We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Casado: a16z's conversation with Tutor Intelligence
Martin Casado promised "you'll learn something from this." He called it "one of the more in depth discussions I've ever had on end to end economics of robotics," including why robotics companies tend to verticalize and how hard the long tail is . In the a16z episode, Casado and Raghu Raghuram talk with Tutor co-founder and CEO Josh Gruenstein. Topics include going from a first sales call to a working robot in about three months, and why so many robot startups died building for a single customer . Tutor charges by the hour with no upfront cost. When the AI gets stuck, a human "tutor" takes over, which creates training data . YouTube
Tan, Lütke, Casado: "Weight-Space Programming"
Three investors and founders shared Rikiya Takehi's post. In it, GPT-6 Astra edited Qwen3.5-0.8B's weights directly. It changed 0.016% of them and, according to the post, improved math up to 6x and instruction following 3x, with no gradients or training . Takehi calls this "Weight-Space Programming" (blog) .
- Garry Tan called it an "Interesting new use of smaller models" .
- Tobi Lütke: "this is exactly as crazy as it seems" .
- Casado shared Alex Krentsel's reproduction of the results. Krentsel calls it a step-change capability that has appeared only in the latest models . Casado asked whether it "must be an artifact of efforts to do RSI within the labs" .
Masad: Replit's security advisory on AI agents
Amjad Masad promoted @lhchavez's "Security advisory: Make managed cloud environments the default for AI agents." His framing: agents running on employee machines are becoming "a systemic security risk." He noted that Goldman Sachs has reportedly restricted tools like Claude Code . The advisory points to supply-chain attacks and prompt injection hidden in GitHub issues or emails. It argues that agents' broad privileges make them attractive targets . Replit's recommendation is to run agents by default in managed, isolated cloud environments. Credentials, permissions, network access and production changes would stay outside the agent's reach .
Modal's CEO: Seven Powers and Ben Thompson's "gingerbread man" post
On 20VC, Modal co-founder and CEO Eric called Hamilton Helmer's Seven Powers "a good book" for thinking about moats. He warned that founders read it and "get delusional," when "you never have moats in the early days" . The post he says he "always come[s] back to" is a Ben Thompson piece on Snap's "gingerbread man" strategy. Snap had no moat, so it just ran faster, and he thinks that holds "for almost every startup these days" . Episode
Shaan Puri: Die with Zero
Shaan Puri posted notes from time with Bill Perkins, author of Die with Zero. He admits he missed the book's main point the first time: it is about dying with zero regrets, not zero dollars . He calls it a "Top 5 book title," arguing that the same ideas sold as "live life to the fullest" would be forgettable . Full episode
Alexandr Wang: Wildwood
Alexandr Wang called Wildwood "a true marvel of human artistry, craftsmanship, and storytelling." He said he liked it even more than You Can See Everything . He noticed that it, like Hoppers, treats death as natural and inevitable . He also pointed to a Buddhist/Taoist theme of non-intervention: trying to impose our own ideas of goodness on the world can do harm . "Huge fan, go see it!" Post
Shorter picks
- Sarah Guo shared Colossus's Michael Kratsios profile, calling it a look at "important behind the scenes work" by "a civil servant who understands technology" . Harry Lambert wrote the piece . Patrick O'Shaughnessy had already shared it.
- Tobi Lütke shared MLStreetTalk's clip "What if there is no hard problem?" with Joscha Bach (@Plinz) . He called interviews with Joscha "the most fascinating podcasts" and Bach "one of the few modern real philosophers" .
- Vinod Khosla shared "Going founder mode on cancer," about GitLab co-founder Sid Sijbrandij (@sytses). He calls it "An extraordinary story of agency" .
- Michael Seibel said he wanted to clip a podcast where Sequoia's Doug Leone called finding product-market fit "a dark art" that investors can't help with, and put it in YC's Bookface and user manual. His point: product-market fit is "your responsibility" .
- Vinod Khosla shared a Khosla Ventures post/video featuring Affirm founder and CEO Max Levchin, who explains why he made AI adoption non-negotiable at Affirm; Khosla frames AI-native work as changing how people work and what one person can accomplish. Link
- Khosla shared a post/video featuring Opendoor CEO @nejatian on leading by example: the CEO should be the company’s most AI-native person, because culture follows leaders’ actions. Link
- Khosla shared a post/video featuring Eight Sleep co-founder and CEO @m_franceschetti on AI multiplying people: employees can become “micro-CEOs,” managing teams of agents and owning more output. Link
- Khosla shared a post/video featuring Vercel founder and CEO @rauchg on software discovery shifting from people searching to agents deciding, making developer experience into agent experience. Link
- Khosla shared a post/video featuring ClickHouse CEO Aaron Katz on rebuilding ClickHouse for an agent-first world: agents, rather than people, may choose the next database, changing product design, onboarding and distribution. Link
Jason Lemkin linked to a SaaStr video post featuring his comments on API costs . His accompanying takeaway was that charging agents for every API call could reduce usage and lead agents to route around the provider, so providers should consider that before charging for agent access .
