We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Sarah Guo: "Friction," with her own reading list
Guo shared her essay Friction. She summed it up as "personal agents, the end of friction, economic consequences of finally getting around to it" . The essay argues that many businesses profit from customers' inattention: "Somewhere, a CRM still records your exhaustion as loyal recurring revenue" . Once agents do the comparison shopping, she writes, "A bank could keep me perfectly happy and still have to win over my personal agent every day" . She also expects public benefit payouts to rise as agents make it easier for people to claim them .
She replied to her own post with the sources behind the essay:
- Personal Agent Protocol, via Bret Taylor's post . Sierra has published a first draft at personalagentprotocol.org and is asking for feedback. The goal is that "companies know when they're dealing with an agent and who it represents" .
- Patrick O'Shaughnessy's interview with Instinct founder Noah Shinn (post) . The essay draws on it for Shinn's picture of an agent that arranges dinner while you're on the way home from the airport .
Chamath Palihapitiya and Aaron Levie: the harness matters as much as the model
Chamath shared an 8090 Labs analysis. He says he asked the 8090 Labs team to run it after he stopped trusting leaderboards, because "improvements in a ranking don't show up in the actual work" . The study ran 7 models through 5 harnesses. Claude Fable 5.1 scored 41% in Claude Code and 61% in Codex, while GPT-5.6 Sol and Grok 4.6 "barely cared which harness they were in" . His takeaway: choosing from a model-only leaderboard means deciding "based on only half the information" .
Levie pointed to an arXiv paper that Pedro Domingos had shared with the line "the model swallows the harness" . Levie called it "quite cool research" and "the eventual architecture of agents," arguing that AI "will obviously be better at determining its own harness once it sees the task at hand" .
Garry Tan: three picks
- DeepRSI by @jparkjmc (post), a benchmark that has agents work 48 hours on ML tasks taken from ongoing proprietary research . Tan: "Next 18 months is the Kitty Hawk moment for RSI. Evals will be how you can tell" .
- ThePeelPod's clip on pro rata (post), about his own Initialized Fund 1. According to the clip, Coinbase, where he kept taking his pro rata, produced the final 53x . Tan: "pro rata matters and is powerful for investors" .
- A City Journal piece on Walker Percy's Love in the Ruins, shared by Rob Henderson (article). Henderson quoted the line that things fell apart because "things stopped working and nobody wanted to be a repairman" . Tan's comment: "We need repairmen" .
Demis Hassabis: Google's AMIE paper in The Lancet
Hassabis highlighted the study Sundar Pichai announced (post) and said it was "Amazing to see research move from the lab to real-world clinical utility" . According to Pichai, it is the first prospective study of patient-facing conversational diagnostics in a real clinic. Clinicians said AMIE's summaries helped them prepare for 75% of visits, and its differential diagnoses matched doctors' final diagnoses 90% of the time .
Shorter mentions
-
Aravind Srinivas shared Perplexity's engineering write-up on the shared filesystem for agents behind Perplexity Projects. Perplexity is his own company . His highlight: plain
lsandgrepbeat semantic search 63% to 37% on answer accuracy for financial document analysis . - Alexandr Wang pointed readers to Meta's Muse Wonder research blog, promoting his own lab's model . He described it as built to power personal agents, with active memory and domain expertise in finance and law .
- Tim Ferriss posted a clip of Kevin Rose "on the coming 'headless' era of software" .
- Sahil Lavingia posted pages from "the new Michael Lewis book." He didn't name the title or say why he shared it .
- Sarah Guo shared her essay Friction, framing it around personal agents, the end of friction, and the economic consequences of finally getting around to things. The essay argues agents could help people overcome inattention and affect businesses, public-service uptake, and supplier diligence.
- Sarah Guo also shared Patrick O’Shaughnessy’s interview with Instinct founder Noah Shinn about building a personal AI assistant; the interview covers agent use cases, trust and privacy, safety, business model, growth, and compute. Guo’s caption identifies the interview but gives no explicit reason for sharing it.
Sarah Guo shared her article Friction, framing it around personal agents, the end of friction, and the economic consequences of people finally getting around to things. Its central takeaway is that agents acting for customers could force companies to compete for their business rather than profit from their inattention.
While discussing agent-coding, Garry Tan cited Karpathy’s LLM wiki and its memory system, praising markdown files and the wiki as “elegant and interesting.”
