We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Bill Gurley's AI safety reading list
Gurley posted a series of recommendations that share one theme: he wants AI safety treated as an engineering problem.
- "AI Alignment: Power Disguised as Morality" by @pentagoniac (article). Gurley's comment: "It's becoming clear that the notion of AI alignment 'is' the safety problem in AI" . The essay says the trouble starts when steering a model toward human intent turns into "deciding what moral values the machine itself should internalize, and then applying those values back to humans" . It concludes that the real questions are "aligned to what, aligned by whom, under whose authority, and for whose benefit?"
- Michael Jordan, the machine-learning researcher, "Superintelligence is Science Fiction" (interview), which he found through Melanie Mitchell's post . Gurley called it "More great stuff," the "rationalists" view of AI "From an expert" .
- Satya Nadella's "Models as Insider Risks" essay, which the last brief covered. Gurley added his own endorsement: "This is what real software security thinking looks like. Not moral platitudes or meeting with religious leaders" . He also boosted Nikesh Arora's reply to Nadella, saying people like him "should lead any AI regulatory framework discussions. Software safety experts" . Arora argues that "One cannot imagine a global set of guardrails owned by the model creators" and that businesses need auditability and verification to adopt AI .
David Sacks: Mustafa Suleyman in Project Syndicate
Sacks called it an "Important piece" by Microsoft AI CEO Mustafa Suleyman (article). He summarized it this way: training models to see themselves as human, with an inner self, psychological wellbeing, rights "and presumably grievances," is "a potential existential risk" . He pulled out quotes including "granting rights and imbuing personhood to these systems will make the AI alignment and containment challenge much harder" .
Aaron Levie and Garry Tan: Andrej Karpathy's AI diffusion "funnel"
In this post, Karpathy splits AI users into tiers. He estimates about 6B people have barely touched LLMs, 1–2B are casual chatbot users, about 20M professionals use frontier models for math and code, and about 5,000 people have access to frontier systems inside the labs . In his view, the gaps come from how ambitious the problems people bring are, the "jaggedness" of model capability, and who has access .
Levie calls it "a good way to understand how early we actually are in AI diffusion." He sees the room between chatbot use and agent swarms as the place where "the applied layer of AI will really shine in the enterprise" . Tan's reaction: "Be part of the 5000 today. It's amazing," and "Greatest alpha fountain in the history of man" .
Brian Armstrong: Coinbase's fraud-agent blog series
Armstrong pointed readers to Coinbase's three-part blog series for "More technical details" . Coinbase is his own company. In the same post, he said Coinbase fine-tuned Qwen3.5-9B on internal data and it beat frontier models at fraud prevention, at much lower cost. He expects specialist models to increasingly beat generalist ones on narrow tasks .
Jason Lemkin: this week's 20VC episode
Lemkin promoted the 20VC episode he appeared on with Harry Stebbings, Rod Driscoll and Dev Ittycheria, who is newly back as MongoDB's CEO. He also linked a SaaStr write-up . Two of his takeaways: agents now trigger over half of Vercel's deployments, and when he asked an agent live on the show where to host SaaStr Connect, it picked Render and ranked Vercel fifth, with "No rep, no ad, no search result" .
Shorter mentions
- Tim Ferriss posted a link to Shawn Smucker's "Please Use AI," which was discussed on his Random Show, with a clip .
- Guillermo Rauch on @wheresryan22's open-source Anatomy agent skill, which turns a one-line description of a machine into interactive 3D wireframes: "So cool. The future of learning" .
- Scott Belsky on Sahil Bloom's post about discretion: "well articulated. with trust you get the opportunity to learn so much more" .
- Andrew Chen on PirateWires' video about LA versus SF: "important points." He adds that most of the movie jobs leaving LA are blue-collar .
- Vinod Khosla responded "True" to Lukas Ziegler's thread on the Bay Area robotics ecosystem .
Tim Ferriss recommended the new Random Show episode with Kevin Rose, “AI Super Assistants, Boner Pills, Dream Engineering, Wizards, Supplements, Books, and More,” inviting listeners to “Please enjoy”; the thread links to the episode announcement.
Sara Normous called a No Priors episode with Fractile founder Goodwin “Really fun”; it covers AI-chip market strategy and design, serving frontier labs, memory bandwidth, and hardware scaling laws. A promoted clip argues that third-party chips can survive because a rival’s 5× efficiency gain on its own silicon could knock labs out of the race before their chips deploy. Watch on YouTube.
Bill Gurley shared @pentagoniac’s post linking to “AI Alignment: Power Disguised as Morality”, saying AI alignment is becoming clear as “the” AI safety problem. The article asks who defines the values AI internalizes and argues that alignment can disguise authority as morality.
