Google DeepMind maintains complete foundation model parity across consumer and enterprise products, deploying identical Gemini 3.1 models to Google Cloud customers on the same day they launch internally for Search, YouTube, and Gmail. (1:18)
DeepMind expanded its generative media and autonomous agent lines with Gemma 4 open weights (2B-30B parameters), native speech-to-speech Gemini Live, Genie 3 interactive 3D world models, and Gemini Robotics ER 1.6 deployed on Boston Dynamics Spot robots. (4:12)
Google Cloud structures its Enterprise Agent platform around four operational pillars—build, scale, govern, and optimize—using live stack trace inspection to enforce policies and dynamically inject execution traces into the context window for self-improvement. (23:01)
Replit reported that efficient Flash-tier models match the capabilities of frontier models from six months prior at a fraction of the cost, compressing the traditional multi-stage SDLC into a rapid create-and-republish loop driven by autonomous background agents. (29:20)
Detailed notes
## Background
Published on YouTube by Google Cloud Tech on June 25, 2026, this session from Google Cloud Next features leadership from Google DeepMind, Google Cloud, and Replit. According to the description, the presentation provides an inside look at recent Gemini research breakthroughs, highlighting advancements in multimodal intelligence, context window processing, and agentic workflows built on Vertex AI.
## TL;DR
Google DeepMind and Google Cloud introduce the Gemini 3.1 model family—spanning Pro, Flash, and Flash-Lite—alongside frontier initiatives like Gemma 4, Gemini Live native audio, Genie 3 world modeling, and Gemini Robotics ER 1.6. The presentation explores how Google maintains strict parity between internal consumer products and Google Cloud's enterprise agent platform across the core pillars of building, scaling, governing, and optimizing agents. Concluding with a real-world case study, Replit demonstrates how tiered Gemini models compress the software development lifecycle, democratize software creation for knowledge workers, and power custom internal business tools.
## Atmosphere & style
The session carries the polished, forward-looking energy of an enterprise tech keynote blended with the candid rapport of a collaborative fireside chat. The dynamic between researchers, cloud product leads, and partner builders feels cohesive and practical, punctuated by occasional moments of unscripted levity, such as laughing off faint background music filtering into the room. Technical discussions on latency budgets, stack traces, and governance balance visionary demonstrations of interactive world generation and robotics, giving viewers an authoritative yet grounded view of autonomous agent deployment.
## Bulletin
- [01:18](https://www.youtube.com/watch?v=92LvgAcR6fI&t=78s) Google DeepMind maintains complete foundation model parity across consumer and enterprise products, deploying identical Gemini 3.1 models to Google Cloud customers on the same day they launch internally for Search, YouTube, and Gmail.
- [04:12](https://www.youtube.com/watch?v=92LvgAcR6fI&t=252s) DeepMind expanded its generative media and autonomous agent lines with Gemma 4 open weights (2B-30B parameters), native speech-to-speech Gemini Live, Genie 3 interactive 3D world models, and Gemini Robotics ER 1.6 deployed on Boston Dynamics Spot robots.
- [23:01](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1381s) Google Cloud structures its Enterprise Agent platform around four operational pillars—build, scale, govern, and optimize—using live stack trace inspection to enforce policies and dynamically inject execution traces into the context window for self-improvement.
- [29:20](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1760s) Replit reported that efficient Flash-tier models match the capabilities of frontier models from six months prior at a fraction of the cost, compressing the traditional multi-stage SDLC into a rapid create-and-republish loop driven by autonomous background agents.
## Key takeaways
- [01:18](https://www.youtube.com/watch?v=92LvgAcR6fI&t=78s) Google DeepMind has evolved from an independent research lab into the unified 'engine room of Google,' centralizing all company-wide AI model development under its original startup-style leadership.
- [01:18](https://www.youtube.com/watch?v=92LvgAcR6fI&t=78s) The Gemini 3.1 family is organized into three tiers—Pro for complex reasoning and agent orchestration, Flash as the balanced workhorse, and Flash-Lite for low-latency, high-throughput tasks.
- [01:18](https://www.youtube.com/watch?v=92LvgAcR6fI&t=78s) Gemini models feature native multimodality built from the ground up, allowing direct and interleaved processing of video, audio, text, images, and code without stitching separate models together.
- [04:12](https://www.youtube.com/watch?v=92LvgAcR6fI&t=252s) Gemini Live achieves real-time conversational capabilities by processing and emitting audio tokens directly, eliminating speech-to-text and text-to-speech latency to support emotional steering and proactive audio.
- [06:49](https://www.youtube.com/watch?v=92LvgAcR6fI&t=409s) Genie 3 generates navigable 3D virtual environments from a text or image prompt, providing spatial simulation grounds that feed physical embodiment models like Gemini Robotics ER 1.6.
- [16:07](https://www.youtube.com/watch?v=92LvgAcR6fI&t=967s) Google's vertical integration across custom Tensor Processing Units (TPUs), global low-latency networking, and AI data centers drives compute efficiency that translates directly into cost-effective models like Gemini Flash.
- [14:21](https://www.youtube.com/watch?v=92LvgAcR6fI&t=861s) External developer feedback on Google Cloud provides critical signal for DeepMind's path to AGI because cloud customers possess market choice, unlike internal product teams acting as a captive audience.
- [10:42](https://www.youtube.com/watch?v=92LvgAcR6fI&t=642s) Deploying enterprise agents requires moving beyond raw foundation models to provide secure data access controls, strict credentialing, and end-to-end auditability across proprietary data layers.
- [19:42](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1182s) Enterprise model selection aligns to three operational boundaries: frontier reasoning for complex analysis, latency-bound targets like emergency dispatch (Flash), and infinite-volume scaling like internet-scale content moderation (Flash-Lite).
- [21:49](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1309s) Context engineering outmatches micromanagement; treating models like high-performing new hires by communicating overarching business goals and providing execution traces yields continuous agent self-improvement.
