AI Builders
Digest

NVIDIA's Bryan Catanzaro highlights that open technologies are essential for AI's diverse applications, similar to the internet. With Nemotron 3 Ultra leading US open-weight models, key advances include 4-bit pretraining for massive efficiency gains, hybrid transformer/state space architectures for better reasoning, Mixture of Experts with latent innovations, multi-token prediction for faster inference, and multi-teacher distillation for specialized capabilities. The focus is on agentic workflows and running at computational limits through thoughtful efficiency rather than raw scale. "If you accept as the truth that we're gonna be running at the limit, then what that means is that the way to get more intelligence is to be more efficient." Open collaboration, including the Nemotron Coalition, drives progress alongside global contributions.

Anthropic's Fable 5 dominates conversations with builders experimenting over the long weekend. Cat Wu shares tips for using Claude Code with computer use to set up data integrations. Thariq details discovering unknowns to prompt better with HTML artifacts. Dan Shipper highlights massive token usage scenarios for personal apps and bug backlogs. Peter Steinberger feeds his tweets to Fable for roasting and explores imagegen for design iteration. Multiple builders report mind-blowing demos and productivity leaps.

Vercel CEO Guillermo Rauch emphasizes agentic self-improvement: agents introspect past runs, spot inefficiencies, and generate new prompts/skills. Built-in observability when deploying to Vercel supports this. Replit CEO Amjad Masad promotes video generation capabilities on the platform.

Box CEO Aaron Levie argues the AI battle is for context. Effective agents need domain expertise, tools, and governed knowledge. Applied AI layers organize knowledge, route between models, and handle change management for strong moats. Platforms capturing best context will excel.

Zara Zhang notes people are less willing to buy pure tools as they can build with coding agents. What they pay for is the feeling of hiring true expertise. Swyx contrasts beautiful "tools for thought" canvases being outpaced by effective but plain CLIs that handle commodity thinking.

Nikunj Kothari praises Gemini for unifying capabilities in one API key: Flash for fast tasks, image gen, search grounding, realtime audio/video. Simplifies many side projects.

Linear's Nan Yu celebrates Tailwind's victory as it's adopted by the company behind Bootstrap. Also notes fields without good training data were likely full of crap.

Garry Tan discusses AI transforming specialist wait times and care quality. Advocates building housing over performative actions and proper treatment funding.