Anthropic Claude team member Boris Cherny
Boris Cherny says Claude Slides has become his way of making slide decks. His concrete example is no longer having to search for an iPhone-frame SVG to assemble a slide. On the coding side, Claude Code Review runs on every one of his PRs. He says it catches almost every bug before merge and credits it with helping the team move faster while improving quality. This is his assessment of his own workflow; the post provides no independent test data.
Vercel CEO Guillermo Rauch
Guillermo Rauch shares a performance claim for TypeScript compiled with scriptc. He says it outperforms tsgo by 2–4 times. The post does not specify the test environment or workload, so the multiplier remains the result he reports. Separately, he rebuilt the Roadster demonstration as a playable simulation. He says it retains many details of the real-world demonstration and provides a link to try it.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari looks back on the path from a thesis to an investment. He wrote a prediction about a sector seven months ago and subsequently met roughly 50 startups in it. He has now found a company whose Series A his team has the opportunity to lead, although the post does not name it. He uses the experience to encourage people to take a shot and let opportunities emerge through action. In a separate joking post, he describes founders evading questions about how revenue is actually calculated, drawing attention to the definition of the number.
Y Combinator CEO Garry Tan
Garry Tan recommends a hosted entry point for people who want to try GBrain but do not know where to start. He describes it as an easier way to get going and includes a link. When starting a new project, he also highlights using Capy with AsideAI. In his account, agents can code the first version and help set up DevOps in the same pass. What he values is combining initial development and operational setup in one workflow.
Replit CEO Amjad Masad
Amjad Masad says Replit can be deployed into a customer’s own cloud so that data stays in the customer’s environment. He positions this deployment option for high-stakes settings such as government, finance and healthcare. The post stresses customer ownership of data without detailing the technical architecture. In another update, he compares mathematicians’ reactions to change to emotional stages running from denial to acceptance. He thinks this transition is happening in a compressed period and places the current moment near acceptance.
Every CEO Dan Shipper
Dan Shipper shares a team demonstration in which Lee Knowlton turned one of his Codex articles into a visual novel. It is a concrete example of converting existing text into another kind of experience. He also argues that generation is now inexpensive and high quality. In his view, the next major capability improvement needs to come in editing and oversight. His emphasis is on modifying and checking what has been generated.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux shares ChatGPT Space material featuring the actual product team and asks what users would like to see next. He describes it as occasionally awkward but endearing precisely because the real team is involved. In the same set of updates, he says slides support was added the previous day. It is currently available in Pro, Business and Enterprise. The team is using this availability to learn from feedback.
Anthropic’s official Claude account
Claude’s official account shares a medical-education example: a cardiology fellow built a 3D tool to help trainees learn to read coronary angiograms. The post provides no clinical validation results. In a separate demonstration, Haiku 5.5 wrote twenty haiku and then drew every frame in JavaScript.
SPC general partner Aditya Agarwal
Aditya Agarwal announces that Sriram Krishnan will join SPC the following week for a fireside chat and Q&A. He highlights work on Azure, scaling Meta’s mobile advertising business and leading a16z’s international expansion. Sriram subsequently went to the White House and helped write the U.S. AI Action Plan. The event brings together these technology, business and policy experiences through one guest. Agarwal also provides an RSVP link; the post does not give an exact date, time or street address.
Box CEO Aaron Levie
Aaron Levie argues that AI is approaching a practical threshold at which it can train models better than humans, with early signs of recursive self-improvement emerging. He points to stronger models, better execution frameworks and infrastructure for long-running agents, and growing data supplies. Multiple agents can also test far more tasks, experiments and approaches in parallel than one person could. He expects these factors to reinforce one another and accelerate model progress. Even if they do not directly increase its pace, he thinks they could rapidly raise the frontier’s ceiling. He cautions against ignoring this potential change; it remains his forecast.
No Priors — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
**Key takeaway: Greg Brockman argues that as software generation gets cheaper, verification, data and security may become the real bottlenecks.** OpenAI co-founder and president Greg Brockman has a mathematics background. He connects AI research, drug development and continuously running agents by asking where the full workflow remains constrained after capabilities improve. First, scientific breakthroughs need to be understandable to people. Brockman says AI has helped resolve nearly 400 important open mathematics problems, while not yet producing entirely new theories. He values insights inside proofs rather than handing people answers nobody understands. Having another AI identify evidence or flaws is one mechanism he discusses for human oversight. Second, mathematical capability does not automatically remove biological data shortages. He considers drug discovery a priority, but says each step must be examined, including data, wet labs, simulation and regulatory filings. AI can help draft submissions; missing experimental data must still be obtained. He explicitly says neural networks have not reached the point of replacing clinical trials, and long-term outcomes cannot simply be established through fast simulation. Third, an always-working form changes how people use AI. He describes Dots integrating email, Slack, calendars and project context to assemble unfinished commitments overnight and flag scheduling conflicts. Differences between consumer and enterprise assistants partly reflect affordable compute budgets. Enterprises also have administrators, security teams and controls over feature rollout. Interfaces can be generated on demand, but users still need to understand permissions and unavailable functionality. Fourth, security investment may exceed software investment. He says, “Security spend should actually vastly outstrip software spend.” His reasoning is that reliability becomes scarcer as generating software becomes cheaper. He describes OpenAI pausing the projects of about 25 production engineers to use its strongest model at the time to find and fix infrastructure issues until the model could find little more. The process must be repeated with the next model. Formal verification and generating secure code from the outset are directions he hopes to advance. Brockman calls the present a window for defenders and urges them to pay down security debt before frontier capabilities fully diffuse. His final requirement is to preserve human judgment, values and control so stronger machines serve the outcomes people want.