AI Builders
Digest

AI engineer Swyx reconstructed a WSJ chart comparing OpenAI's $850B valuation at ~$30B ARR with Ant's ~$900B valuation at ~$44B ARR, noting Ant's revenue would be ~$8-10B lower using OpenAI's methodology. This highlights the aggressive multiples in AI versus fintech.

Roblox product leader Peter Yang outlined a three-phase AI adoption trajectory: coding is the first frontier where AI is already delivering, knowledge work is the second, and personal agents represent the third wave. This framing suggests a natural progression from technical to personal use cases.

Box CEO Aaron Levie notes that both Anthropic and OpenAI are launching initiatives to help enterprises deploy AI agents. Beyond the models, there's real work in upgrading IT systems, providing context, modernizing workflows, managing human-agent relationships, and driving adoption—creating massive opportunities for new jobs and firms.

FPV Ventures partner Nikunj Kothari argues that startups from 2023-2025 are realizing that fancy launch videos and distribution focus might get VC funding but still burn cash if retention is neglected. Momentum is not a moat, and the seed-to-Series A gap is widening, leading to more acquihires.

South Park Commons GP Aditya Agarwal emphasized that they never desired to be an accelerator, because velocity without grounding towards true north is not interesting. This reflects a philosophy of purposeful building over mere speed.

OpenAI CEO Sam Altman expressed excitement for voice models getting great, noting that people are already starting to change how they interface with AI—shifting toward voice. This signals a major UX evolution in AI products.

The Takeaway: Waymo has entered exponential scaling, giving 20 million fully autonomous rides, with 10 million in the last 7 months. Waymo co-CEO Dmitri Dolgov revealed that their foundation model is an end-to-end multimodal world action model, powering driver, simulator, and critic. Safety is non-negotiable: the Waymo Driver is 13x safer than human drivers for serious injury collisions, preventing a serious injury every 8 days. A memorable moment: LiDAR detected a pedestrian's footsteps under a bus, allowing the AI to predict and react. Waymo is expanding to 11 cities and internationally to London and Tokyo this year.

Replit CEO Amjad Masad shared a story of an entrepreneur who used Replit to build their product and successfully attracted investors and landed meetings. This showcases Replit as a launchpad for non-technical founders to get to funding.

Amjad Masad highlighted a great use of AI for education: a multi-modal learning platform for deaf students, built using Replit. This demonstrates AI's potential to create inclusive educational tools.

Vercel CEO Guillermo Rauch announced npx deepsec, an open-source agent orchestrator for deep security reviews. It can find critical vulnerabilities in minutes that would take teams months. Optimized for Vercel Sandbox, it harnesses thousands of agents scrutinizing code in parallel. He invites OSS projects to DM for a sponsored run.

YC CEO Garry Tan's GBrain is a unified tool combining memory layer, code tool, and search engine under one graph with one query interface. v0.27 adds support for non-Anthropic/OpenAI embeddings and LLMs. Multi-modal embeddings and photo OCR coming soon. He uses it daily with a 100k markdown file and OpenClaw+Hermes Agent setup.

FPV Ventures partner Nikunj Kothari calls Gemini Flash 'criminally cheap and good,' with 1M context windows and structured outputs. It's his most used model in production workloads. He also praises Google's new live voice model as mindblowingly good.

OpenClaw creator Peter Steinberger released Crabbox 0.5.0, enabling ephemeral remote CI boxes with WebVNC for reproducing issues. Agents can set up exact state, test, fix, and post videos on PRs. Now supports desktop/browser leases, AWS Windows with WSL2, screenshots, and app launch.

Anthropic Engineering introduced auto mode for Claude Code, using model-based classifiers to delegate approvals. It catches overeager actions like deleting remote branches or exfiltrating data. A two-stage classifier (fast filter then reasoning) reduces false positives to 0.4% while blocking most dangerous actions. It also includes prompt injection defense. Auto mode is designed to replace --dangerously-skip-permissions safely.