- [25:01](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1501s) Software development interfaces are pivoting from manual line-by-line coding to agent management, expanding the addressable developer base from tens of millions of coders to 1.5 to 2 billion knowledge workers.
- [29:20](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1760s) Replit runs entirely on Google Cloud Run and Google Cloud Storage, allowing simple user prototypes and MVPs to scale instantaneously to millions of users without manual infrastructure configuration.
- [31:17](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1877s) Roughly 50% of the applications generated on Replit represent novel, unseen use cases, requiring continuous online A/B testing and task completion tracking rather than static test suites to benchmark agent quality.
- [34:33](https://www.youtube.com/watch?v=92LvgAcR6fI&t=2073s) Bespoke internal software is actively replacing off-the-shelf SaaS, with Replit building approximately 80% of its internal operations tooling across HR, sales, and finance directly on its own agentic platform.
## Detailed walkthrough
### [00:14](https://www.youtube.com/watch?v=92LvgAcR6fI&t=14s) Welcome and Introductions
Michael Gerstenhaber, VP of Product Management for the Gemini Enterprise Agent platform, opens the session at Google Cloud Next 2026 by welcoming attendees and introducing his co-presenters: David Thacker, VP of Product at Google DeepMind, and Michele Catasta, President and Head of AI at Replit. Setting the stage for the presentation, Catasta succinctly defines Replit's overarching ambition, explaining, “In case you haven't heard of us yet, we are a platform to democratize software creation for everyone.”
David Thacker lays out the roadmap for the talk, describing it as a blended panel and presentation. DeepMind kicks off the session with an insider look across their frontier model portfolio, followed by Gerstenhaber outlining how Google Cloud operationalizes and productizes these breakthroughs for enterprise developers, and concluding with Catasta sharing how Replit leverages this underlying technology to power software creation workflows.
[01:18](https://www.youtube.com/watch?v=92LvgAcR6fI&t=78s) Google DeepMind's Evolution and the Gemini 3.1 Model Lineup
David Thacker traces Google DeepMind's roots back to its 2010 founding in London with the audacious goal of achieving artificial general intelligence (AGI)—a concept largely dismissed as pure science fiction at the time. Google acquired DeepMind in 2014, intentionally allowing the laboratory to operate independently for its first five to six years. Over the past three years, however, DeepMind has been brought directly into the core of Google as what Thacker terms the “engine room of Google.” While AI model development across the entire company has now been centralized within DeepMind, the team continues to operate with the nimble culture of a startup under its original founder and CEO.
The flagship result of this consolidated development is the Gemini family, which is just over two years old and now at version 3.1. Thacker emphasizes that Google does not maintain separate tiers of models for internal teams and external enterprise users: the exact same foundation models that power Google Search, YouTube, and Gmail are pushed directly to Google Cloud customers, frequently on the very same day they launch internally. Gemini 3.1 is specifically tuned around core pillars including complex reasoning, agentic multi-step planning, versatile tool use, and software engineering. Thacker reflects on this model parity, noting: “The same models that we use for Google products are the same models that we put into our Cloud platform for customers to use.”
```mermaid
flowchart TD
accTitle: Gemini 3.1 Model Family and Deployment Flow
accDescr: Diagram illustrating Gemini 3.1 model tiers deployed simultaneously across Google Products and Google Cloud.
- **Gemini 3.1 Pro**: The largest and most capable tier; optimized for STEM problem-solving, advanced mathematics, software engineering, and multi-step agent orchestration.
- **Gemini 3.1 Flash**: The primary workhorse model; strikes an optimal balance between execution speed, operational cost, and high-level reasoning.
- **Gemini 3.1 Flash-Lite**: The smallest, fastest footprint designed for low latency, high throughput, and cost-sensitive tasks.
- **Internal Adoption**: Tens of thousands of Google software engineers use Gemini daily for internal code authoring and systems development.
- **Native Multimodality**: Built from the ground up to natively process interleaved video, audio, text, images, and raw code.
[04:12](https://www.youtube.com/watch?v=92LvgAcR6fI&t=252s) Open Weights and Expressive Media: Gemma 4, Gemini Live, and Lyria 3 Pro
Beyond the Gemini frontier line, Thacker highlights DeepMind's broad portfolio of science, robotics, open-weight, and generative media models. DeepMind launched Gemma 4, their latest generation of lightweight open models derived directly from the Gemini 3 architecture. Ranging from 2 billion up to 30 billion parameters, Gemma 4 is tailored for local on-device deployment and specialized task fine-tuning, featuring native audio and video understanding alongside broad multilingual coverage.
Simultaneously, DeepMind has prioritized speech with Gemini Live, a fully native audio-to-audio model that processes and emits audio tokens directly. This architecture bypasses the latency and nuance penalties of separate speech-to-text and text-to-speech pipelines, enabling real-time conversational agents capable of emotional steering and what Thacker calls "proactive audio"—where the model dynamically senses and mirrors human conversational affect. On the creative front, DeepMind introduced Lyria 3 Pro, a state-of-the-art music model capable of synthesizing a complete song up to three minutes long with full vocal arrangements from a single text or image prompt. Emphasizing the team's belief in conversational interfaces, Thacker explains: “We really think we're really bullish on voice as a modality to engage with intelligence and with AI models.”
- **Gemma 4**: Open-weight models spanning 2B to 30B parameters, optimized for efficient on-device execution and task-specific fine-tuning with built-in audio and video comprehension.
- **Gemini Live**: Native speech-to-speech processing with direct audio token generation, low latency, nuanced stylistic steering, and emotional mirroring.
- **Lyria 3 Pro**: High-fidelity music generation creating up to 3-minute compositions, complete with full vocals and instrumentation, directly from image or text prompts.
