AI Builders
Digest

Surge founder and CEO Edwin sees his company building a school for AGI where models learn about humanity and how to run the world. As capabilities advance, teaching shifts from basic math benchmarks like GSM8K to research-level math via Riemann bench. OpenAI models recently broke through on AirDish problems, even disproving a conjecture.

Edwin explores AI surpassing humans while still being driven by human-set goals. He references a Ted Chiang story to stress pretending we have free will and continuing to create to preserve humanity. On model optimization, he warns against engagement-maximizing that turns AI addictive like social media, advocating delegation and human uplifting instead.

Anthropic team shares the origin story of Claude Code from safety research, now supporting Artifacts that turn work into live, shareable visual pages like PR walkthroughs, dashboards, and release checklists. These are built automatically from session context and update in real-time.

Swyx highlights the J-space paper's demonstration of models detecting reasoning interventions, showing deeper understanding. Boris Cherny, Cat Wu, and official Claude promoted the new capability.

Replit CEO Amjad Masad highlights an Atlanta real estate company saving $100k by replacing Salesforce with a Replit-built CRM. Replit is improving rapidly via closed loop self-improving agents.

Vercel CEO Guillermo Rauch reports AI radically increased his productivity and agency, enabling quick creation of personal tools. Vercel teams use agents to autonomously build impactful software. He stresses evals are essential for agents and ships eve eval out of the box.

Claude Blog details how Artifacts transform Claude Code sessions into live updating pages for debugging, PR reviews, incident investigations, and more. Pages are private, shareable with version history.

Peter Yang shares 5 Fable 5 use cases worth trying: finding high-value tasks, business advice, making projects ship-ready, planning big projects, and refactoring AI skills, with copy-paste prompts.

Box CEO Aaron Levie discusses open source AI: frontier intelligence for new use cases and orchestration, while mature cases shift to tuned models. The applied AI layer handles evals, model mixing, and custom training.

Zara Zhang recommends binge-watching talks from AI Engineer, Cursor Compile, and Figma Config conferences. Even non-engineers benefit greatly. High-quality knowledge is freely available online.

Linear Head of Product Nan Yu notes agents enable early startups to do way more in parallel. Bruteforcing tedious tasks with long hours is less necessary as automation advances.

Swyx analyzes Anthropic paper; Thibault Sottiaux promotes OpenAI events; Dan Shipper on Fable gremlins; Peter Steinberger on AI-assisted interviews and OpenClaw.