OpenAI’s Thibault Sottiaux
OpenAI’s Thibault Sottiaux says the company has assembled a cyber team and plans to share more of its work over the coming weeks and months. He separately invites people interested in security work to contact Clint Gibler, without specifying roles, locations or requirements in the post.
Peter Steinberger of OpenClaw and OpenAI
Peter Steinberger of OpenClaw and OpenAI questions whether individual usage limits still make sense. As capabilities improve, he argues, people want an agent representing a project, running loops and evolving its codebase while humans curate the automations. This raises a project-level agent and quota question rather than announcing a new billing or usage-limit policy.
FPV Ventures partner Nikunj Kothari
FPV Ventures partner Nikunj Kothari recaps an event with Peter Walker and other guests and aims to hold it weekly. He seeks speakers who can discuss a field casually but deeply for 45 minutes, particularly people not yet household names whom he thinks could become widely known within 12 months. Suggested speakers can be sent to him by direct message.
Vercel CEO Guillermo Rauch
Guillermo Rauch reports that Vercel Security Reviews identified 6,817 vulnerabilities in 24 hours, including 580 critical or high findings, with 1,411 already patched by an agent. He says the service is entirely free until November 10. Separately, he argues that much human-written software is slow, vulnerable, underspecified or incorrect and predicts AI will rewrite it repeatedly. He expects proof, formal verification, load testing and hardening, including replacement of memory-unsafe foundations with more reliable, optimized ones. He also points to documentation, teaching, diagrams, animations and slide decks; the vulnerability figures are his operational report, while the broader engineering claims describe his expectations.
FirstMark investor Matt Turck
FirstMark investor Matt Turck describes the unified AI he wants as a customer: one central brain accessible by text or speech, for purchasing or support. It should take useful agentic action as well as answer, and be available around the clock. He presents the combination of Dialpad and Ada as a new entrant pursuing that vision. He discloses being an investor since the Series A and expresses enthusiasm for Mimurchison and the team. This is his desired customer experience and investment perspective, rather than confirmation that the full vision has been delivered.
Y Combinator CEO Garry Tan
Y Combinator CEO Garry Tan says GStack’s /review skill remains useful for preventing bugs. The brief tool assessment supplies no defect counts, comparison test or detailed workflow.
AI Engineer organizer Swyx
Swyx highlights Thinking Machines’ inkling base model, claiming 3x speed and 5x lower cost with the same FrontierCode 2 evaluation results. He calls it distillation without tradeoffs and considers the model underrated. He expects more such work, particularly on US-pretrained models. The post does not detail benchmark settings, so the figures and forecast remain his claims.
Designer Ryo Lu
Ryo Lu, whose bio lists design work at Cursor, Notion and Stripe, reflects on an AI Engineer Shanghai talk: when we can make anything, do we still know what we want? He wants tools that help people find their own answers rather than skip the process of forming them. His goal is not only more output but more room to think, care and make things together, with further thoughts to come.
AI tutorial author Peter Yang
AI tutorial author Peter Yang says he does not use AI to call local businesses but is willing to use it for Xfinity customer support. He complains that the phone system prevents reaching a person when an address has an active outage. This is his service experience and usage preference, without new technical details about a calling agent.
Builder Zara Zhang
Builder Zara Zhang notices that most people sharing personal websites in a thread are developers or designers. She asks whether only those groups need websites or only those groups know how to make them. Her question separates demand from the ability to build, with the observation limited to that thread.
Replit CEO Amjad Masad
Replit CEO Amjad Masad announces a partnership with Japan’s Sumitomo Mitsui Finance & Leasing, or SMFL. He says Replit powers SMFL’s core software products and has accelerated development. As an example, he claims one person completed a project in one week that previously required six people for six months. The partners plan to extend AI-powered software creation to Japanese enterprises, addressing labor shortages, innovation and new business opportunities. The comparison is his customer-case account, without a project scope or acceptance methodology in the post.
OpenAI Core Products product leader Cat Wu
Cat Wu announces that she is joining OpenAI to lead product management for Core Products. She credits OpenAI with demonstrating scaling, bringing AI to more than a billion people and treating strong products as central to its mission. She plans to work with Thibault Sottiaux, the PM team and colleagues across the company on agents that support individuals and teams while keeping users in control. She hopes to free time for the people and work users care about and invites feedback.
AI & I by Every — The Inside Story of the 50 Who Built Modern AI With Kevin Roose
Key takeaway: AI can handle retrieval and line editing, while the author still decides why a story matters and which struggles must remain personal.
Tech journalist Kevin Roose interviewed about 150 people connected to the industry for The AGI Chronicles, seeking to record decisions behind modern AI. Despite writing an AI book in 2021, he still wanted to understand how the transformer moved from Google to OpenAI, the choices behind releasing ChatGPT and the origins of reasoning models. Confidential records and disappearing chats made preserving that history urgent to him.
Deep access required trust. The first roughly 25 interviews were hardest. Most sources could speak on background, without their names being directly associated with particular information. Recommendations among acquaintances and people wanting to supply a different side of a story then helped the reporting snowball. He emphasizes reading papers and earlier interviews first and arriving with specific questions rather than asking for a surface overview.
He wrote the book’s words himself and used AI elsewhere. Through Claude Code, he organized a council of Claudes with perspectives such as a skeptic, a technical historian and an editor focused on narrative craft, then had a master editor synthesize their responses. He says this caught factual errors missed by him and his human editors: “A couple years ago, you know, we had to fact check the AIs, and now the AIs fact check me.” This describes his writing experience rather than establishing that model-assisted checks need no further verification.
NotebookLM helped trace information across a sprawling archive, such as who told a story about GPT-4 pretraining or which three people to contact about an event at Anthropic. AI helps with line edits, structure and pacing. He still needs human editors to judge whether a story is worth pursuing, encourage another visit to sources and recognize when an idea is not yet resolved. He mainly writes in Google Docs to collaborate with producers.
Roose distinguishes weightlifting from forklifting. In the former, struggling is part of learning; in the latter, the aim is simply to move a heavy object from A to B. He keeps parenting personal while delegating startup paperwork, insurance and administrative chores to tools. He is also exploring video that maintains journalistic standards while fitting YouTube viewing habits and still serving listeners with the screen off. His regulatory and industry-risk views remain judgments. More directly applicable to creators are retaining responsibility for writing, organizing traceable research and deciding what deserves personal practice.