[06:49](https://www.youtube.com/watch?v=92LvgAcR6fI&t=409s) Autonomous Workflows and Embodied AI: Deep Research, Genie 3, and Gemini Robotics ER
Thacker introduces the Gemini Deep Research agent, launched just prior to the event. Designed to eliminate the labor of open-ended exploratory research, Deep Research can be triggered via a single API call to scour both the open web and private enterprise data repositories. Instead of simply generating a wall of text, the agent synthesizes comprehensive reports embedded with dynamically generated charts and infographics to seed downstream business workflows.
The presentation then turns to simulated worlds with Genie 3, DeepMind's foundational world model available to AI Ultra subscribers. A brief moment of levity occurs when background audio filters into the room, prompting Thacker to joke, “That's some great music. Sounds like Jump by Van Halen.” He then explains how Genie 3 allows a user to type a text prompt or submit an image to generate a fully interactive, walkable 3D environment navigable via keyboard.
```mermaid
flowchart LR
accTitle: Genie 3 Simulation to Physical World Loop
accDescr: How Genie 3 world modeling feeds into robotics embodied reasoning in the physical world.
subgraph Simulation["World Simulation (Genie 3)"]
Prompt["Text / Image Prompt"] --> GenWorld["Interactive 3D World"]
GenWorld --> SimAgent["Agent Spatial Learning"]
end
subgraph Physical["Physical Embodiment (Gemini ER 1.6)"]
Vision["Multimodal Vision"] --> ER["Embodied Reasoning Core"]
ER --> Actuation["Hardware Actuation<br>(e.g., Boston Dynamics Spot)"]
Actuation --> Tasks["Facility Inspection<br>Gauge Reading & Object Counting"]
end
SimAgent -.->|"Simulation Training Data"| ER
```
These spatial simulations serve as a training ground for DeepMind's robotics initiatives, bridging digital intelligence and real-world hardware. Thacker details Gemini Robotics ER 1.6, where "ER" signifies Embodied Reasoning. This model acts as a physical agent's brain, pairing Gemini's visual recognition and logical reasoning with mechanical actuation. Thacker highlights their hardware partnership with Boston Dynamics, who integrated ER 1.6 into their Spot quadruped robot to execute autonomous industrial inspections. Summarizing the ultimate goal of connecting DeepMind's model development with the Google Cloud platform, Thacker states: “We really want to see our AI transform every industry and every geography.”
- **Gemini Deep Research Agent**: Autonomous multi-source research agent grounding on enterprise data and public web sources to produce structured briefs with custom data visualizations.
- **Genie 3 World Model**: Real-time generation of controllable, interactive 3D virtual environments from single image or text inputs, used for entertainment, education, gaming, and agent spatial simulation.
- **Gemini Robotics ER 1.6**: Embodied Reasoning platform equipping physical robots with multimodal scene understanding, planning, and real-time physical actuation.
- **Boston Dynamics Deployment**: Spot quadruped robots running ER 1.6 to navigate industrial facilities, autonomously count inventory, and inspect analog gauges that previously required human intervention.
[09:56](https://www.youtube.com/watch?v=92LvgAcR6fI&t=596s) Bringing Frontier Research to Google Cloud
David Thacker emphasizes that Google DeepMind is bringing the very same core models that power flagship Google consumer products directly into the Google Cloud platform. DeepMind works closely with the Google Cloud team because developer and enterprise feedback provides critical, diverse signal for refining and advancing systems toward artificial general intelligence (AGI). Thacker notes that he and Michael Gerstenhaber talk at least once a day, while their broader teams are in constant coordination. “We’ve been really excited to see as soon as we launch a new model in the Google Cloud platform, it gets adoption right away,” Thacker explains, before handing the stage over to Gerstenhaber to detail how Cloud deploys these capabilities for enterprise environments.
[10:42](https://www.youtube.com/watch?v=92LvgAcR6fI&t=642s) Hyperscale Infrastructure and Enterprise Agent Architectures
Gerstenhaber reflects on Google’s long history and the compounding infrastructure advantages that feed directly into the current AI wave. DeepMind pioneered the core model technology, while Google Cloud developed a high-performance global network originally built to deliver cat videos from New York to Tokyo in 75 milliseconds. That exact low-latency infrastructure is now mission-critical for agents: an airplane technician working on a jet engine needs split-second answers from a field manual through real-time voice, keeping their hands free rather than paging through documentation. Decades of work on hyperscale reliability, coupled with visual assets like Google Maps and Street View, directly reinforce Google's multimodal reasoning capabilities.
Moving from consumer models to enterprise agents presents unique hurdles. Having launched models across the industry—including Claude 3 during his time at Anthropic and Gemini 3 upon arriving at Google—Gerstenhaber points out that models like Gemini 2.5 and Gemini 3 already exceed what is required for many operational workflows, even as Gemini 3.1 sets the frontier for elite tasks like world-class code authorship. Financial institutions use Gemini for Know Your Customer (KYC) automation, and biotech and life sciences organizations run complex statistical data analyses on clinical trial datasets.
Yet, generic intelligence alone is insufficient: out-of-the-box models do not inherently know an enterprise's proprietary customer background records or locked clinical trial databases bound for FDA submissions. Enterprise agents require a reliable platform that provides governed data retrieval, rigorous credentialing, and end-to-end auditability.
```mermaid
flowchart LR
accTitle: Enterprise AI Agent Architecture
accDescr: Pipeline showing data retrieval through an audit and trust boundary to an autonomous agent engine at cloud scale.
subgraph INGEST["1. Governed Data Layer"]
A["Enterprise Data\n(KYC, Clinical Data)"]
B["Credentialing &\nAccess Control"]
A --> B
end
subgraph ENGINE["2. Agent Platform"]
C["Gemini Reasoning\nEngine"]
D["Synthetic Data &\nClassification"]
B --> C
C --> D
end
subgraph SCALE["3. Trust & Observability"]
E["Decision Auditing"]
F["Cloud-Scale Execution\n(Human-Out-Of-The-Loop)"]
D --> E
E --> F
end
```
- Grounded context requires integrating locked clinical trial databases and enterprise KYC records securely into the agent's workflow.
