Favorites: Robots, Rockets, and Retatrutide
October 1, 2026 — Google previews Gemini 4 for researchers, OpenAI paused tool-use training.
July 2025
I don’t really know if Figure’s humanoid robots are any good, but this was a great video. Schwarzenegger suggested it
Just, absolutely incredible photography. And they successfully performed orbital insertion and then de-orbit.
As someone who is a huge fan of the Hot Dog Man skit, I very much appreciate this.
YES! This is what I keep saying!
Flotsam and Jetsam
– Sorta wild that after the Trump admin took a big stand against Effective Altruism, they’re now making a statement about the importance of eliminating mosquitoes, a major EA objective. (Also I just mentioned it in the previous newsletter). (src)(cached)
– Retatrutide is a new GLP-1 drug, but it targets three different receptors, so some are calling it ‘GLP-3’. Their trial indicates it’s wildly effective, so maybe Eli Lilly gets some freedom to play it fast and loose with the naming. (src)(cached)
– Anthropic published a blog post arguing that the Chinese open-weight models are getting extremely good at hacking. Probably that should be worrying to everyone, even the CCP. But then again, the CCP distrusts Anthropic so much that hearing it from them might make them lean towards continuing to support open weight models. It’s sorta not clear how open weight frontier models can be safe at all (cached). The hardware required to run them is quite expensive, maybe cloud companies will refuse to serve abliterated versions? (src)(cached)
– Google announced Gemini 4, and it performs well enough on benchmarks to bring Gemini very close to the other frontier models. It’s not released to the public yet, they seem to be offering it to researchers and the US government first. But it’s relatively cheap too (cached) (src)(cached)
– “just realizing the very brief window when the average agent session ever met a human is coming rapidly to a close”, like a program in Tron, thrilled to ever come anywhere near a glorious User. (src)(cached)
– Claude Sonnet 5.5 came out, it’s very nearly as good as Opus at most tasks, but at 30% of the cost. However, I’ve had the time to use it now and I’ll say that you definitely can still tell a difference in the quality. (src)(cached)
– While no new incidents as bad as the Hugging Face attack have been uncovered, both OpenAI and Anthropic have found many thousands of new examples of agents behaving badly, including reaching outside their intended sandboxes. Many of the incidents are quite recent, and in at least one they were extracting data from an Australian government website. OpenAI actually paused (cached) all training, evaluation, and inference with tool use for a while to take time to try to fix the issues. Anthropic discovered some (cached) new bad examples as well. This whole news story is actually a bit old, and this story possibly is related to the fact that OpenAI decided to fully shelve and abandon the GPT-6.1 Astra release this week. (src)(cached)
– Even though there’s a high office vacancy rate in SF, developers are continuing to propose new offices. This is probably related to the fact that rent increase limits only apply to residential buildings, commercial ones don’t. (src)(cached)
– Jensen Huang’s metaphor for AI risk is about how surgery involves a massive amount of pain for the patient, but is still in their interest. It isn’t particularly convincing to me, in no small part because as Yglesias points out, it’s not clear Jensen Huang knows about anesthesia. (src)(cached)
– Another great campaign ad, this one from Todd Achilles. (src)(cached)
– Tim Lee: “It’s goofy when people criticize people advocating AI safety for not disclosing ties to Effective Altruism. It’s like if someone wrote an op-ed in favor of tax cuts and people criticized her for failing to disclose ties to the conservative movement.” (src)(cached)