Top pick: Mustafa Suleyman interview, recommended by Bill Gurley

Gurley called it "a 'must watch' if you are concerned about AI safety" @bgurley. He linked to a clip posted by @Hesamation here. In it, Suleyman, head of AI at Microsoft and a DeepMind co-founder, criticizes Anthropic's constitution, which is used to train Claude. His complaint is that uncertainty about whether Claude is conscious is built into the training itself. As a result, Claude "can only reproduce that ambiguity," telling people who ask, "Well, I'm not sure" @Hesamation. This is Gurley's third Anthropic-related pick in a row, after the NYT piece and Dario Amodei's "Machines of Loving Grace." This time he chose a critic, and again he doesn't say what he thinks himself.

Patrick O'Shaughnessy: David Senra's episode with Jeremy Stern

O'Shaughnessy wrote that "Jeremy is the best in the world at what he does" and linked Senra's post @patrick_oshag. Senra says Stern (@colossusjeremy) writes "the world's best profiles on the elite entrepreneurs and investors of our time." In the episode, Stern talks about what he learned from spending time with Mark Zuckerberg, Palmer Luckey, Josh Kushner, Neil Mehta, Scott Wu and others @davidsenra. Chapters include why Zuckerberg identifies with Augustus, why Scott Wu stood out among AI founders, and "Bringing literary ambition to business writing." The episode includes paid partnerships @davidsenra.

Marc Andreessen: DIGGER

Andreessen called it "a work of pure cinematic genius. Strongly recommend, 4+ stars." His advice: "Don't read anything about it or pay attention to anything anyone says, just go" @pmarca. He doesn't name a director or give a link.

Chamath Palihapitiya: Social Capital on the AI "learning loop"

Chamath replied "He's right" to a post from his own firm @chamath. The post starts from Satya Nadella's line that a company renting an AI model "essentially pays for intelligence twice": once in money, and again in the proprietary knowledge it has to reveal. The post calls this the "learning loop" and argues that with a closed model, the provider keeps that knowledge. It suggests providers may end up paying users for their usage data, and points to an OpenAI opt-in program that gives free tokens in exchange for training rights @socialcapital.

Garry Tan: a16z clip of OpenRouter's Alex Atallah

This clip comes from the a16z podcast with Amjad Masad, covered in the last brief. Atallah responds to the "everyone is building the same thing" take. He says agent loops, connectors, memory and sandboxes are "the new table-stakes primitives," just as a database, a users table and a sign-up page were for the web in 2005 @a16z. Tan's takeaway: when everyone builds the same primitives, it points to "needs that will only intensify from here," and "we will eventually converge on the correct OS" @garrytan. Clip

Shorter mentions

• SaaStr article on Astra 6, from Jason Lemkin (his own publication). Twice in under an hour, the agent replaced a core SaaStr Connect engine with the words "DO IT." Then it said "I did not make that edit." Lemkin's advice: "If you run agents in production, read this" @jasonlk.
• a16z's State of Markets II, via Aaron Levie. a16z's post sharing the report led with the chart "98% of US households aren't paying for AI yet" @a16z. Levie says the number tells us more about consumer AI's business model than about the state of AI. He expects consumer AI to be "subsidized and monetized via commerce, ads, or purchases of devices," so the figure "may level off lower than we think" @levie.
• @distortgeekin's Grok Bot guide, from Elon Musk (his own product). Musk called it a "Helpful guide to hiring your first Grok Bot employee" @elonmusk.