Patrick Collison: Matt Faherty's report on the NSF

Collison called Matt Faherty's report on the National Science Foundation "excellent." His reason goes beyond this one report: "I wish there were more deep investigations and assessments like this of government agencies in the US, the EU, and elsewhere." @patrickc

Balaji Srinivasan: Vernor Vinge's True Names

In his TOKEN2049 talk "The Zcash Moment," Srinivasan cited "a great science fiction book… called True Names" by Vernor Vinge. He summed up its premise: people go online under an alias, and "if the government discovers the connection between your NIM and your real name, they'll go to your house and kill you." TOKEN2049 He used it to show why Satoshi's anonymity was "another form of decentralization." Because the pseudonym wasn't tied to centralized records, attackers had nothing to go after beyond the name itself, so there was no "radar lock." TOKEN2049

David Sacks: two replies on the AI bill's FLOPs threshold

Sacks wrote that a new "Anthropic-written safety bill" would define frontier models at 10²⁵ FLOPs. He argued that line is outdated, since the frontier "actually is today" at 10²⁷. He also objected that the count is cumulative and includes fine-tuning and reinforcement learning (RL). @DavidSacks He then pointed readers to two replies:

• @KryptonCGO notes that Meta's Llama 3.1 405B used 3.8×10²⁵ FLOPs and has been downloadable since July 2024. Under the cutoff, "a two-year-old open-weight model" would count as frontier. @DavidSacks +1
• @sagivo writes that because the count is cumulative, his weekly post-training on small models could mean "filing compliance paperwork for models that are nowhere near frontier." @DavidSacks +1

Harrison Chase: Factory's blog posts on model routing

On 20VC, the LangChain founder said "factory does a tremendous job" at routing (sending each task to the right model). He added that Factory has "written a number of technical blogs that are very well done" about how it does this. His view is that routing has to be "harness aware," meaning it depends on the agent's setup and on the domain. 20VC with Harry Stebbings

Guillermo Rauch: Tim Sweeney on AI-assisted development

Rauch quoted Sweeney's post and called it "Well said!" @rauchg Sweeney says about 70% of his time writing Unreal Engine 1 went to "typing and refactoring." He also says AI has surprised him as "a sort of insta-Carmack" that can solve hard problems. He concludes: "What matters is the end product, whether it's a game or an engine." @TimSweeneyEpic

Shorter mentions

• Elon Musk recommended Gad Saad's book on "suicidal empathy." He said the degree to which people in the West fall prey to it "continues to astonish me." His post doesn't give a title or link. @elonmusk
• Paul Graham linked @lefineder's post on Norman land grants after 1066. It says 57 followers were given control of half of England's land, and English landholders held 8% of landed wealth. @lefineder Graham's takeaway: before you could get rich by creating wealth, you did it by winning "a military or political dispute." @paulg
• Marc Andreessen named You Can See Everything, Primetime and Digger as making 2026 "the best year for film since 2015." His 2015 examples are Sicario, Ex Machina and Mad Max: Fury Road. "At long last, movies are back." @pmarca
• Martin Casado shared Ekimko's post of clips showing an ASCII MMORPG, a game built entirely from colored text. The clips are credited to TikTok user @Ansenjeo. @Ekimko Casado called it "Sooo incredible… reliving the early Internet all over again. But juiced 10000x." @martin_casado
• Jason Lemkin promoted his own SaaStr piece on the cuts at Cloudflare (20%), Atlassian (10%) and Bending Spoons (75%), so this is self-promotion. His takeaways: "Small teams ship more," and "the cuts worked where AI agents already did the work." @jasonlk