- Platforms must manage fine-grained credentialing, data access control, and information retrieval boundaries.
- Observability and auditability are non-negotiable: platforms must track agent decisions, data access events, and generated synthetic data classifications.
- “So it is only with trusts in front and trust at the end of that pipeline that the middle is going to scale without a human in the loop, because if it's scaling with a human in the loop, it's really only scaling to the size of your teams and not the size of the Cloud,” Gerstenhaber argues.
[14:21](https://www.youtube.com/watch?v=92LvgAcR6fI&t=861s) The DeepMind-Cloud Feedback Loop
Gerstenhaber shifts the presentation into an informal dialogue with Thacker, asking how direct customer feedback flows back into DeepMind to make the models smarter, faster, and more efficient. Thacker reveals that Gemini now has millions of developers building on the platform. To steer model releases, DeepMind frequently runs public and private previews across a diverse spectrum of users—from early-stage startups to massive Fortune 500 enterprises.
This real-world usage determines which model checkpoints launch and how specific capabilities are steered. While internal Google product teams are a massive testing ground, Thacker candidly points out their limitations: “I think it's great to have all the Google products using Gemini, but the Google teams are kind of a captive audience. They have to use our models. And so what's great about Cloud customers is you all have choices.” Because external developers can choose competing providers, their willingness to adopt Gemini directly forces DeepMind to build the most capable and practical models on the market.
[16:07](https://www.youtube.com/watch?v=92LvgAcR6fI&t=967s) Google's Full-Stack Advantage and Native Multimodality
Gerstenhaber asks Thacker where DeepMind deliberately focuses its technical differentiation beyond benchmark chasing and standard "hill climbing." Thacker identifies Google’s vertically integrated, full-stack architecture as an unmatched advantage. DeepMind collaborates directly with internal chip designers building custom Tensor Processing Units (TPUs), as well as the Cloud infrastructure teams engineering next-generation AI data centers optimized for high-throughput training and inference.
Running Gemini in Google Search—delivering AI Overviews and AI Mode to more than two billion users globally—forces the entire stack to operate with ultra-low latency and extreme compute efficiency. These efficiency gains directly yield fast, cost-effective models like Gemini Flash for Cloud customers. Furthermore, DeepMind prioritizes native multimodality from day one to emulate how the human brain processes real-world inputs:
- Gemini processes raw multimodal inputs natively rather than stitching separate vision and audio models together.
- Users can input end-to-end video (such as a full video recording of a conference panel) and receive coherent, highly nuanced critiques and analysis.
- Generative capabilities span text, images, video, and music generation.
- Real-world workflows pipeline multimodal inputs with grounding tools: an agent can retrieve a user's location via Google Search grounding, fetch local weather data, and synthesize a personalized video of that user standing in the rain.
[18:56](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1136s) Software Engineering and Architectural Discipline
Thacker directs the conversation toward software development, noting that generative AI is transforming the software development lifecycle (SDLC) and dramatically speeding up implementation. He asks Gerstenhaber how he advises Cloud customers on determining what to build versus what *not* to build in an environment where code generation has become nearly effortless.
Gerstenhaber welcomes the question, drawing on his seven years at Datadog prior to Google and Anthropic. At Datadog, he oversaw monitoring solutions across diverse environments like Kubernetes, observing that whether serving Capital One or Airbnb, organizations often face remarkably similar core architectural problems despite operating in vastly different business domains.
[19:42](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1182s) Selecting the Right Model: Frontier, Latency, and Infinite Volume
Working with enterprise customers on AI departs completely from traditional infrastructure discussions about pods and clusters; it is fundamentally mission-oriented. Engagements with clients like Thomson Reuters center on points of law and parsing specific contract clauses to assess whether they favor an advocate or an adversary, while work with major banks zeroes in on anti-money laundering and Know Your Customer (KYC) workflows. When customers ask what they should build, the key is determining what proprietary knowledge they possess that no one else in the world has, and establishing how to trust the model. Across these deployments, three distinct operational boundary conditions emerge:
```mermaid
flowchart TD
accTitle: Three Model Boundary Conditions
accDescr: Categorization of AI use cases across Frontier of Intelligence, Latency-Bound, and Infinite Volume.
subgraph Boundaries["Model Boundary Conditions"]
B1["Frontier of Intelligence<br/>Complex reasoning, deep analysis"]
B2["Latency-Bound Operations<br/>Urgent real-time policy execution"]
B3["Infinite Volume Scaling<br/>Platform-wide content processing"]
end
First is the frontier of intelligence. While coding benchmarks constantly push forward, routine catalog extraction tasks—such as verifying whether green chairs are green—reach a natural ceiling where the chairs do not get any greener, meaning frontier reasoning is unnecessary. Second are latency-bound scenarios, such as a 911 operator or a DoorDash courier needing to evaluate and execute a policy in real time during an urgent situation. These environments demand the maximum possible intelligence delivered within a fixed time budget, which is where Gemini Flash excels. Third are infinite-volume workloads, typified by massive online platforms attempting to moderate the entire internet without knowing incoming post counts or severity; for this, Gemini Flash-Lite—originally engineered for the scale of Google Search—provides the throughput required to handle unbounded volume.
- Routine metadata extraction does not require frontier model capabilities once a baseline quality is met.
- Latency-sensitive real-time policy applications rely on Gemini Flash to maximize intelligence within strict time budgets.
- Flash-Lite was specifically architected for Google Search scale to handle infinite, unpredictable request volumes affordably.
---
[21:49](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1309s) Coaching Agents and Context Engineering Over Micromanagement
Foundation models routinely outperform humans across domains where an individual lacks deep, multi-decade experience. Because of this, enterprise teams must resist the urge to micromanage how a model solves a problem. Rather than prescribing procedural steps, organizations succeed by communicating overarching goals, strategies, and business context, then trusting the system to execute. This mirrors onboarding a talented recruit:
> “When you hire a brilliant intern or not even an intern, a new hire from Princeton with one year of experience or something, they're going to pick up what you're putting down pretty quickly.”
