FPV Ventures partner Nikunj Kothari
Nikunj Kothari offered two reflections on the startup environment: long-term commitment is harder to find, and defensibility is harder to reduce to one factor. He sees founders and executives moving toward obvious winners and worries about the companies left behind. Application-layer advantages often come from thousands of details across product, technology, and GTM, many of them invisible. Ultimately, founders’ execution speed and adaptability remain important. He also suggested that B2B founders aggregate trending X topics for LinkedIn, which he believes often lags by a week or two, reducing reliance on expensive CPMs. He and Lola invited people navigating transitions to a breakfast conversation.
Box CEO Aaron Levie
Aaron Levie believes shared standards and practices can address AI safety and security for the foreseeable future. He supports rapid industry collaboration. As capabilities advance, greater oversight, testing, liability, and regulation will eventually emerge. At the current stage, however, he thinks companies can still align collectively. He particularly values supporting technological progress while preserving competition.
Builder Zara Zhang
Zara Zhang made a marketing website for her water bottle. She said Opus 5.5 handled the whole process in six prompts. It used ElevenLabs to generate music. Meshy produced the 3D assets, and the OpenAI API generated images. The demo shows a web project combining several generative tools.
Sam Altman
Sam Altman announced Dots, positioned as AI that works around the clock to give users back time and attention. He also introduced 6.1 Sol at one-fifth of Astra’s price, with a 95% cache-read discount. He spoke positively about its capabilities. In a separate usage impression, he said Ultrafast was so fast that he did not want to return to the previous experience. The announcements center on persistent work, price, and responsiveness.
Thibault Sottiaux, Codex and ChatGPT at OpenAI
Thibault Sottiaux expects millions of dots to come online within days. He wants to learn what different users enjoy and what still falls short. In his own experience, teaching a dot preferences and current concerns produced a noticeable improvement after two or three days. He said it learns quickly and can independently take on ambitious tasks. The team plans to learn from primary-dot usage before releasing the ability to create a full team of dots.
Vercel CEO Guillermo Rauch
Guillermo Rauch announced that AI SDK reached 30 million weekly downloads. He used fframes, fx, and Opus to make a milestone video, asking fx to examine the data and highlight 5, 10, 20, and 30 million downloads. The video was written in Rust and rendered on the GPU. Another chart post highlighted an inverse relationship between product defects and downloads. He presented product-growth data alongside programmable video creation.
Y Combinator CEO Garry Tan
Garry Tan merged a version of OpenClaw’s test-audit skill into GStack. He explicitly called test bloat a real problem. The reuse focuses on test auditing. It brings an existing skill from another project into his toolkit. Garry sees readily available model intelligence as a way to address such problems.
Official blog · Claude in Chrome is generally available
Claude in Chrome is generally available on every paid Claude plan. It uses existing browser logins to read pages, click, type, and fill forms, reaching internal dashboards, legacy systems, and vendor portals without connectors. A new mechanism automatically approves actions judged safe, and users can disable it in settings. Enterprise administrators can restrict allowed domains. Local files and other applications still require the desktop app; other Chromium browsers and mobile are not yet supported.
Safeguards combine model training, injection probes on tool results, and pre-execution checks that actions match the user’s request. Anthropic notes: “Prompt injection remains a moving target.” On a newer evaluation with stronger attacks, success rates without additional safeguards were 17.6% for Opus 4.5 and 3.8% for Opus 5. With probes and safety classifiers, no attacks succeeded against Sonnet 5, Opus 5, or Mythos 5, while Fable 5 saw 0.3%. These are results under the reported evaluation conditions. The older evaluation was retired after tested models reached 0%, so the two evaluations should not be treated as one comparison.