AI Builders
Digest

The Takeaway: As AI makes software radically cheap, security and reliability become the real value—security spend should vastly outstrip software spend. Meanwhile AI is already generating genuinely new knowledge, with math breakthroughs opening the door to scientific discovery in areas like drug development.

OpenAI cofounder and president Greg Brockman argues that recent math results prove AI can now resolve century-old open problems and produce new knowledge. “It’s just impossible to deny… it is definitively yes, possible.” The real excitement is what a human equipped with these tools can achieve that was previously unthinkable. The same capabilities are being transferred to biology. Drug discovery ranks as the clearest top priority: from first principles, neural nets that excel at natural language should also crack the “language of biology.” Data bottlenecks, wet-lab innovation, and regulatory processes will determine the pace.

On cybersecurity, Brockman says we are in a “defender’s window”—peak frontier capabilities exist but are not yet broadly diffused. Defenders must use this moment to harden systems. OpenAI has made a billion-dollar commitment to frontline defenders and internally paused 25% of production engineers to saturate vulnerabilities with its strongest models. He states: “Cyber spend, security spend should actually vastly outstrip software spend.” Once software is nearly free, the value shifts to software that is reliable, secure, and trustworthy.

He also highlights DOTS, a proactive always-on agent that notices calendar conflicts, unfinished tasks, and context across inbox and Slack without being asked. The best UI, he argues, is no UI—generative interfaces that mold to the moment. The through-line remains human agency: “Humans should be firmly in the driver’s seat, should have control over the future, and this technology should amplify human agency.”

Box CEO Aaron Levie notes that we have crossed a threshold where it is already practical for AI to be better at training models than humans. Model capability exceeding most human engineers, advances in harness and infrastructure for long-running agents, growing data, and the ability to parallelize agents across far more tasks and tests than any person could manage create an incredible flywheel.

This flywheel will almost certainly accelerate model progress far faster than before—or at minimum push the frontier ceiling at an insane rate. “If you aren’t pricing in what RSI will lead to, that’s probably a mistake.”

Replit CEO Amjad Masad observes that mathematicians are moving through denial, anger, bargaining, depression, and acceptance in a highly compressed window. We are at the tail end of depression and entering acceptance. Separately, he announced that Replit can now be deployed to a customer’s own cloud so data never leaves home—designed for high-stakes environments such as government, finance, and healthcare.

Boris Cherny of the Claude team at Anthropic says he now makes every slide deck with Claude Slides—no more googling for “iphone frame svg.” Claude Code Review runs on all his PRs and catches almost every bug before merge; it is part of the secret sauce that lets the team move fast while raising the quality bar.

Thibault Sottiaux, who leads Codex and ChatGPT at OpenAI, introduced ChatGPT Space—endearingly awkward at times because it is the actual team building it. Slides support was added yesterday and is currently available to Pro, Business, and Enterprise users while the team learns from feedback. He also urged users to update the app for the latest experience.

Vercel CEO Guillermo Rauch rebuilt the Roadster demo as a playable, highly accurate simulation that retains many real-world details. He also noted that TypeScript compiled with scriptc outperforms tsgo by 2-4x—“TypeScript is the new TypeScript.”

Y Combinator President & CEO Garry Tan says his favorite part of starting a new project is pairing Capy with AsideAI so agents can both code the first version and set up all the DevOps in one fell swoop. He also pointed to the easiest hosted way to try GBrain.

Every CEO Dan Shipper observes that generation is now cheap and high quality; the big capability bump still needed is editing and oversight. He also shared a team demo that turned a Codex piece into a visual novel.

FPV Ventures partner Nikunj Kothari shared that a prediction he wrote seven months ago came full circle: after meeting roughly 50 startups in the space, he found the one and is leading its Series A—a perfect way to end the week.

OpenAI CEO Sam Altman showed the back of the Navier-Stokes milestone shirt, which prints the last line of the chain-of-thought from the solution. He also called the new Roadster “quite amazing” and expressed happiness at progress from Nick.

The official Claude account highlighted weekend creations: Haiku 5.5 wrote twenty haiku then drew every frame in JavaScript, and a cardiology fellow built a 3D tool to help trainees learn to read coronary angiograms.

AI tutorial creator Peter Yang took content from docs to slides to video and finally to a rap battle as the ultimate form of communication, asking whether he should make another one in the same style.

Swyx of smol.ai and Latent Space was struck by a video explanation that reached near-Kurzgesagt clarity and urgently wants the prompt(s) behind it. He also celebrated launching a show he had been waiting almost a year to release.