OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux said that over the next 28 days, the team will either ship one clear improvement each day for most Codex and Work users or make a full reset. He made the bar specific: a change should matter to most users and deliver a clear improvement. The post sets a cadence for the next 28 days but does not list the features planned for each day. It also does not explain what a full reset means. Sottiaux closed with “Let the improvements begin,” making this a commitment whose delivery can be checked day by day.
AI educator Peter Yang
Peter Yang said ChatGPT’s product structure has become complicated and named three areas he would simplify: Work versus Codex, Spaces versus Pages versus Sites, and the model and effort picker. He called the launch of Work as a standalone brand a mistake and compared it with Claude folding Cowork back into Chat. He said he is looking forward to a better primary AI interface. In another post, he relayed Granola co-founder Sam’s response to users bypassing the product UI and using MCP instead: as a UI designer, Sam feels some sadness, but accepts that large enterprises will inevitably rely on internal tools. A product that does not work with those tools can be a non-starter for enterprise use. Sam said the team has accepted that, in some workflows, Granola can capture context for use through an internal agent. Yang also shared Sam’s product-building advice: shipping version one is not enough; teams should check whether users get value from it every day.
Vercel CEO Guillermo Rauch
Guillermo Rauch said that as models get faster, the overhead of an agent harness matters more. He said the next release will improve session storage and retrieval performance and bring cloud durability to libfx. In another post, he revisited Vercel’s migration of Turborepo from Go to Rust: the migration is complete, but its return on investment was debated internally. Rust offered better low-level operating-system access, while Go was fast, well designed, and easy to iterate on, making the human cost of migration substantial. He believes that as agents write more code, the balance between what is best for people and what is best for the business is changing; he also does not think Rust is the final toolchain, since it predates the agent wave. Rauch also distinguishes between a README written by hand for people and internal documentation written in AI English for agents. He sees blogs, tweets, and READMEs as storytelling that connects people, beyond questions of formatting.
Box CEO Aaron Levie
Aaron Levie sees AI creating new jobs around enterprise deployment. The demand includes AI engineers building enterprise applications, forward-deployed engineers bringing agents into companies, and new firms providing deployment services. He thinks published job statistics may undercount transitions in existing roles, as data, research, and software teams move toward AI work. Banks, life-sciences companies, manufacturers, and law firms are hiring or repositioning technical staff to support agent deployment. Levie says it is easier to see new work once it is actually happening than to predict it by focusing only on what AI might replace. He says these changes are now becoming visible.
Matt Turck, investor at FirstMark Capital
Matt Turck pointed to an aspect of AI that remains poorly understood outside the field: models are “grown” rather than built like conventional software. He added that people do not really know how models work. Because their formation remains difficult to explain, he sees interpretability research as urgent. The post names Eric Ho and GoodfireAI but does not describe their specific work. It directly links limited understanding of how models form to the need for interpretability research.
Nikunj Kothari, partner at FPV Ventures
Nikunj Kothari argues that a Zoom call, or even a meeting with only the CEO, can miss important signals when evaluating an investment. He prefers visiting a founder’s office, where he can observe team energy, how colleagues work together, how co-founders complement one another, and what the beta product and collaboration look like. He acknowledges that investment decisions are often time constrained, but still asks to meet founders at their offices for final meetings. He says many founders tell him he is the only investor who has asked. He also questions bringing founders into an investor’s conference room to repeat the same pitch deck.
Dan Shipper, CEO of Every
Dan Shipper says dot has become his primary AI interface over the past few weeks, and he named his dot Boo. He guesses that a year from now he will still interact with ChatGPT through an always-on persistent agent, but the current dot may have disappeared. He compares this with the early days of OpenClaw, when users liked the small personality quirks of their Claws. Those quirks faded in importance once a functionally stronger option appeared. His point is that agent use may continue even if dot’s current personalized identity does not.