Just as a new hire at Datadog cannot perform specialized time-series arithmetic on day one without onboarding, models need targeted background information. You cannot script every customer interaction ahead of time; you equip the worker with context and trust them to adapt in the field. Crucially, while underlying foundation models are immutable, software agents possess the ability to achieve continuous self-improvement through runtime evaluation and dynamic memory injection.
---
[23:01](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1381s) The Google Enterprise Agent Platform: Build, Scale, Govern, and Optimize
Building an agent involves far more than picking an LLM. Google Cloud organizes its Enterprise Agent platform around four operational pillars: *build*, *scale*, *govern*, and *optimize*. While engines and deployment harnesses have long addressed build and scale by simplifying deployment and plumbing in observability, governance and optimization represent the critical additions that function hand in hand.
```mermaid
flowchart LR
accTitle: Agent Governance and Optimization Flywheel
accDescr: Real-time execution tracing feeds into governance policy checks and optimization context injection.
subgraph Runtime["Agent Execution"]
A["Agent Action"] --> B["Stack Trace Inspection"]
end
subgraph Control["Govern & Optimize Flywheel"]
B --> C{"Policy Breached?"}
C -- "Yes" --> D["Halt Action"]
C -- "No / Evaluation" --> E["Identify Successes & Failures"]
E --> F["Inject Execution Trace into Context"]
F --> G["Extrapolate & Self-Improve"]
end
G -.-> A
```
The platform observes agent actions in real time by analyzing live stack traces. Governance enables immediate intervention if an agent breaches an established enterprise policy. Optimization uses these traces to teach the agent on the fly by feeding past mistakes and successes back into its context window:
> “Context engineering isn't a process of writing a lot of English with your hands. It's really just throwing a trace into the context window and say, don't do that ever again.”
From that single trace, the agent extrapolates principles for future behavior, creating a compounding flywheel that safeguards customer interactions while steadily elevating agent productivity.
---
[25:01](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1501s) Replit's Agent Architecture and Democratizing Engineering
Michele Catasta shares how Replit has utilized Gemini across its developer platform since the project's inception. Replit specifically values the tiered design of the Gemini family, noting that intelligence cannot be treated as a single monolithic package. Pushing the raw frontier with flagship models while deploying Flash and Flash-Lite as cost-effective "Swiss army knives" enables Replit to enrich product interactions with intelligent features without inflating user budgets. Running Replit’s entire infrastructure on Google Cloud Platform (GCP) and accessing Gemini through Vertex AI provides unified operational oversight, significantly reducing maintenance overhead for Site Reliability Engineers (SREs) compared to juggling fragmented model providers.
This infrastructure supports a fundamental market expansion. Software engineering is no longer confined to individuals who understand Kubernetes, carry pagers, and wake up at odd hours of the night for production incidents. By abstracting systems engineering, the target audience expands from the roughly 30 to 50 million professional coders worldwide to an addressable base of 1.5 to 2 billion knowledge workers.
- Replit utilizes Gemini Flash and Flash-Lite across product surfaces to maintain high delight while protecting operating margins.
- Unifying model access on Vertex AI within GCP centralizes operational telemetry, simplifying maintenance for SRE teams.
- The shift removes deep infrastructure prerequisites like Kubernetes knowledge, opening development tools to mainstream knowledge workers.
---
[26:51](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1611s) Managing Subagents and Evolving the Knowledge Worker Application Layer
Software development is actively pivoting away from manual code editing toward agent management. Replit advanced this shift by launching its agent platform in early March, redesigning the primary user experience so users act less like manual line-by-line builders and more like project managers:
> “You spend quite a lot of time planning, and then we spin up agents in the ground that accomplish the job. So even though the product hides most of the complexity, in reality, our users as of today are managers of agents already.”
While elite software engineers may continue running multiple terminal shells and parallel IDEs, the broader public requires interfaces that mask underlying complexity. Users must be able to orchestrate multiple agents executing tasks in parallel without shouldering the cognitive load of tracking tens or hundreds of low-level subagent threads.
At the application layer, raw model horsepower must be translated into frictionless outcomes for general knowledge workers who simply need a job done. Consequently, Replit is looking beyond strictly software-centric paradigms toward a universal workspace for knowledge tasks. While this includes spinning up full applications and autonomous agents, it also extends into generating slide decks and motion videos, empowering users to execute any complex digital workflow from a single unified prompt.
[29:20](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1760s) Cloud Infrastructure and Compressing the SDLC on Replit
Michele Catasta explains how Replit enables users to express their creativity through software creation while entirely abstracting away the friction of writing code, maintaining codebases, and publishing applications. Under the hood, this relies directly on enterprise-grade infrastructure: Replit runs on Google Cloud Run and Google Cloud Storage. Because of this foundation, even when a user spins up a simple MVP or prototype, it is immediately capable of scaling to millions of users overnight without additional configuration.
When asked how software development life cycles will change in the future, Catasta argues that the traditional SDLC studied in college—spanning requirements gathering, implementation, testing, and deployment—is effectively collapsing into a tight, two-step feedback loop: creating software and republishing it. Everything else runs continuously behind the scenes. Replit deploys autonomous background agents, such as a production agent that constantly monitors deployments, evaluates query speeds, analyzes error logs, and recommends concrete fixes. Catasta highlights a newly launched feature that tracks dependencies for published CVE vulnerabilities, automatically warning users and generating remediation tasks that developers can resolve and deploy in a single click:
> “So the cycle that we studied in college, several different steps, you write your requirements and then you create the software and write tests I think is all being compressed into this creation experience that you have on Replit.”
