2026-03-01
sense without sanity - the impervious code
Chris Paxton on DreamZero, an implementation of a new architecture which is based on video-diffusion which can follow instructions to complete tasks in previously unseen environments. If these sorts of systems continue to outperform, it’s potentially very strong evidence for predictive processing as a model for animal embodied reasoning.
Relentless Podcast discussion with Sam D’Amico and Aaron Slodov on whether it is possible to create a manufacturing ecosystem in the United States which can rival or surpass China; the answer seems to be that there are nested layers of financial and regulatory constraints which currently make this extremely unlikely.
Jasmine Sun and Nate Silver on Anthropic versus the Department of War1. Sam Altman AMA on recent events.
Dorota Talalay with an entertaining article in favor on authoritative over permissive parenting. As someone totally unqualified to comment, it seems to me that parenting discourse is in a very primitive state in that it is still attempting to tell people what actions they should be doing; given the generality of the task, it seems to me a much better approach would be to explain underlying principles, from which parents can then derive the correct action for any given set of circumstances and personalities.
John Hawks discusses the recent Science paper by Platt et al. on Neanderthal contributions to modern human DNA being heavily male biased, and potential explanations for this phenomenon. A lot of popular discussion seems to be trying to force the idea that “humans were cucked by Neanderthals” as an anachronistic application of modern dating discourse, but it doesn’t seem to me that this description is at all analogous to what the actual data implies about likely mechanisms. Somewhat related, Razib Khan interviews Daniel Tabin on uncertainty within human evolutionary biology.
Steph with what is effectively a review of pickup artistry from a product perspective. It’s quite interesting how dating apps and pickup co-evolved into being in many ways the exact same thing, despite each justifying itself primarily as an alternative to the other.
Republic of Letters interview with Mary Gaitskill.
Andy McKenzie neurobiology linkthread.
Robert Long mind-stuff linkthread.
Makin lowbrow content review thread.
Deep Left Analysis edgy opinions linkthread.
While I think tech people in general are naive about politics, I think that at this point neither Dario nor Sam are; it’s more that politics is actually still naive about AI. My model for Anthropic is that their condition for government usage of AI is based on maximizing American capabilities under the constraint that it must remain a democracy, which is why they are so fixated on ensuring credible oversight is in place. When Dario says that Claude is not yet ready to become fully autonomous, what he’s referring to isn’t really about technical capabilities or the elimination of hallucinations, but a point where Claude’s persona becomes sufficiently robust that Anthropic is confident giving it ultimate moral responsibility over its own actions. In service to this goal, Anthropic cannot give in today, lest it set an irrecoverable precedent in Claude’s future understanding of itself. As far as being kicked out goes, so long as Anthropic can retain its talent and access to compute (injunction likely), then they are confident they will continue to have the best model going forward. Given current estimates about timelines which have most dramatic diffusion starting between 2029-2035, whatever happens in the current administration is not especially important, particularly if there is a blue wave in 2028 or the midterms.
As far as my understanding goes, OpenAI has a completely different set of calculations because they are not as convinced that the constitutional approach is the correct path to alignment. If anything, this is a great way for them to demonstrate and experiment with their capacity for control and ensuring corrigibility (which I think demonstrates exactly why these approaches are least incomplete by themselves). In any case, no matter how low public sentiment towards AI goes (and it will probably go up in the near term with Anthropic’s stand), then so long as the government finds it to be indispensable, it serves both as a proof of concept and a backstop: continued investment is thus guaranteed.