AI Builders Digest

Stanford’s AI coding course puts students inside real open-source projects

Mihail Eric replaced easy coding exercises with contributions to large open-source projects, where maintainers will review students’ work and teach them how complex software is designed. His updated course also covers preparing codebases for agents, code review, security, and background agents. At Monaco, his rule for choosing tools is: “You buy to learn and you build to keep.”

• #agents
• #education
• #open-source

Levie expects parallel agents to change the scale of knowledge work

Aaron Levie argues that assigning work to many agents at once could produce gains beyond the 20–50% improvements he associates with chat-based AI. He sees coding as an early example and expects similar effects in cybersecurity and science. His practical advice is to retry failed AI workflows every three to six months as models improve at tool use and longer tasks.

• #agents
• #productivity

Nan Yu: Start with the dumbest design that works

Nan Yu says simple software is something you arrive at through iteration; the right starting point is the dumbest architecture and design that gets the job done. Clever designs make users work harder to understand them. AI has made building easier when the goal is clear, but can make an unclear goal more expensive by helping teams ship a pile of junk.

• #products
• #software

Tan says a replicated engineering result improved with particle-1

Garry Tan says his gstack-evals project replicated a Jev Engineering article’s result and found particle-1 performed even better, though his posts give no figures or method. He also argues that easier software changes make lingering bugs less tolerable and point toward personal software built when needed.

• #evals
• #agents
• #software

Rauch builds a record player that simulates the groove itself

Guillermo Rauch built a WebGL2 and Web Audio record player that turns a song into a simulated physical groove. A virtual stylus follows that groove 384,000 times a second, producing crackle, pops, and pitch variation as effects of the simulation. The code is linked for others to remix and deploy.

• #products
• #open-source

Zara Zhang questions what “vibe coding” implies about effort

Zara Zhang has nearly stopped using “vibe coding” because it can imply that work made with AI lacks care or effort. She argues that people can put heart and soul into a product while using AI coding tools, and wants a better term for that work.

• #software
• #products

Kothari says showing your work beats asking to network

Nikunj Kothari’s advice for meeting people in San Francisco is to bring evidence of effort: a pointed message, a shipped side project, or a meetup you organized. He says generic requests to “swap notes” get lost amid low-effort outreach, while specific work gives busy people a reason to make time.

• #startups
• #careers