```mermaid
flowchart LR
accTitle: Compressed Software Development Loop
accDescr: Diagram showing user workflow on Replit interacting with background production and security agents.
subgraph UserLoop["User Workflow"]
A["Create Software"] --> B["Republish Application"]
end
C -.->|"Recommends Fixes"| A
D -.->|"Generates Remediation Tasks"| A
B -.->|"Runs on Cloud Run & Cloud Storage"| E["Scale to Millions of Users"]
```
[31:17](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1877s) Model Orchestration and Rigorous Evals
Catasta discusses the evolution of how Replit orchestrates models across different capability tiers, contrasting their early heuristic-driven testing with their current quantitative approach. Initially, evaluation of early Gemini 1.5 Flash and Flash-Lite models was largely "based on vibes." Over time, they recognized that efficient, lightweight models offer an immense economic advantage: a Flash-tier model typically matches the performance of a frontier model from just six months prior at a fraction of the cost. This rapid price-to-performance shift reshapes the product roadmap, allowing features that were initially cost-prohibitive for global long-tail users to become democratized six months later:
> “If you think about it, a Flash model is usually six months behind where the frontier was. So if you use Flash today, you're basically using a way more expensive model from six months ago in the frontier.”
Today, Replit tests agent configurations through rigorous nightly automated rollouts and continuous A/B testing in production. These evaluations determine exactly which sub-tasks and accessory functions can be safely routed to smaller, lower-intelligence models rather than top-tier frontier models.
Evaluating arbitrary coding workloads presents unique challenges because roughly half of all applications created on Replit are unique queries never seen before on the platform—analogous to Google Search query distribution. Since automated tests cannot anticipate every custom application architecture, Replit relies on offline benchmarks (sharing feedback directly with Google's Gemini team) coupled with online tracking of user sentiment and sustained engagement. User sentiment and task completion success rates have climbed steadily over the past eighteen months, driven by both model improvements and more capable orchestration scaffolding. Catasta expects these success metrics to begin plateauing within the next twelve months as models fully master the standard software development playbooks for web applications.
- Flash models track approximately six months behind frontier models while dramatically cutting compute costs.
- Replit conducts continuous online A/B testing and nightly redeployments of agent configurations.
- Approximately 50% of the applications created on Replit represent completely novel, unique use cases.
[34:33](https://www.youtube.com/watch?v=92LvgAcR6fI&t=2073s) Emerging Use Cases: Functional Prototypes and Internal SaaS
Addressing common developer patterns across both enterprise teams and weekend hobbyists, Catasta identifies two dominant trends. First, users leverage Replit for zero-to-one prototyping. Rather than assembling static landing pages, UI mockups, slide decks, or six-page memos, creators walk into stakeholder meetings with fully functional applications accessible via a live URL, allowing decision-makers to evaluate real interactive software on day one.
Second, Replit has become an engine for custom internal business tooling, fundamentally reshaping how internal SaaS is delivered across the industry. Organizations increasingly build bespoke internal applications rather than purchasing off-the-shelf software packages. Within Replit itself, the team actively dogfoods this pattern:
> “So all our internal tools ranging from HR sales finance, everything is built on our platform and we see this happening across the entire industry.”
Catasta notes that roughly 80% of the internal operational software used daily across Replit's HR, sales, and finance operations was built directly on Replit. The session concludes with the hosts thanking the Replit team for their partnership, active model testing, and feedback loop with Google DeepMind and the Gemini engineering teams.
## Notable quotes
> [23:01](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1381s) Context engineering isn't a process of writing a lot of English with your hands. It's really just throwing a trace into the context window and say, don't do that ever again. — Michael Gerstenhaber
> [10:42](https://www.youtube.com/watch?v=92LvgAcR6fI&t=642s) if it's scaling with a human in the loop, it's really only scaling to the size of your teams and not the size of the Cloud — Michael Gerstenhaber
> [31:17](https://www.youtube.com/watch?v=92LvgAcR6fI&t=1877s) If you think about it, a Flash model is usually six months behind where the frontier was. So if you use Flash today, you're basically using a way more expensive model from six months ago in the frontier. — Michele Catasta
## Mentioned
Michael Gerstenhaber, David Thacker, Michele Catasta, Google DeepMind, Google Cloud, Vertex AI, Replit, Gemini 3.1, Gemini 3.1 Pro, Gemini 3.1 Flash, Gemini 3.1 Flash-Lite, Gemini 2.5, Gemini 3, Gemma 4, Gemini Live, Lyria 3 Pro, Gemini Deep Research, Genie 3, Gemini Robotics ER 1.6, Boston Dynamics, Spot, Anthropic, Claude 3, Datadog, Thomson Reuters, DoorDash, Google Cloud Run, Google Cloud Storage, Kubernetes
## Video description
Join Google DeepMind for an exclusive look at what’s new with Gemini and the research breakthroughs defining the next generation of multimodal intelligence. We’ll showcase new workflows and demos designed to empower and inspire developers to build and scale agentic applications with Gemini. Discover what Gemini 3’s upgraded model reasoning, expanded context window processing, and enhanced multimodal tool use can enable in Vertex AI. And learn more about the research advancements for Gemini that are driving agent design.
Watch more Google Cloud Next 2026 Sessions → https://goo.gle/gcloud-next-2026
Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
Speakers: David Thacker, Michael Gerstenhaber, Michele Catasta
At Google Cloud Next 2026, leaders from Google DeepMind, Google Cloud, and Replit argued that foundation models are transitioning from standalone consumer intelligence into tightly integrated, multi-tier enterprise agent architectures. They highlighted how full-stack cloud infrastructure, rigorous governance, and cost-effective model routing are fundamentally restructuring enterprise workflows and software development.
