AI Builders
Digest

Cerebras founder and CEO Andrew Feldman emphasizes that AI speed doesn't just improve existing tools but unlocks entirely new business models. Just as Netflix evolved from DVDs to a movie studio, fast AI will fundamentally reorganize work. Feldman notes that once models became smart enough for daily use in 2025, inference speed became critical, with Cerebras' wafer-scale architecture delivering 15-20x performance over GPUs.

Multiple AI builders shared experiences with tools like Cursor, Replit, and Codex. Replit CEO Amjad Masad highlighted users building MVPs in a weekend with Replit Dial and getting Apple approval on first try. Peter Steinberger shared practical Codex tips for big refactors, including maintaining a scratch-log for decisions and tradeoffs.

The Takeaway: Radical hardware innovation delivers explosive growth when market timing aligns. Cerebras founder and CEO Andrew Feldman shares the journey from long-term bet on wafer-scale chips to becoming the fastest AI inference platform. Demand exploded in 2025 as models became practical, leading to a $20B+ deal with OpenAI and AWS partnership.

Feldman stresses that radical improvement requires a fundamentally different design: "You're not going to get 15 or 20 times better than the GPU with a minor modification." He discusses early skepticism, manufacturing challenges, and how sovereign customers like G42 helped bridge to mainstream. Cerebras now focuses on delivery while preserving a fearless culture. Feldman believes speed will enable new business models beyond incremental efficiency.

Replit CEO Amjad Masad shared user feedback that building with Replit Dial was dramatically faster than Cursor, with one user completing an MVP over a weekend and getting Apple approval on the first try.

Peter Steinberger from OpenAI built a GitHub dashboard to see repos, open Issues/PRs, latest releases, and commits. He also recommends having Codex maintain a scratch-log during big refactors to document decisions, tradeoffs, and review fixes.

Vercel CEO Guillermo Rauch analyzed 1400 replies showing OpenAI catching up to Anthropic, with 'Codex' mentioned more than 'Claude Code' but Anthropic models dominating mentions. He asked builders to share their proudest AI creations with links.

Anthropic's Thariq demonstrated how the "please save me money" prompt effectively identifies cost savings in legacy codebases and services.

Box CEO Aaron Levie argues people confuse task automation with job elimination. As AI handles tasks, job definitions expand to more volume, higher quality, or new un-automated work, applying to coding, legal, sales, and marketing.

Y Combinator President Garry Tan released gbrain-evals showing SOTA performance in reranking and embedding. He also shared fine-tuning a Qwen3.5-397B model quickly and pledged $50K to fight against efforts pushing startups out of San Francisco.

South Park Commons' Aditya Agarwal noted that just six months ago we were happy chatting with LLMs, but by May 2026 these models have produced more code than humans have written throughout history.

FPV Ventures partner Nikunj Kothari reminded B2B companies to seize opportunities and use strong narratives to stand out amidst the slop.