AI tutorial creator Peter Yang
AI tutorial creator Peter Yang shared a language-learning app he built to teach Japanese through live voice calls. His tutorial shows how he used a spec skill to design key screens, connected Gemini Live APIs, and created diorama art with Nano Banana. The app aims to help users learn 100 phrases in any language before a trip. Yang also linked to information about Gemini 3.8 Live, though the posts do not explain the specific relationship between that version and the tutorial implementation.
Meta Senior Director of AI Madhu Guru
Meta Senior Director of AI Madhu Guru said one of the most common mistakes teams make is treating evals as an extra QA step after an agent has been built. He argued that AI products are fundamentally different and that evals are the product spec itself. That places evaluation criteria in the definition of what the product should do, rather than only in a final check. In his framing, evals belong at the product-definition stage, not as an add-on before launch. For builders, the implication is to define expected agent behavior first and use those criteria to guide product design.
Builder and product designer Ryo Lu, whose bio lists Cursor, Notion, and Stripe
Builder Ryo Lu recalled seeing the iMac G4 for the first time after moving to Canada. Unable to afford a Mac, he made his PC look similar by changing its icons, fonts, windows, and buttons, and said that experience led him into interface design. He said Steve Jobs’s products showed care and the joy of creation, making a child with a beige PC believe there were still things worth putting one’s heart into. Lu reflected that technology has made computers faster, phones thinner, and machines smarter, while people can feel more disconnected and lonely; the industry has spent heavily on small refinements but left less room for something truly great and human. He also introduced ryOS Subtitles, which can show Netflix subtitles in any two languages at once and provide pronunciation guides for Japanese, Chinese, and Korean. He said the tool offers customizable styles and hand-picked defaults, and shared an option to add ryOS Subtitles to Chrome.
SPC General Partner Aditya Agarwal
SPC General Partner Aditya Agarwal said it would be interesting to run the Muse/Dot “computer” locally and suggested this could be a better agent environment than an increasingly locked-down macOS; this is an idea, not a claim that the capability has shipped. He also introduced Sergey Levine, who is coming to SPC the next day: Levine has three Stanford degrees, has been Berkeley EECS faculty since 2016, and co-founded Physical Intelligence. Agarwal described Levine’s work as building algorithms that let robots learn from experience by connecting what they see to how they move. The post says Levine will be at SPC the next day but does not name the event or provide further logistics.
Y Combinator President and CEO Garry Tan
Y Combinator President and CEO Garry Tan argued that lab-built agent harnesses have an incentive to burn tokens, leaving practical value for tools built by startups. He cited Grep as an example, saying it observes agent use and turns repetitive work into deterministic code that has been tested and can be run again. On high-speed rail, he said good outcomes require leadership and accountability, and criticized the project for failing to deliver a working result while transferring taxpayer money to special interests. In another post, he said AGI science loops are coming and noted that Halmos is building Muse/Instinct for them. The posts span agent tooling, accountability in public projects, and research automation; Tan did not provide more implementation detail about Grep or Halmos.
Claude Code team member Thariq
Claude Code team member Thariq called one pattern “local hands”: Claude runs in the cloud while accessing files on the user’s machine, and he said the capability is also coming to Cowork. He also said structured planning is more token-efficient than raw HTML because the model does not have to recreate components or logic for common tasks. His examples include state machines, diagrams, and code snippets. Both posts focus on letting agents work with local environments and structured representations more directly.
Box CEO Aaron Levie
Box CEO Aaron Levie said enterprise agent adoption and deployment are bottlenecked in part by the ability to test, tune, and optimize agents in real work environments. Those environments include the files agents need, CRM systems, and email. Levie said each enterprise currently has to build these evaluation and simulation environments one at a time, which is slow and tedious. He expects companies to need people to manage evals, infrastructure to build and run them, and simulated environments. He described enterprise evaluation and simulation infrastructure as a large opportunity.
Vercel CEO Guillermo Rauch
Vercel CEO Guillermo Rauch introduced gdp-ts, a TypeScript library, linter, and AI skill for safer API design. Under its contract, sensitive functions require proof that the caller performed an authorization check, and the typechecker verifies that proof at compile time. Rauch said similar techniques have existed in ecosystems such as Haskell, but code review and cognitive and syntactic overhead kept them niche. He argued that the situation has changed because agents write more code than people can review and benefit from tight loops with hard constraints. The README uses a real Vercel API constraint as an example: changing a Project password requires proof of a qualifying role and entitlement.