The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More
TL;DR
Google DeepMind's Logan Kilpatrick and Tulsee Doshi detail the launch of Gemini 3.5 Flash, Omni video generation, and Spark agent features, emphasizing a strategic pivot toward cost-adjusted performance and standardized agent infrastructure ('anti-gravity') across Google's product ecosystem rather than competing solely on absolute model capability.
⚡ Flash-First Strategy & The Performance Frontier 3 insights
Prioritizing Speed and Cost Over Raw Scale
Google launched Gemini 3.5 Flash first—3x faster and significantly cheaper than larger models—to optimize for latency and cost at consumer scale, recognizing that users won't tolerate wait times even for marginally better quality.
Skipping the Ultra Tier
The team abandoned the 'Ultra' branding despite scaling up Pro capabilities internally, finding that Flash and 'Flashlight' variants better serve billions of real-world users and Google's mission to democratize AI access.
Bidirectional Model Development
Rather than one-way distillation from large to small, DeepMind scales recipes bidirectionally, using Pro to improve Flash while simultaneously scaling Flash techniques up to enhance Pro.
🔗 Agent Infrastructure & Anti-Gravity 3 insights
The Agent Harness as Foundation
DeepMind now ships models integrated with 'anti-gravity'—a standardized agent harness that provides the infrastructure layer for agentic behaviors, replacing the old practice of dumping isolated models on product teams.
Sub-Agent Orchestration
The harness enables complex multi-agent workflows previewed via the '/teamwork' feature, where sub-agents complete tasks autonomously and power upcoming consumer experiences like Gemini Spark.
Cross-Product Symbiosis
Models are co-designed with the harness to ensure they function simultaneously across Search, the Gemini app, Cloud, and AI Studio, creating consistent agentic experiences across Google's diverse product surfaces.
🎬 Multimodal Expansion 2 insights
Omni Video Generation Launch
Gemini Omni Flash introduces video generation and editing with avatar insertion, representing Google's first major push toward 'any modality in, any modality out' starting with creative video tools.
Gemini Live Upgrades
Real-time conversation capabilities now feature faster response times, significantly improved background noise detection, and smarter interaction patterns designed to feel like a natural partner.
Bottom Line
Google is betting that winning the AI era requires optimizing the cost-performance frontier and embedding standardized agentic infrastructure across its entire product ecosystem rather than merely competing to build the most capable single model.
More from Cognitive Revolution
View all
Three Kinds of Software Survive: Tasklet's Andrew Lee on Competing to be a Horizontal Platform
Tasklet CEO Andrew Lee reveals a complete architectural rebuild shifting from workflow automation to a general-purpose AI agent platform, emphasizing file-based context management and aggressive summarization to control token costs, while outlining a strategic pivot toward becoming a horizontal platform capable of integrating any frontier model as competition intensifies with API providers like Anthropic.
Milliseconds to Match: Criteo's AdTech AI & the Future of Commerce w/ Diarmuid Gill & Liva Ralaivola
Criteo's CTO Diarmuid Gill and VP of Research Liva Ralaivola detail how their AI infrastructure makes millisecond-level ad bidding decisions across billions of anonymous profiles, while explaining their new OpenAI partnership to combine large language models with real-time commerce data for accurate product recommendations.
"Descript Isn't a Slop Machine": Laura Burkhauser on the AI Tools Creators Love and Hate
Descript CEO Laura Burkhauser distinguishes 'slop'—mass-produced algorithmic arbitrage for profit—from necessary 'bad art' created while learning new mediums. She reveals a clear hierarchy in creator acceptance of AI tools: universal love for deterministic features like Studio Sound, frustration with agentic assistants like Underlord, and visceral opposition to generative video models, while outlining Descript's strategy to serve creators without becoming a content mill.
The RL Fine-Tuning Playbook: CoreWeave's Kyle Corbitt on GRPO, Rubrics, Environments, Reward Hacking
Kyle Corbitt explains that unlike supervised fine-tuning (SFT), which destructively overwrites model weights and causes catastrophic forgetting, reinforcement learning (RL) optimizes performance by minimally adjusting logits within the model's existing reasoning pathways—delivering higher performance ceilings and lower inference costs for specific tasks, though frontier models may still dominate creative domains.