AI Builders
Digest

Kevin Weil, VP Science at OpenAI, shared that not only is AI solving more open problems, but its proofs are getting more elegant as the models improve.

This observation highlights how AI is evolving from raw computation toward more refined, insightful problem-solving in mathematics and science.

Box CEO Aaron Levie offered a grounded contrarian perspective on AI agent productivity: the ultimate rate limiter will be security, compliance, governance, and the ability to review agent work to ensure regulatory compatibility.

We've been in la-la land assuming enterprises will let agents freely 'vibe code' everything; the last 48 hours show why companies won't take on every risk with every piece of technology. There's no free lunch—firms must build systems, processes, and controls so agents can't run wild on any data at any time.

Aditya Agarwal, General Partner at South Park Commons and Co-Founder of Bevel Health, former early engineer at Facebook and CTO at Dropbox, reflected on early Facebook's "fixing bugs in production" philosophy that once sounded crazy to outsiders.

He suggests the modern equivalent for dynamic software might be simply making your source code public. With software evolving so rapidly, this approach may not be as batshit crazy as it seems—or perhaps it still is.

The Takeaway: AI-powered manager agents will soon enable a single founder to run a billion-dollar company by orchestrating self-running operations across product development, sales, support, and more.

Andrew, CEO and founder of The General Intelligence Company of New York—his second venture after dropping out of college at 18 to lead as CEO—has a clear mission: build organizations that run themselves with AI.

Starting with Co-Founder, their cloud-based AI assistant with advanced memory that automates emails, candidate searches, and operations (used by over 8,000 startups), they realized an assistant alone doesn't make a company. A full company needs to create a product, sell it, support it, and operate it autonomously.

Their new Co-Founder CTO delivers the first fully autonomous engineering department in the cloud. It uses manager agents to delegate to coding agents like Claude, browser agents for testing, and Vercel for deployments in preview environments—no local dev required. Features go from ticket to tested, merged PR automatically, with infrastructure monitoring and auto-fixes. Internally, their team is 400% faster, averaging 25 PRs per engineer per day, sometimes spending more on tokens than salaries. As Andrew notes, "we've started spending more on tokens than on salaries depending on the day."

Over the next six months, customer support agents and a full CRO revenue stack will complete the picture, turning the one-person billion-dollar company from vision to reality.

Swyx, AI builder and affiliate with Cognition, Temporal.io, and Latent Space podcast, wondered what happened with no observed growth in Codex during March and shared supporting data.

This might indicate a slowdown or shift in certain AI coding tool adoption trends.