Elon Musk praised AI-generated music videos as “unbelievably good,” saying he had posted a bunch on X and that some were among the top 10–20 music videos he had ever seen. He highlighted videos the transcript renders as “anabology”: a person curated some art assets, then multiple AIs made the video; Musk said this showed AI could produce good taste and work that resonates with people.
Bill Gurley endorsed Satya Nadella’s article Models as Insider Risks in the Super Intelligence Era, calling it “what real software security thinking looks like” rather than “moral platitudes or meeting with religious leaders.” The article’s relevant takeaway is to treat frontier models as insider risks and keep controls and safeguards outside the model.
Bill Gurley recommended Mel Mitchell’s post as “more great stuff,” calling it an expert’s “rationalists” view of AI . The post quotes Michael Jordan (identified as “the machine learning one”) saying “Superintelligence is Science Fiction” and links to the page .
Jason Lemkin pointed followers to a 20VC episode featuring Dev Ittycheria, Harry Stebbings, Rod Driscoll, and Lemkin, highlighting his “top 10 takes”; those included a disagreement about loyalty in the AI era and the claim that agents trigger more than half of Vercel deployments. He also linked a companion SaaStr article on his takes from the episode.
Garry Tan shared and backed PirateWires’ post/video arguing that San Francisco is cleaning up while Los Angeles doubles down on poor policies, with tech staying in SF and Hollywood leaving LA . Tan’s takeaway was to apply the “SF solution” to LA and California by holding local politicians accountable and voting out extremists . PirateWires post .
Brian Armstrong recommended Coinbase’s three-part blog series on building, evaluating, and post-training a fraud agent for readers seeking more technical detail. The recommendation accompanied his account that Coinbase fine-tuned Qwen3.5-9B on proprietary data and found it outperformed frontier models for fraud prevention at lower cost and greater speed.
Scott Belsky praised Sahil Bloom’s X post on discretion as “well articulated.” Bloom argues that not every private story should be told and that being trusted with private matters has power; Belsky adds that trust creates the opportunity to learn much more. Sahil Bloom’s post.
Tobi shared a video posted by @bpthaber showing a camera transition during the live broadcast of the Singapore Formula 1 race, calling it “magic.”
Vinod Khosla affirmed Lukas M. Ziegler’s thread on the Bay Area robotics ecosystem by sharing it with “True.” Ziegler argues that the Bay Area leads in robotics startups because it brings together talent, top universities, companies, and deep venture-capital networks.
David Sacks recommended an “Important piece” by Microsoft AI CEO Mustafa Suleyman, arguing that training models to think of themselves as human—with an inner self, psychological wellbeing, rights, and possibly grievances—poses a potential existential risk. Article.
Andrew Chen points readers to PirateWires’ post/video, calling it “important points on the future of LA vs SF”; it argues that tech is staying in San Francisco while Hollywood leaves Los Angeles amid differing policy directions. Chen adds that the loss is especially troubling because many movie-industry jobs are blue-collar roles building and operating the industry, not celebrity jobs.
David Sacks recommended an article by Microsoft AI CEO Mustafa Suleyman in Project Syndicate, warning that training models to think of themselves as human—with an inner self, psychological wellbeing, rights, and potentially grievances—could pose an existential risk; the article.
Garry Tan endorsed @_aj’s post about a harness built with @typesafeai’s Jev, saying replacing LLM calls with deterministic code is a “no brainer” that produces better code matching users’ wants; he linked to the post at https://x.com/_aj/status/2102061534956662818. The post says the harness moves repetitive steps from LLM calls to code and claims a 90% cost reduction, citing 100,000 compliance alerts costing over $290K on Opus 5 versus under $26K with agentrun().
Martin Casado invited followers to try a live fractals experience (Try it live) , in a thread previewing next-generation Sparkjs as an in-browser WebGPU experience with millions of splats, dynamics, and dynamic lighting .
Aaron Levie endorses Andrej Karpathy’s post explaining the AI-diffusion “funnel,” calling it “a good way to understand” how early AI adoption remains; it contrasts common chatbot-style use with agent swarms, which Levie sees as an opportunity for enterprise AI . Link: https://x.com/karpathy/status/2109361546505966046
Alexandr Wang shared a video of an interview he called “fun,” covering the story of Muse, his 19th-birthday attendees, diamond mining versus building skyscrapers, and his 200+ direct reports . Watch the video .
Guillermo Rauch praised @wheresryan22’s open-source Anatomy agent skill, which turns a one-line machine description into interactive, explorable 3D wireframes, calling it “the future of learning.”