Interview, Fireside Chat, Podcast
Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness
- The "Gemini 3.5 era" is identified as the current phase where the agentic layer is "actually becoming true," with the industry moving from a "little bit early" stage to products "rebase to become sort of agentic native" capabilities.
- "Anti-gravity" is positioned as a through-line technology intended to power agentic features across Search, the Gemini app, Cloud, and AI Studio, serving as the core engine for "24-7 always on agent" and "autonomous coding agents."
- Google's strategic path over the next 12 months involves shifting focus from maximizing "eyeball time" to maximizing customer outcomes, with a prediction that agentic capabilities will increase search interactions despite potential business cannibalization.
- Current product experiences are described as being in the "crawl" phase for the majority of the 13 billion user base, while the "Gemini app" and "anti-gravity" represent the closest steps toward "walk" or "running" capabilities.
- Over the next one to two years, ecosystem value creation regarding human behavior and agent usage is expected to become "somewhat clear," whereas outcomes three to five years out remain "much less clear."
- The speaker predicts that "Flash" models, specifically "3.5 flash," will deliver superior coding performance compared to any previous "pro model," accelerating value realization from DeepMind's pre-training efforts.
- "Long running agents" and "coding agents" are established as key KPIs for DeepMind, acting as an accelerant for the broader business, though "10,000 TPUs" usage requires human oversight to prevent nonsensical resource allocation.
- "Narrow super intelligence" is currently evident in coding, with a trajectory toward "vertical superintelligence" in domains like science, math, and finance where verifiability is high, preceding the realization of general AGI.
- By the end of 2025, the capability to "vibe code video games" is expected to be accessible to everyone, though current development relies on a "coding agent plus some sort of game engine" rather than standalone world models.
- The "Omni" model and its "Omni Flash" iteration represent a shift from traditional action-conditioned video models to a single model with world understanding, with significantly more capable versions expected beyond the current first iteration.
- Rapid development via "identity coding" has enabled the Gemini Mac OS app to be delivered faster than any previous Google team, contributing to the build of "350,000 android apps" in AI Studio, the majority of which are personal applications.
- The "model eats the harness" narrative is predicted to play out over the next 12 months, where models will "digest" scaffolding, shifting the competitive advantage from building custom harnesses to the models' native capabilities.
- Independent companies face "even more opportunity than there was" 24 months ago due to coding capabilities allowing startups to run faster, countering earlier concerns that agentic AI would reduce their viability.
- While "20% of apps built in AI Studio" were previously games, the ecosystem is expected to diversify, though "finance-related stuff" currently comprises around 20% of usage, with science cited as a high-potential domain for rapid, impactful gains.
- Challenges in "sprite generation" and the need for an "orchestration layer" indicate that "vibe coding video games" will require significant product scaffolding and tooling to become reusable and replayable experiences.
- Google's unique scale allows for the deployment of Gemini to billion-user products, a problem the speaker notes "only two companies in the world have," while anticipating that future world model definitions will blur as technology improves.