2026-07-16
walden
ChinaTalk interview with Kevin Xu and Matt Sheehan discussing to what extent the events surrounding the release of Claude Mythos might affect the Chinese strategy of releasing open-weights models. Anyway, Z.ai announces Kimi K3.
Boaz Barak on how the good AI future isn’t the one where the “good guys” win, but rather the decentralized future where individuals retain their individual agency.
Tyler Cowen transcript on a talk speculating on different societal changes that are likely to result or be required post-AGI, primarily from the perspective of how individual lives of intellectuals are likely to change. One aspect worth mentioning is the need to rebuild our existing institutions; on that note, Matt Ridley on government services as suffering from Baumol’s cost disease, Elliot Paschal on using Claude to proofread contracts for likely disputes, and Brian Potter with a statistical analysis of how often federal laws tend to diverge between their intended purpose and actual usage.
Julianne Weiss interviewed in Asterisk Mag Dagger on Aid on the Hill and their efforts to retain PEPFAR and other forms of foreign aid.
Matt Reardon on backlash to Effective Altruism due to evangelism typically being given and received under a framing of moral realism which generates feelings of obligation and guilt, arguing that EA would be much better received if it was described in a more anti-realist manner, such as a hobby of optimizing the balance of happiness and suffering in the world1.
Mario Giancotti with a fantasy short story structured around the concept of being misled by survivorship bias. This is plausibly also a work of AI-inspired fiction in that it also serves as a demonstrates the inefficiency of reinforcement learning2.
Sam Kriss story about America forever and ever, meritocracy and freedom and growth infinitely without memory. Unironically, these ideas indeed predated America, and should continue to exist long afterwards.
Arbogast in the Futurist Letters reviewing Obsession as a successor to It Follows, tracing the course of dating discourse from the 2010’s fear of sexual desire3 to the 2020’s fear of being unworthy of love.
LRB new issue articles: David Runciman on American gerontocracy, Laleh Khalili on (respectable) women at port and sea, and political profiles of Einstein and Joyce.
While on this topic, coincidentally Jack Thompson has a repudiation of Overmuser’s justification for negative utilitarianism, which is based off a definition of suffering which does not actually appear to exist; and Both Sides Brigade with some claims about arguments about moral anti-realism from evolutionary origins which I don’t quite understand, for which Daniel Greco’s comment seems straightforwardly correct (Edit: also Liam Robins on how virtue ethics and utilitarianism are the same thing under anti-realism).
Not really related, but an interesting release by Pantograph of a Minecraft model using purely video recording data for training, using a description of video end-states as the goal for reinforcement learning. It’s not clear to me to what extent this will work for embodied robotics, given the significantly greater degrees of freedom required for real-world interactions, and differing POVs, which will often not include important elements such as the positioning of one’s hands and legs. But this does seem like it might be a promising means of extending the capabilities of robots into basic planning and task completion, assuming they have already been trained to be capable of fine-motor control. In any case, Pantograph also had a talk at Manifest on the potential implications if we can get general robotics.
Edit: I was thinking about this some more, as to how this is distinct from World Action Models such as DreamZero, and it seems to me that the distinction is exactly this aspect that it would function more as an action layer on top of a model which has already learned motor control, whereas WAM’s are trying to learn both at the same time, which presumably requires significantly more training data. My understanding is that there is a sort of divide between US and Chinese robotics labs, where the former are focusing purely on learning motor skills via teleoperation (and succeeding), whereas the US labs are trying to go fully end-to-end to learn motor skills and actions at the same time (and mostly failing). It’s interesting to consider how Pantograph seems to be more of a ideologically driven research company, because in a world where robotics goes through multiple stages of training, from fine-motor skills, to action learning, into fine-tuning; it seems to me that they would create significant value by enabling the middle step, yet with video data being abundantly available, they will not be able to capture any of it. Meanwhile, if teleoperation data ends up being tuned to specific hardware, this means that the bulk of the value will accrue to whichever manufacturer can make the most capable model, especially if the fine-tuning data flywheel can get set up before physical limitations cap out. The primary question is who will end up capturing the value from this final bit of fine-tuning, as to whether it will be controlled by the manufacturer (and colocated contractors), or by the applications themselves (or integrators), which depends on the extent to which fine-tuning needs to be done close to the application itself (or can be simulated), as well as how generalizable such training would be, which would affect the manufacturers leverage regarding the sharing and pooling of such data.
While on the topic of gender discourse, QC litigating 2010’s purity culture discourse. Also, Defending Feminism writing on the rationale behind questionable claims about dressing in a sexually attractive manner is “for oneself” (which reminds me of an old Piotr Pachota post on the topic of “sexiness”; very interesting that a significant faction within contemporary feminism seems to be converging on a sort of red-pill adjacent pragmatism with the goal of delivering maximally good outcomes for women, an apparent recognition of the downsides of over-ambitious ideology).