DeepMind's Model Tiering and Native Multimodality (1:18)
Google DeepMind VP of Product David Thacker detailed the centralized strategy behind the Gemini 3.1 family, emphasizing that external enterprise clients receive the identical foundation models deployed internally for flagship Google products. The portfolio is tiered by operational demand: Gemini 3.1 Pro addresses complex reasoning, advanced STEM calculations, and agent orchestration; Flash serves as the efficiency workhorse; and Flash-Lite provides high-throughput processing for cost-sensitive, massive-scale tasks. Beyond core text and code models, the ecosystem spans open-weight Gemma 4 models (2B to 30B parameters), native speech-to-speech interaction via Gemini Live, Genie 3 interactive 3D world simulation, and embodied reasoning through Gemini Robotics ER 1.6 running on Boston Dynamics Spot robots.
Thacker underscored that Google's key technical differentiator lies in native multimodality developed across a vertically integrated stack, linking DeepMind's research with proprietary Tensor Processing Units (TPUs) and custom AI data center networking. Rather than stitching discrete pipelines for text, vision, and speech, models handle interleaved inputs natively, enabling complex tasks such as ingesting full video recordings for nuanced critiques or combining location data with search grounding to synthesize dynamic video outputs.
Enterprise Agent Platforms and Operational Boundary Conditions (10:42)
Michael Gerstenhaber, VP of Product Management for Google's Gemini Enterprise Agent platform, observed that frontier intelligence alone does not resolve enterprise adoption challenges. High-stakes domains like clinical trial evaluations or banking compliance require governed data layers, strict credentialing, and verifiable audit trails. To scale without human bottlenecks, platforms must build trust at both data ingestion and output boundaries.
Gerstenhaber categorized enterprise deployments into three distinct boundary conditions rather than generic computing tiers: the frontier of intelligence for complex reasoning like Thomson Reuters' legal contract analysis; latency-bound operations where emergency dispatchers or couriers require real-time policy execution within tight time budgets via Gemini Flash; and infinite-volume operations such as internet-scale content moderation powered by Flash-Lite. Rather than micromanaging model execution with step-by-step procedural scripts, Gerstenhaber advised context engineering—providing high-level operational goals and feeding runtime execution traces back into the context window so agents can self-correct.
Software Development Lifecycles Compressed into Agent Workflows (25:01)
Michele Catasta, President and Head of AI at Replit, explained how the platform leverages Gemini across Vertex AI to expand software creation from tens of millions of professional engineers to an estimated 1.5 to 2 billion knowledge workers. Operating on Google Cloud Run and Google Cloud Storage, the platform abstracts infrastructure management such as Kubernetes. Users increasingly function as project managers directing autonomous subagents that handle implementation, while background agents continuously monitor runtime performance, query speeds, error logs, and security vulnerabilities like CVE tracking.
According to Catasta, this architecture collapses the traditional multi-stage development lifecycle of requirements gathering, testing, and deployment into a continuous cycle of software creation and republishing. He noted two prominent industry trends: rapid zero-to-one prototyping, where teams bypass static slide decks by presenting fully functional web applications, and the rise of bespoke internal tooling replacing commercial SaaS. Within Replit, roughly 80% of internal operational software across HR, sales, and finance was built directly on its own platform.
Economic Realities and Quantitative Model Routing (31:17)
Replit's orchestration strategy relies on shifting workloads away from expensive frontier models toward efficient tiers like Flash and Flash-Lite. Catasta stated that efficient models typically match the capability of frontier models from six months prior at a fraction of the cost, making previously cost-prohibitive capabilities economically viable for long-tail users. Replit evaluates these trade-offs through nightly automated configuration rollouts and live production A/B testing.
Because roughly half of the applications generated on Replit represent unique, unprecedented queries similar to the long-tail distribution of Google Search, the team combines offline benchmark feedback loops with Google's Gemini team alongside online metrics tracking user sentiment and task completion. Catasta reported that task success rates have climbed steadily over the past eighteen months, though he anticipates these metrics will plateau within the next twelve months as models fully master standard web application architectures.
While foundation models are rapidly commoditizing routine coding and knowledge workflows, sustaining enterprise adoption depends on governed execution, low-latency infrastructure, and systematic cost optimization. Presenters emphasized that managing autonomous agents—rather than manual programming—will define the future of software creation, though long-term performance ceilings across novel problem architectures remain an active area of evaluation.
Source quotes
[23:01] Michael Gerstenhaber: Context engineering isn't a process of writing a lot of English with your hands. It's really just throwing a trace into the context window and say, don't do that ever again.
[10:42] Michael Gerstenhaber: if it's scaling with a human in the loop, it's really only scaling to the size of your teams and not the size of the Cloud
[31:17] Michele Catasta: If you think about it, a Flash model is usually six months behind where the frontier was. So if you use Flash today, you're basically using a way more expensive model from six months ago in the frontier.
Original transcript
[0:14] Hi, everybody.
[0:15] My name is Michael gerstenhaber.
[0:16] I'm the VP of product management
for the Gemini Enterprise Agent
[0:20] platform.
[0:21] And I'm here today with David
Thacker and Michael from Replit.
[0:27] But I'll let them
introduce themselves.
[0:29] Go ahead.
[0:29] Hi, everyone.
[0:30] Great to see you all.
[0:31] I'm David Thacker, VP of
product at Google DeepMind,
[0:34] based in California.
[0:35] But I'm responsible
for the product teams
[0:37] that are building our models.
[0:38] And I also run DeepMind's
applied AI team.
[0:42] Michael catasta, President
and head of AI at Replit.
[0:45] In case you haven't
heard of us yet,
[0:47] we are a platform to democratize
software creation for everyone.
[0:52] All right.
[0:53] Great So I think today we've
got a bit of a combination panel
[0:57] presentation I'm going
to start off and give you
[1:00] just a glimpse of what
we've been working on
[1:02] at DeepMind across the
board, and then I'll
[1:04] transition to Michael.
[1:05] We'll talk about
some of the things
[1:06] we're doing in Google Cloud
to productize these models.
[1:08] And then we'll hear from
Replit on some of the ways
[1:10] that they're using
our technology.
[1:12] So here's us starting
off with Google DeepMind.
[1:18] So I always like to
start off by talking
[1:20] about the mission of DeepMind.
