AI engineering builder Swyx
Swyx recalls pitching something like Claude Code in 2019 and then not building it for seven years. He uses the experience to poke fun at missed-opportunity stories, pointing to the gap between an idea and execution without detailing the proposal or why he did not pursue it.
Practical AI tutorial and interview creator Peter Yang
Peter Yang lists four hiring practices he wants changed: filtering by years of experience or company names, favoring résumés and cover letters over proof of work, treating the number of people managed as an accomplishment, and choosing endless interviews over short work trials. He favors real projects that demonstrate what a candidate has done. For management, he rejects headcount alone as evidence of contribution. For assessment, he suggests short trials without prescribing a universal length or implementation. In another post, he criticizes videos built around invented ARR, or annual recurring revenue, saying that creators making up figures might as well identify themselves as satirical media. Across the two posts, his emphasis is on connecting displayed credentials or numbers to work actually done.
Former Claude Code team member Thariq
Thariq proposes a new way to distribute open-source software so users can customize it through agents while automatically merging upstream changes. He thinks coding agents can perform merges but do not yet complete that process automatically. His own experiment is an AI-native macOS Dock with writing, building and other modes to help him focus. An agent can customize mode behavior and add modes. The project uses Jev to classify activity and notifications in real time and act on them. He says this is the kind of human–computer interaction work he has wanted to pursue, but changing operating-system internals has been time-consuming. With many edge cases remaining, he is not sharing it yet and does not recommend using it; it is a personal experiment.
Replit CEO Amjad Masad
Amjad Masad says that if he were investing in frontier AI, he would underwrite it mostly on new discoveries and their applications in areas such as medicine. His rationale is that sub-frontier models are rapidly saturating coding and much knowledge work. This remains his investment and capability-tiering view; the post supplies neither model comparisons nor medical outcomes.
Vercel CEO Guillermo Rauch
Guillermo Rauch shows a Dockerfile example running an HTTP service with Bun. It starts from the oven/bun:alpine image, uses Bun.serve on port 80 and returns “we are so back.” He first aliases now to vercel, then runs now. The post ends with Ready and an access link, showing a short path from container configuration to a deployment command. It does not list plans, resource limits or a more complete container setup.
Box CEO Aaron Levie
Aaron Levie argues that most enterprise workflows have not meaningfully adopted agents and estimates current deployment across the economy at less than 1%. He distinguishes a single LLM pass that makes a decision or supplies feedback from a workflow that actually runs automatically. He attributes the gap partly to enterprise adoption rates but mainly to AI capability. With progress in computer use, long-running agents, harnesses and supporting infrastructure, he thinks more complex workflows are becoming feasible. He envisions systems improving execution rates over many instances of a workflow goal, with agent swarms applying different amounts of compute. The emphasis is on organizations redesigning underlying processes, rather than individual end users prompting each task. He sees opportunities in finance, life sciences, healthcare and other information-heavy fields; the less-than-1% figure and subsequent changes remain his estimate and expectations.
Y Combinator CEO Garry Tan
Garry Tan says he is slimming down GStack to reduce token use and task latency. Another post shows how his GitHub contribution graph changed after using the Capy software factory. He says Capy Captains keeps 10–20 subagents running on GStack and GBrain, producing a heavy flow of pull requests. He presents the graph as a personal illustration of development intensity. Together, the posts show his interest in expanding parallel execution while reducing consumption and waiting; the posts do not equate pull-request counts with accepted deliveries.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck challenges the common advice that too many co-founders is necessarily bad. He lists seven co-founders for Anthropic, eleven for OpenAI and seven for Databricks. By placing these examples alongside familiar concerns about dilution, founder friction and analysis paralysis, he questions judging a startup team by the count alone.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari also cites seven co-founders each at Anthropic and Databricks to reject a fixed “right way” to identify or build startups. He argues that the best companies are outliers and that the approach that works for a particular team matters. Another post points out that familiar industry terms can receive wildly different definitions. His eight examples are long horizon, out of distribution, especially data, intelligence per watt, RL environments, “forward deployed,” revenue, “frontier” and self-improving. He acknowledges that some are properly subjective but is still surprised by the variation. Even when he has a reasonable grasp, he needs to ask more deeply what the other person means. Both posts challenge treating simplified labels as sufficient grounds for judgment, rather than propose new universal definitions.
Every CEO Dan Shipper
Dan Shipper predicts that video creation and game making will become widely distributed foundations of knowledge work, much as document and slide creation are today. His emphasis is on these forms moving from specialist work toward everyday expressive tools for more knowledge workers. The post supplies no timetable or specific tool path.