AI Builders
Digest

Google DeepMind Gemini co-lead Oriol Vinyals argues world models are a critical path to AGI, with Google uniquely focused on multimodal understanding. Omni demonstrates impressive video editing via language interaction, though pure visual knowledge extraction remains a breakthrough needed. Vinyals stresses memory systems (working and episodic), multi-agent orchestration, and continual learning, predicting acceleration in robotics via simulation. Post-training is still a greenfield, with RL poised for expansion beyond coding/math. Key quote: "We're at the foothills of AGI."

Box CEO Aaron Levie notes that AI has made creating and finding security issues far easier, but reviewing, responding to, and fixing them still requires significant human judgment and triage. This is Jevons paradox in action, leading to a security engineer boom. Engineers aren't going away - they're becoming more essential.

Swyx shares Kakuna, an agent focused on hardening codebases with checklist skills. Plan with it then let it run for a day - it returns same functionality with boring work done plus self-audit. Focuses on subagent parallelism and "mullet factory" approach: party in front for unique features, dark in the back for production principles.

YC President & CEO Garry Tan released GBrain v0.40.0 giving OpenClaw/Hermes Agent voice capabilities based on Gemini Live. Mars as friend, Venus as EA with large context and tool use. Open source under MIT license - his gift so everyone can have a personal AI.

Peter Yang is obsessed with how best solo founders and engineers use agents to 10x output. He asks about AI stacks, end-to-end building, and managing multiple agents. Advice includes learning Codex/Claude Code, shipping side projects, and building GitHub history.

FPV Ventures partner Nikunj Kothari announced leading the Series A of a special company (not AI). Perfect way to start the weekend. He stresses this time is too important to not be doing your life's best work.

Claude Blog announces Dreaming for self-improvement by reviewing sessions, Outcomes for rubric-based self-correction, and multi-agent orchestration. Teams like Harvey, Netflix, and Spiral are using them to boost completion rates and handle complex work.

Peter Yang offers 6 things to take back control amid mass layoffs: read signals, learn Codex/Claude Code, build side projects, develop GitHub history, become top 10% at your craft, and let the market determine your value. Entrepreneurship may be the safest job in AI era.

Garry Tan explains that in "bar is zero" markets, Geoffrey Moore's chasm doesn't apply. Buyers accept 60% solutions because alternatives don't exist. Example: YC's 9 Mothers counter-drone defense.

Swyx co-signs a framework on what kinds of learning transformers do well today and their limitations. References earlier work with Ankit on adversarial world models to approach the Kolmogorov-limit generator of reality. Scaling inefficient paradigms may be outclassed by truth-seeking solutions.

Google Labs refreshed their site to easily find latest innovations including I/O announcements. Team shares most underrated features.

Matt Turck highlights continuous compounding progress behind the scenes at OpenAI driving recent AI advances.

Peter Steinberger (Polyagentmorous ClawFather at OpenClaw) welcomes native GitHub 10 PR per person limit, which they previously enforced with bots.

Ex-OpenAI CPO Kevin Weil loves the quote "Make no little plans" as a powerful mental framework.

Every CEO Dan Shipper will discuss his piece "After Automation!" exploring opportunities beyond automation.

Claude's Problem Solvers series features founders like Genspark AI CTO Kay Zhu using Claude. Team makes the difference in fast-moving markets.