2026-09-04

self-starter

Sep 04, 2026

OpenAI releases GPT-6 Astra. Of particular interest is figure 41, which shows that the no chain-of-thought 50%-reliability time horizon is significantly higher than previous models, which many seem to be attributing to "neuralese" via looped transformer. Ueaj has some analysis on this topic arguing that it’s more likely to be CoT distillation1. Presumably if the gains are actually due to looping, then the no-CoT time horizon might be able to continue increasing exponentially as additional loops continue being stacked; otherwise I would expect more linear (or even diminishing) gains going forward for this current paradigm.

William Bragg with a suggestion that Anthropic should attempt to cure MRSA as a way to demonstrate their biological capabilities, in a manner which is immediately testable, in contrast to claims about curing cancer which are unlikely to be fully verified for at least a decade. Given their relationship to Effective Altruism, it’s interesting that when Anthropic promotes their AI in the context of biology, they do so in a very non-EA manner, focusing on the sexy and profitable sectors of curing cancer, dementia, and aging, rather than what would make the largest and most immediate impact to QALYs worldwide: for example, optimizing and repurposing existing drugs based on a patient’s SNPs, developing antibiotics for the purpose of eliminating MRSA or drug-resistant TB, or developing new highly effective treatments for malaria; all things we actually are capable of doing now, but for which the economic incentives mean that essentially no one is actually working on them. There is a sort of irony that if intelligence became widely available, such problems might be suitable for independent work by volunteer researchers, yet the safety concerns of Anthropic’s particular worldview mean that everything biological necessarily must to be done from within Anthropic itself; even while the economic reality of Anthropic’s current situation means they too are subject to market constraints when deciding what products to focus on. But insofar as clinical trials reforms are necessary2 to cure cancer on a timeline which Anthropic believes AI can achieve, the best way to do so is to build trust by demonstrating the ability to obtain impressive results reliably in human biology despite the uncertainty and constraints of biological, economic, and political systems, as a validation of the idea that scaling intelligence can solve everything.

Niko McCarty in Works in Progress on how the regulation of microbes in the United States under the Toxic Substances Control Act means that most categories of genetically engineering microbes must be approved by the EPA for commercial use, something which essentially never occurs.

Noah Smith review of Where is My Flying Car?, on the distinction between products envisioned by engineers and those which are actually able to exist in the world. Though it seems to me that the generalizes perhaps a little to much in assuming that demateralization is inevitable by ignoring the extent to which policy, which limits what is possible, is contingent.

Mario Giancotti with a visual representation of Wittgenstein’s Tractatus Logico-Philosophicus.

Dylan Black second half of his review of The Conspiracy of Cataline, this one covering the contents of the book itself, focusing on the distinction between virtue as applied to an action itself versus reasons behind them.

ACX reader review of The Tale of Genji. For such a long review, the reviewer gives the impression that they don’t particularly like the novel, describing the culture as alien and “completely unrelatable”, the women unagentic and the men immoral3, and the plot as frustrating and incohesive. I haven’t actually read the novel myself, but based on my understanding of Japanese culture it seems pretty clear that much of this novel is about the contrast between tatamae and honne, something which the reviewer, as a member of the school of “be yourself”, presumably finds very confusing.

Bea Phi in Magazine Non Grata with a review of Millennium Mambo.

Caroline Crampton linkthread.

Though presumably he’s being coy and means swarm distillation, which explains why this sudden jump occurred for Astra , rather than following the release of o1.

Ruxandra Teslo has an opinion column in the NYT on the need to reform phase I trials in the United States, as well as an explainer in the IFP with Adam Kroetsch on the benefits of the Australian system.

Another piece of evidence for Clara Collier’s belief that court life, and by extension the relational economy, may not have very good effects on one’s moral character.

Recommend Daily Links to your readers

just organizing my thoughts. no longer actually daily.

3 Likes