Insights, research, and perspectives from the forefront of artificial general intelligence

Coding agents aren't just "more autonomous" or "less autonomous"—they vary along two distinct dimensions: the scope of changes they produce (incremental to greenfield) and how frequently humans review their work. Drawing on insights from Gemini Build Day, where 200+ builders spent 18 hours pushing the boundaries, this memo maps the landscape from GitHub Copilot to Google Antigravity and explains why context engineering—not raw model capability—is becoming the key differentiator.

Native multimodal architectures have hit an inflection point. Drawing from our Gemini 3 x Graphon AI hackathon with 200+ builders, we analyze three core theses: why native fusion creates defensible moats over late-fusion approaches, how embodied applications unlock proprietary data flywheels, and why prosumer controllability drives 10x better retention than consumer magic. Includes field notes, technical challenges, and opportunities for builders at the infrastructure and application layers.

Google's Gemini 3 crosses 1500 Elo and claims the LMArena crown with native multimodality, 1M token context, and sparse MoE architecture. We break down the research innovations—from early-fusion transformers to the new Antigravity IDE—and provide 25 curated project ideas that leverage Gemini's unique capabilities for our December 13th Build Day.

Frontier APIs win generic tasks, but specialized SLMs win where structural constraints favor local deployment. Drawing on academic research, 2025 breakthroughs, and our AWS hackathon, we map four investment theses: proprietary data flywheels, enterprise packaging, hardware-architecture co-design, and edge-cloud orchestration.

Self-evolving agents are reshaping how we think about the future of intelligence and the opportunities ahead. Our latest AGI House Hackathon surfaced insights that sharpen our investment thesis across the full stack.