Michael Seibel recommended preserving a clip from a podcast featuring Doug Leone of Sequoia, who said finding product-market fit is a “dark art” investors cannot solve; Seibel wanted to put the clip in Bookface and the user manual because it candidly sets founders’ expectations and challenges investors who claim they can deliver product-market fit.
Sarah Guo praised Humans& for “continuing to share their cool ideas” and linked its post on weight teleportation. The post presents faster trainer-to-sampler weight updates and says the technique is open-sourced.
Sarah Guo recommended Colossus’s profile of Michael Kratsios, describing it as covering important behind-the-scenes work by a civil servant who understands technology . The feature is by Harry Lambert and examines how the fight over AI regulation and U.S. science reform runs through Kratsios’s office (article ).
Tobi praised interviews with Joscha as producing “the most fascinating podcasts” and called him “one of the few modern real philosophers.” He linked an MLStreetTalk video post titled “What if there is no hard problem?” that tags @Plinz.
Vinod Khosla endorses Khosla Ventures’ “Going founder mode on cancer”, highlighting GitLab co-founder and executive chair @sytses’s approach to cancer—questioning assumptions, running experiments, and continuing to search—as an “extraordinary story of agency.”
- Modal co-founder and CEO Eric recommends Seven Powers as a good book for thinking about moats, while cautioning that founders can overstate their startups’ moats, especially early on.
- Eric says he keeps returning to Ben Thompson’s blog post about Snap’s “gingerbread man” strategy: Snap lacked a moat and instead moved faster, a lesson Eric says applies to nearly every startup today.
Tim Ferriss shared “How to Build a Strong Neck”, a video attributed to Dr. Andrew Huberman (@hubermanlab).
@tobi endorsed @rikiyatakehi’s post about “Weight-Space Programming,” replying, “Yea. this is exactly as crazy as it seems,” and linking the post . The post links to a blog describing frontier LLMs directly editing neural-network weights : https://rikiyat.github.io/blog/weight-space-programming/.
Scott Belsky called Howard’s post—“I select people for their strengths, not their lack of weaknesses”—an “important insight,” adding that everyone is a package deal and most outlier strengths have a corresponding weakness. Howard’s post.
Amjad Masad shared @lhchavez’s Security advisory: Make managed cloud environments the default for AI agents, presenting it as outlining agent risks and a path forward. The advisory recommends making approved, managed cloud-agent environments the standard for autonomous development, citing the excessive access agents can inherit on employee workstations; it calls for verified isolation, task-scoped permissions, and independent containment. It cautions that cloud execution needs verified security controls and that the reviewed sources do not establish comparative breach rates for local versus cloud agents.
Chamath Palihapitiya shared @PhilosophyOfPhy’s video/post about the Sun carrying Earth and the planets around the Milky Way, tracing corkscrew paths through space; he reacted, “We are part of something so awe inspiring.”
Elon Musk shared @mattyp’s “What’s New in Grok Bot” video, captioning it “Grok improvements.” The demo’s listed topics include shopping cards and product pages, order tracking, payments, custom modes, Team Bot setup and discovery, Slack connector asks, and Vault contacts.
Elon Musk endorsed Jason Bud’s X post, which advises treating Grok Bot like a teammate and giving it an ambitious task normally assigned to a new hire; the post links to Scott Billows’s original post. Musk’s stated reaction was “This is the way.”
Jason Lemkin points readers to a SaaStr deep dive and conversation with Databricks’ CRO and Sam Blond; the post calls the CRO Rob Gabrisko, while the URL slug names him Ron Gabrisko. The shared takeaway is to stop selling slides and instead show customers a solution already built for them—or one that can be built quickly—using an FDE motion.
Garry Tan shared @rikiyatakehi’s post on direct manipulation of smaller-model weights, calling it an “interesting new use.” The post calls the approach “Weight-Space Programming” and links to its blog; it reports that GPT-6 Astra changed 0.016% of Qwen3.5-0.8B’s weights, improving math up to 6× and instruction following 3× without gradients or training.
Amjad Masad shared @lhchavez’s “Security advisory: Make managed cloud environments the default for AI agents” and stated that agents should run in dedicated, organization-controlled environments with access limited to each task’s needs. The advisory’s case for managed cloud environments is that organizations can isolate agent execution and control credentials, permissions, network access, and production changes.
@jasonlk shared his new SaaStr article, “Why You Don't Want Humans Doing Outbound Anymore”, arguing that humans stop reading the outbound messages they send at volume, while agents read every account each time. He illustrates the point with a rep pitching him using a 2024 sequence that had not been revisited.
Shaan Puri praises Bill Perkins’s book Die with Zero, calling its title one of his “Top 5” and explaining the book’s point as dying with zero regrets, not zero dollars. He also shares the full YouTube episode from his conversation with Perkins: watch the episode.
