TickerTain
TickerTain
NewsroomShortsPortfolioConvergence
NewsroomShortsPortfolioConvergence
←
▶ 17:57 · AI Infrastructure · Post-training gains (3.5 Flash) surpass pre-trained Pro models; pre-training windows matter
episode briefing
Sequoia Capital

Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness

2026-06-11 · 2 company · 8 thematic
sentiment
2 bull0 bear0 neu
speakers
logan kilpatrick

Logan Kilpatrick runs Google AI Studio and the Gemini API at Google DeepMind, focusing on developer platforms and the next generation of AI builders. He previously worked at OpenAI on developer relations.

episode shorts · 3

"So much alpha in authenticity" | Google DeepMind's Logan Kilpa…

AI didn't just make him faster. It made him more ambitious. | G…

Google's agentic shift | Logan Kilpatrick

now playing · AI Infrastructure
AI Economics & Business Modelstailwindscore 6/10logan kilpatrick
Agentic AI is positive-sum for search/ecosystem; maximize outcomes not eyeballs
Contrary to cannibalization fears, AI answers increased search volume (positive-sum). Google's success metric shifts from maximizing eyeball time to maximizing customer outcomes — agents do…
Enterprise AI Adoptiontailwindscore 7/10logan kilpatrick
Enterprises only see coding agents working today; science and finance verticals coming this year
Enterprises report only coding agents work reliably today, but longer-horizon agents are spiking in capability (3-day autonomous runs). Verticals with verifiability — math, finance, science…
AI Infrastructuretailwindscore 7/10logan kilpatrick
Post-training gains (3.5 Flash) surpass pre-trained Pro models; pre-training windows matter
Gemini 3.5 Flash surpassed all previous Pro models purely through post-training gains, demonstrating that pre-training compute windows and cluster availability create lagging external perce…
AI Coding Agentstailwindscore 9/10logan kilpatrick
Coding agents are narrow superintelligence; model eats the harness in 12 months
Coding has become the first vertical where AI achieves narrow superintelligence, enabling developers to tackle 10x more ambitious problems. The external agent harness (scaffolding) will be…
World Models & Simulationtailwindscore 8/10logan kilpatrick
Omni model blurs world model definition; single multimodal model replaces ensemble
Google's Omni model is a single unified model (not routing) that understands the world across video, audio, image, and text — blurring the line between world models and generative models. T…
Generative Mediatailwindscore 7/10logan kilpatrick
Omni model edits video while preserving human authenticity; not AI avatars
Omni's video editing preserves the creator's words, voice, and personhood while altering non-personal elements (sets, objects) — amplifying authenticity rather than replacing humans with AI…
AI Applicationstailwindscore 8/10logan kilpatrick
Vibe coding enables personal software; 350k Android apps built in AI Studio in one week
AI Studio saw 350,000 Android apps created in a single week — most solving personal problems that wouldn't have been built before. Native mobile capabilities (sensors, messaging, OS integra…
AI Agentstailwindscore 8/10logan kilpatrick
Anti-gravity agent harness powers all Google products; model eats scaffolding
Google's anti-gravity harness — spanning IDE, CLI, SDK, and managed API — is the shared agent infrastructure powering Search, Gemini app, Cloud, and Android Studio. The pattern repeats: sca…