell them about it.
https://defensesindepth.bio/10-big-projects-for-reducing-bio-x-risk/ these seem fine, and I applaud making a "wish list" in biosecurity public. i'd have expected a lot more...technical meat, though? like, nitty-gritty with specific threat models of how engineered pandemics are expected to work and what techniques will counteract them. these ideas seem to be based on a belief that we're very close to doomed when it comes to technical or policy safeguards, or even most types of preparedness. if you're talking about DIY respirators, your assumptions are extremely pessimistic -- maybe too pessimistic! yes, that sort of thing was useful in COVID, but COVID was a complete snafu. do you think we don't have time to stand up a more functional institution before the Next Big One hits???
https://en.wikipedia.org/wiki/Annie_Oakley her life was much more interesting than i realized. orphaned. horrific childhood. learned to shoot so she wouldn't starve. cutest marriage ever (he fell in love when she beat him at a marksmanship contest). taught 15,000 women to shoot, firmly believed that firearms were important for women's independence.
https://en.wikipedia.org/wiki/Calamity_Jane by contrast...she seems like kind of a loser? it's not clear why she's famous at all, apart from featuring in some highly fictionalized Westerns.
https://www.lesswrong.com/posts/cWFsCFyCttsiJwn2j/bringing-more-expertise-to-bear-on-alignment this is about connecting AI alignment to more "straight machine learning" research fields such as complexity theory, statistical learning theory, "computational mechanics" (the geometry of representations in neural networks), "singular learning theory" (mapping properties of a model to properties of its training data), and more. I think this is promising, if-and-only-if we aren't trying to prevent a catastrophe in like <2 years.
as an "AI moderate"/"foom skeptic"/whatever, i think it's entirely possible that it makes sense to start multi-year, heavily theoretical alignment projects, in addition to less-theoretical, more-urgent stuff around near-term misuse or accidental-catastrophe risks (policy, evals/monitoring, resilience stuff like "level up your cybersecurity" and "try to make formal verification real").
however, if you think we are literally going to die before 2030, starting a theoretical research program in a new paradigm is clearly futile.
my intuition is also that it's...sort of salutary? and neglected? to try to do more "serious" and ambitious theoretical research. like, people would be tempted to not bother with that even if it were a good idea. that's not in itself a reason to do theory, but it's on my mind.
https://arxiv.org/pdf/2508.14390v1 this is a bad paper (uses out of date models, narratizes results not in the data) but it does engage with the question of whether AI models are calibrated in the sense that they'll "bet harder" on questions they are in fact more likely to get right, if instructed to maximize their expected value in a betting game. (this is somewhat different from asking them to give their confidence as a percentage, from a prompting/persona perspective.) the answer is that Llama3.1 and Qwen2.5 are not calibrated on MMLU-pro, but are on the easier Trivia-QA...
https://arxiv.org/pdf/2311.08298 mostly eliciting confidence/calibration from AIs focuses on how to get them to be more calibrated, or whether you can use white-box methods like training a model or using logits to predict their odds of getting things right. I'm interested in calibration as an onramp to thinking about machine consciousness ("do they know what they know?") but the field isn't so much asking questions that would be useful for that.
https://arxiv.org/abs/2207.05221 "Language Models (Mostly) Know What They Know" is an awfully confident title for a paper that does not fucking prove that.
https://arxiv.org/abs/2601.17406 analyzes characteristics of AI-generated code from different models. mostly this is about language in the PR rather than in the code itself; one exception is that Claude Code has more conditional statements per line of code than other agents.
i'm inclined to believe that large rounds of big-company layoffs do not, in real life, happen when it's economically rational, but that there are times when it's socially unacceptable to do em, and times when it's socially acceptable. so you get these big waves when everybody's doing layoffs. it wouldn't be so correlated if companies were continuously optimizing the correct profit-maximizing # of employees. and right now "AI" is an excuse to do layoffs. it may also be rational to reduce headcount because of AI, in some companies, but I think the actual considerations are more complicated in practice. and there's a lot of spray-and-pray in practice, where layoffs are random or correlated with something irrelevant to performance, instead of more-productive employees getting preferentially retained.