Swyx, affiliated with SMOL AI, Cognition and Latent Space
Swyx finds Opus useful for writing, but says incidental disallowed words in a prompt can block the work. He wants a production tool rather than a system that passes judgment for him. He says that behavior cannot sit on the critical path of an AI media organization. Separately, he tried using Codex to make animated explainer videos and found the result inadequate. Both posts describe his experience; neither supplies a reproducible prompt or evaluation result.
Google vice president for Labs, Gemini App and AI Studio Josh Woodward
Josh Woodward says the Gemini App Trusted Tester community has passed a major milestone, without specifying its size. He reports strong interest in early features on Web, Android and iOS. He says people who previously signed up are fully included. New applicants can use the signup link in his follow-up post. He also quotes Sundar’s commitment to remaining at the frontier and invites people to help build.
OpenAI Codex and ChatGPT builder Thibault Sottiaux
Thibault Sottiaux announces support for 12-hour caching in the API. He says caches are retained indefinitely for signed-in Codex and Work users, so returning to an older conversation or resuming after several hours carries no penalty and is guaranteed to hit the cache. A separate update addresses reports of faster-than-expected usage consumption over the weekend. The team identified approximately 40,000 accounts whose applied usage limits did not match their selected plans. He says the issue has been corrected and affected users will be contacted individually and given credits. He apologizes for the mistake. These are his public statements about caching and compensation, rather than verification of any individual account.
Practical AI tutorial and interview creator Peter Yang
Peter Yang introduces Mercury Payroll through a paid partnership with Mercury. He describes putting himself on payroll for a one-person S-corp as a painful part of running the business. The product brings payroll into a Mercury account, with a team handling state-tax registration paperwork and live-balance checks flagging potential shortfalls before payday. The accompanying disclosure says Mercury is a fintech company rather than an FDIC-insured bank; banking services are provided by Choice Financial Group and Column N.A. Mercury provides no legal, tax, accounting or investment advice and does not act as an employer, PEO, EOR or plan fiduciary. AI-generated information may be incomplete or inaccurate, and users remain responsible for reviewing and confirming payroll actions before submission or approval.
Google AI tools and experiments team Google Labs
Google Labs uses a three-part series called Dreambeans Roulette to show how Dreambeans might shape a morning. Participants include product manager Sahana R., vice president of product Brian R. and his dog Nala. The sequence is simple: open Dreambeans, choose a story and act on it in the real world. The current installment is Part 1. The post says Dreambeans is available in the US to users aged 18 and over, with more regions to follow.
Vercel CEO Guillermo Rauch
Guillermo Rauch says the team wants to bring a genuinely open-source, runtime-agnostic durability solution to harnesses, following the approach used for AI SDK. Infrastructure behind Eve and Workflows is being extended to Pi, OpenCode and other harnesses, whether run locally, on Vercel or in a self-managed environment. He says Claude Code and Codex can also become durable and observable this way. However, neither is currently designed for embedding, which makes them less efficient and portable than Pi or Fx and adds a sandbox requirement. In another post, he imagines AI identifying the precise gap between what a person understands and does not, enabling progressive learning. Free education that helps people reach their potential is his vision, rather than a reported validation of educational outcomes.
Box CEO Aaron Levie
Aaron Levie sees autonomous systems dedicated to particular tasks or domains as an emerging AI product form. He considers this a major step beyond simple LLM wrappers: users can delegate more sophisticated work while retaining humans in the loop. The first change is better tool and computer use, giving agents granular access to external data and software to perform role-specific workflows and build missing software. The second is stronger model capability in law, biological research, code security and supply chains, although he acknowledges that intuition remains worse than people’s in many areas. The third is improving harnesses, sandboxes, file systems and data pipelines, enabling longer tasks and large-scale processing. He emphasizes the need for UX, implementation and change management to support human oversight and coordination. He points to IT automation, cyber operations, drug discovery, manufacturing and AI research, and expects systems of this kind across many economic roles.
Y Combinator president and CEO Garry Tan
Garry Tan says GStack is now Capy-aware and works with Captains, Capy’s software-factory feature. The short post does not provide configuration steps or a comparative evaluation.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck collects several developments among FirstMark-related companies, including Daytona’s Series B and Baba’s Series A. He says FirstMark led Daytona’s Series B after leading its Series A earlier in the year. He specifically mentions Ivan Burazin, Jukic Vedran and their team. On products, he highlights Gradium AI’s new small open-source on-device voice models, describing them in another post as a text-to-speech suite. He calls them on-device state of the art, without supplying benchmarks or comparison conditions in the posts. He also mentions a New York subway campaign by Synthesia and Rillet.
Product builder Zara Zhang
Zara Zhang argues that distribution may not be the greatest advantage of having an audience at the earliest stage of building a product. More important are testing ideas and angles to see what resonates and finding the first 100 early adopters. She sums up the judgment as signals over distribution.
OpenClaw and OpenAI builder Peter Steinberger
Peter Steinberger describes a division of work in which Astra plans and Sol 6.1 codes. He finds that asking Astra to run Codex CLI in worktrees works much better than relying directly on subagents. A second agent supervises and challenges Astra to keep it on track. This differs from having the same agent review a goal text itself. He wrote an OC plugin for this “smart goal” feature. Separately, he announces a conversation about agents with dok2001 at Cloudflare Connect the following day.
Every CEO Dan Shipper
Dan Shipper recommends Meghan O’Gieblyn’s Will and Attention for its clarity, rigor and self-questioning honesty. He highlights its treatment of relapse into alcoholism, how personal narratives can save people and later fail them, and whether people can trust their desires. His central reading concerns pragmatism: the author initially adopted faith, surrender and sobriety because they worked, rather than as absolute truths. That can create a gap between behavior and actual desires or beliefs, while reopening the question of whether sobriety is useful. He sees the book’s candid account of what drinking offers as an examination of rationalization rather than a celebration of alcohol. He also identifies a tension: returning to inviolable truths and treating abstinence as unconditional may still rest on their practical ability to prevent self-destruction. In a separate post, he reports an approval-policy deadlock: a tool says approval is needed, the policy is set to never, and it cannot explain how the user can change it. That is his usage report, without a full diagnostic record in the post.
SPC general partner and Bevel Health cofounder Aditya Agarwal
Aditya Agarwal compares the progression from inference-time scaling to subagents and agent swarms with CPU threads, multicore CPUs and multinode distributed systems. He finds the parallels striking and describes history as rhyming. The post offers an architectural analogy, rather than evidence that performance or coordination costs are equivalent.