Project Genie just got a real upgrade: you can now start worlds from Google Maps Street View locations, organize creations in a library, and share them externally. They also teased Computational Discovery, where AlphaEvolve and ERA generate and test thousands of code variants to speed up model and algorithm discovery.

He says token costs are about to become a top CIO headache, and nobody in Fortune 500 land feels like they have a clean answer yet. The current playbook is messy: throttle by workload, tier access by user, cap spend by team, or make teams prove the ROI before they get more AI. He also says Gemini 3.5 Flash is a real jump for knowledge work, with Box AI tests showing a 12-point gain on complex document tasks and big wins in healthcare, public sector, and life sciences.

They’re shipping a CDN pricing model that smooths over traffic spikes and viral events, so teams get predictable costs without sacrificing performance or getting shoved onto slower routes or priority tiers. He also teased Claude Managed Agents working with Vercel Sandbox, but the big move here is making infra pricing behave less like a surprise tax.

He says Gemini 3.5 Flash is a real step up: stronger multimodal performance, big gains in agentic coding, and leading scores on a bunch of key benchmarks. The caveat is obvious — it’s still just benchmarks and not cheap — but his read is that Google is back in the race and the three-lab competition is making everyone better.

He says the Bay Area still hasn’t fully priced in the shift from assistants to coworkers to autonomous workers. His take: the missing pieces are long-horizon training data, better task harnesses, and models that can now self-correct well enough to keep pushing into real jobs over the next 10–20 years.

He says the old Microsoft dream of WinFS — a system that actually understands and organizes your files — is now plausible with LLMs, and that’s basically what his GBrain project is aiming at. It’s a neat full-circle take from someone who worked on the original effort in 2003-2005: same problem, new tech, finally enough intelligence to pull it off.

He says the best builders don’t sit on year-long plans — they ship, learn, and iterate 3-4 times to see what sticks. The vibe is very product-lead-in-an-AI-era: 90-day roadmaps, constant experimentation, and keeping the builder muscle from turning to mush.

He lays out a 4-part AI SDLC: start with real test coverage, use `/plan` to carve up hot paths and improve maintainability, let the agent break backward compatibility when needed, then keep spot-checking and steering bugs as it ships. The vibe is clear: AI coding works best when you treat it like an aggressive junior engineer that still needs guardrails, logging, and human judgment.