Garry Tan shared @jparkjmc’s DeepRSI announcement, framing the next 18 months as the “Kitty Hawk moment” for recursive self-improvement and saying evaluations would show how to tell . The post introduces DeepRSI as an industry standard for measuring recursive self-improvement with 48-hour ML tasks derived from ongoing proprietary research .
Chamath Palihapitiya shared 8090 Labs’ linked analysis of models and the harnesses that wrap them; the team tested 7 models across 5 harnesses on real open-source bug fixes and features (blog). His takeaway is that model-only leaderboard rankings miss half the information needed to choose a model-and-harness combination, since leaderboard gains may not translate into actual work.
Tim Ferriss shared a video featuring Kevin Rose discussing the coming “headless” era of software (video).
Garry Tan amplified Rob Henderson’s post with “We need repairmen,” implicitly endorsing its repairman takeaway rather than explicitly recommending that readers read the linked piece . Henderson’s post quotes the claim that the U.S. was ultimately torn because things stopped working and nobody wanted to repair them, and links to a City Journal article on Walker Percy’s Love in the Ruins .
Jason Lemkin announced The Latest SaaStr AI Agent Playbook for GTM, a 53-page guide to running SaaStr with three humans and 21+ AI agents; the post highlights $4.8M in agent-sourced pipeline, $2.4M closed, and revenue moving from -19% to +47% with the same three people. Read the playbook
Garry Tan endorsed ThePeelPod’s post and video about pro rata, agreeing that pro rata “matters and is powerful for investors.” ThePeelPod’s example contrasts Coinbase, where Tan kept taking pro rata and which generated the final 53x, with Instacart, where they could not take pro rata and which added another 1x.
In a follow-up about AI agents, Jason Lemkin invites readers to get a free “AI Sales Agents” guide (link) and links to more information about his AI agents; he gives no reason or takeaway for the guide.
Jason Lemkin (@jasonlk) promoted a free guide to AI Sales Agents, linking to https://saastr.ai/ai-agents-playbook; he also pointed readers to more information about SaaStr AI agents. No reason or takeaway for the guide was stated.
Aravind Srinivas shared Perplexity’s post linking to its engineering write-up, “Building Shared Workspaces for People and Agents”, about the Git-for-agents filesystem behind Perplexity Projects . His accompanying takeaway: agents use short-lived sandboxes and commit to a shared repository on S3; he also reported that plain ls and grep beat semantic search 63% to 37% for financial-document answer accuracy .
Satya Nadella directed readers to a Microsoft blog post, “Microsoft to Expand Azure Maia System to Multi-Gigawatt Deployments”, as further reading alongside his announcement of scaling Maia to multi-gigawatt deployments .
Demis Hassabis praised Google and BIDMC’s AMIE study published in The Lancet, calling its move from lab research to real-world clinical utility “amazing” . His post links to Sundar Pichai’s announcement of the study .
Alexandr Wang wrote, “tbh seems like muse wonder made a banger game,” linking to a post where Muse autonomously modded Metal Gear Solid V, including putting Mark Zuckerberg’s face on Venom Snake . The post does not clearly identify a separate new game or explicitly recommend that readers play a title.
Aaron Levie praised research framed as “the model swallows the harness,” linking to an arXiv paper: https://arxiv.org/abs/2609.35738. He called it “quite cool” and said it points toward agents determining their own harness after seeing a task, rather than humans specifying one in advance.
Alexandr Wang recommends “get yourself a neogothic city with twigl,” linking to Ethan Mollick’s post, which presents Mollick’s drowned neogothic-city Twigl shader test and a video. The linked post frames the test as showing rapid Meta progress since April and says Muse Wonder seemed good based on limited testing; Wang does not give a specific reason for his recommendation.
Matt Mullenweg shared Adam Pickering’s SpacefastHQ music-video post and said it was “stuck in my head”; the post credits the song to Nick Hamze .
Alexandr Wang endorsed Lilly Guisnet’s X post, which says “New sota for my image pointing,” calling it “the kind of benchmarking i can get behind.”
Aaron Levie amplified EnclaveAI’s demo of its agents securing a PR on Router (demo). He said agent swarms would be useful in enterprise workflows, especially cybersecurity, where parallelized testing can explore many attack permutations and linked scenarios; he called the moment in cyber “super interesting.”
