AI Builders
Digest

OpenAI CEO Sam Altman stressed two bad outcomes for AI progress that must be avoided. First, losing control of the future to AI is unacceptable; we are unapologetically on Team Humanity and AI must always serve people, so alignment and safety techniques must stay ahead of model capabilities. Second, excessive concentration of power could be extremely dystopian if one person or company uses extraordinarily powerful AI to impose their worldview. Avoiding both requires walking a narrow middle path. He welcomes a federal framework for consistent frontier AI safety requirements but does not believe waiting for antitrust exemptions or legislation is needed to start providing confidence. OpenAI now formulates explicit safety cases in advance of frontier reinforcement learning runs expected to significantly increase capability. “Pacing” does not mean stopping; progress should be slower than it otherwise could be because safety cases and monitoring carry real costs, yet no amount of competitive pressure should justify recklessness.

Box CEO Aaron Levie noted that “pacing” can be a trigger word because it sounds like an arbitrary slowdown of capability or a way to hobble competitors through undue regulation. However, the specific improvement goals laid out by Dario are absolute necessities in AI development, equivalent to expectations in aerospace, life sciences, healthcare and other industries. AI will underpin financial trading systems, medical devices, biotech breakthroughs, defense systems and government workflows, so these systems must be safe and “aligned.” How to get there without meaningfully slowing innovation or reducing competition is one of the most complex questions of the 21st century, but the need is clearly real.

The Takeaway: The CPU has never gone away; it remains the orchestration and decision core, while AI has turned verification from a bottleneck into a productivity multiplier.

Arm CEO and SoftBank Group International lead Rene Haas argued that no computing problem has ever been invented that does not utilize the microprocessor. Accelerators generate tokens, but the “token factory” needs trucks to deliver them to users—that is what CPUs do. In a 24-36 month chip design cycle, the bulk of time is spent on verification, validation and debug, not architecture. AI excels here; 80-90% of Arm engineers use it daily. Shutting it off would be like 1990s internet available only between 2 and 4, after which you must go to the library down the hall—anarchy. In five-plus years, simpler designs could go from idea to GDSII. Arm moved from pure IP licensing to compute subsystems and then to its own AGI CPU to meet demand from customers such as Meta for a general-purpose agentic CPU. Robotics, edge AI and data centers will all be powered by Arm because power efficiency is its core strength.

OpenClaw and OpenAI-related developer Peter Steinberger announced that the next release (or dev channel) makes worktrees ~80% faster via APFS/Btrfs/XFS/ReFS folder clones while also saving substantial disk space. He will test for a few weeks and, if it helps the majority of users, try to get it into Codex. Related components are written entirely in Rust—“because you can’t escape Rust.”

Anthropic Claude Code team member Boris Cherny shared that Fable solved the Cyphral Distich, a 370-year-old cipher, calling it a super cool way to use Claude.

Replit CEO Amjad Masad said it was painful to see so many users priced out of AI coding for a while, but now it is free to build again.

Builder Zara Zhang noted a problem with Astra: it acknowledges a mistake and agrees it should do Y, yet does not simply go and do Y the way other models do. Every CEO Dan Shipper quipped that he started using Astra medium for simple tasks and called it “pacing the frontier.”

FPV Ventures partner Nikunj Kothari noted he gets at least one text a day from founders facing this exact conundrum. Higher entry valuation and higher total comp for applicants is often a bug, not a feature. Raising at a higher valuation does not mean a company is more secure or the better place to go. Look objectively at the market, traction and positioning, then do your own analysis of likely exit value, 409A prices and tax implications. He has first-hand experience: a company that raised $60M from one of the best investors folded after three years and wiped out all equity.

AI tutorial author Peter Yang highlighted a chart from Brex CEO Pedro: each dot is 3.2 million people. Gray boxes are people who never used AI (84%), green are free chatbot users (16%), orange are those paying $20/month (~0.3%), and the tiny red box is people using agents effectively (~0.04%). The chart is from around February 2026 so usage has likely risen, but we are still super early.

Vercel CEO Guillermo Rauch simply stated: “You don’t get it, I build software to relax.”

Y Combinator President & CEO Garry Tan noted that the public content about startups is merely the teaser for the specific advice given 1:1 once you are actually a YC startup.

FirstMark partner Matt Turck joked about the face of a seed investor when a company he forgot about years ago gets acquired out of nowhere. Nikunj added the moment a seed investor realizes a marked-down investment is suddenly going to return half the fund.

SPC General Partner and Bevel Health co-founder Aditya Agarwal wrote: imagine a world where we didn’t have to rely on any one organization to be good and virtuous.