AI Builders
Digest

The Takeaway: There is no such thing as a moat—long-term winners in AI and business must relentlessly develop new services and better ways to serve customers. Booking Holdings CEO Glenn Fogel, a 27-year veteran who grew the company from near collapse to over $100B market cap, shares hard-won lessons from the dot-com bust that parallel today's AI boom. He sees AI as a powerful tool for personalized travel agents like Priceline's Penny, which handles complex itineraries, frequent flyer optimization, and disruption recovery far beyond human capabilities. Fogel stresses understanding industry complexity, regulatory hurdles, and customer needs over hype. "There is no such thing as a moat. There is no such thing as somewhere you're gonna be protected against innovation." He highlights scaling agentic AI while managing token costs and ROI, with customer service costs already down 10% and satisfaction up.

Box CEO Aaron Levie highlights Cursor's research showing frontier models as planners/orchestrators paired with cheaper workhorse models deliver 15X cost improvements. "Few moments in a large task genuinely require frontier intelligence... Once a frontier planner has collapsed the ambiguity into a detailed, explicit instruction, less expensive models simply have to follow it." This routing approach is becoming core for complex agents across coding, finance, legal, and healthcare.

Swyx points to a notable trajectory comparison in the RLM paper from @a1zhang and @lateinteraction. An open secret in frontier model training is training on test lookalikes to game benchmarks, but open weights often hide the datasets. They explore NLP distance metrics on hidden trajectories, supporting that RLMs generalize to unseen tasks sharing latent structures.

Builder Zara Zhang describes two kinds of companies: pre-agent retrofits vs. post-agent natives with tiny teams, project-based work, and closed loops. For hiring: Round 1 in-person no-AI domain test; Round 2 AI-mandatory project evaluated on results and agent chat transcripts.

Claude Code engineer Thariq notes a short-lived bug encountered during late-night coding on his personal account.

Replit CEO Amjad Masad celebrates what appears to be the first physical product shipped by a coding agent. He also reflects on cancellation as a choice and shares tool recommendations.

Vercel CEO Guillermo Rauch draws the big lesson from AI: everything is code—from slide decks and design to promo videos, Excel automation, and possibly the universe.

Peter Yang argues banning Chinese models is a self-own like EVs. Madhu Guru (Meta) says the road to AGI is paved with economically valuable enterprise tasks and notes product sense is key; tokenomics debates now center on open vs closed weights. Matt Turck reacts to Chinese open source models pressuring OpenAI/Anthropic. Nikunj Kothari warns against assuming scale/capital as moats, citing historical failures like Webvan. Garry Tan comments on compute, California policy, and staying frosty. Dan Shipper is hiring for agents at Every. Sam Altman notes improvements. OpenAI's Thibault Sottiaux loves the dynamic workplace.