[1:22] And for those of
you not familiar,
[1:23] DeepMind was a startup
founded in London back in 2010
[1:26] with the mission to build
artificial general intelligence.
[1:29] And back in 2010 nobody
else was doing that.
[1:32] It was truly science
fiction back then,
[1:35] but the company was really
committed to this vision.
[1:37] And in 2014, Google
acquired the company
[1:40] and pretty much left it alone
for the first five or six years.
[1:42] And then really over
the last three years,
[1:44] we've brought DeepMind
into the fold of Google.
[1:46] We now call it the
engine room of Google.
[1:48] We've centralized AI model
development for Google
[1:51] within DeepMind, but
we're still run very much
[1:54] like a startup within Google.
[1:55] Our founder and CEO
still runs the company.
[1:57] And I'll take you through
some of the things
[1:59] we've been working on recently.
[2:01] But I think one of
the things that we're
[2:03] most proud of is as Gemini
or Gemini models, which
[2:05] is just over two years old.
[2:07] Now we're on our version
3.1 of the Gemini models,
[2:11] and I hope most of you are
familiar with the lineup that
[2:15] comes in a variety
of different sizes.
[2:17] Everything from our
largest, most capable model,
[2:19] which is our Pro model, which
is great for powering agents.
[2:22] It's also great for coding
for other types of stem
[2:26] or engineering math type work.
[2:28] We have our Flash
model in the middle,
[2:30] which is a great blend of
performance and efficiency.
[2:32] This is our most popular model
and really our workhorse model.
[2:36] And then finally, we have
our flashlight model,
[2:38] which is our smallest,
fastest, most performant model.
[2:41] And I think what it's
really important to note
[2:42] is that at Google,
we're building
[2:45] these models to run across all
the different Google products.
[2:48] My team works very closely
with every Google product team,
[2:51] including Google Search,
YouTube, Gmail, you name it.
[2:55] They're all powered by Gemini
today and the same models
[2:58] that we use.
[2:58] The of Google products
are the same models
[3:00] that we put into our Cloud
platform for customers to use.
[3:03] Oftentimes, we're putting
them into the Cloud platform
[3:05] the same day we're
launching them
[3:07] across Google product services.
[3:08] So you're getting the benefits
of our intelligence at Google
[3:12] through the Cloud platform.
[3:14] Gemini 3.1.
[3:15] We really focused
in a few areas.
[3:17] I think we were the
model really shines
[3:19] and stands out is in its
reasoning capabilities.
[3:21] Also its ability to
leverage multimodal
[3:25] understanding for different
types of activities.
[3:29] The model has an
incredible ability
[3:30] to analyze video, text,
audio, images, code.
[3:34] From the outset,
we designed Gemini
[3:35] to be a fully multimodal model.
[3:38] We also invested quite a bit
in making the model work really
[3:41] well for powering agents
and agentic type activities.
[3:44] So everything from
complex multi-step
[3:47] planning to the ability to use
various types of different tools
[3:49] across different environments.
[3:51] And finally, coding has been
a big focus for us at Gemini
[3:54] and making this model work
worked incredibly well
[3:56] for coding for a variety
of different reasons.
[3:58] Within Google, we use Gemini,
all of our software engineers,
[4:02] tens of thousands of
software engineers
[4:03] to help develop our software.
[4:06] But also coding is a
foundational building block
[4:08] for a lot of different
types of activities.
[4:12] Beyond Gemini.
[4:13] There's a number of other
recent innovations we've had.
[4:16] I'm just going to go
through a snapshot of what
[4:18] we're doing across the board.
[4:20] DeepMind Gemini is just
one part of our portfolio.
[4:23] We're also heavily
invested in areas
[4:25] as diverse as science
models, robotics models, also
[4:30] generative media models.
[4:31] So I'll talk about a few
of those things right now.
[4:34] But one of our most
recent launches
[4:35] is our a lineup of open
weight models called Gemma.
[4:39] So we're now on the
fourth iteration of Gemma.
[4:41] We just launched just
a couple of weeks ago.
[4:43] These are really incredible
open source models
[4:47] that are derive from the
Gemini three family of models,
[4:51] and they come in a range
of sizes, everything
[4:53] from down to two
billion parameters.
[4:55] Up to up to $30
billion parameters.
[4:58] But really great for running
on device in particular.
[5:02] Incredibly efficient.
[5:03] They're great for if you want
to train a task specific model.
[5:07] The Gemini model are
terrific for this.
[5:09] They support a wide
range of languages.
[5:11] They have audio and video
understanding built in.
[5:14] So not just text.
[5:15] And we're really excited to see
the adoption of these models
[5:19] since we've launched them.
[5:21] Another model capability
we've really invested in
[5:24] is our audio and voice models.
[5:25] So we have a native audio
model called Gemini Live.
[5:28] This allows us to take in audio
tokens and output audio tokens.
[5:33] And so it's designed to
power interactive voice
[5:35] experiences voice agents.
[5:37] If you want to
try this model you
[5:38] can use it in our
Gemini consumer app.
[5:41] We just deployed
this a few weeks ago
[5:43] and have rolled it out broadly,
but much reduced latency.
[5:47] We support things
like expressiveness,
[5:49] so you can really steer
and instruct the model
[5:51] to speak and act
in a certain way.
[5:54] It also has the ability to
react to the human that's
[5:57] talking to in terms of mirroring
emotions and things like that.
[6:01] We call that proactive audio.
[6:04] So we really think we're
really bullish on voice
[6:07] as a modality to engage with
intelligence and with AI models.
[6:11] And so it's been a
big investment for us.
[6:15] On the generative
media side, we've
[6:16] got a full range of
models, everything
[6:18] from image generation models
to video generation models.
[6:21] Lyria is our music
generation model.
[6:23] So we just launched
Lyria three Pro.
[6:26] This allows you based on a
text prompt or even an image
[6:28] to generate a song up
to three minutes long.
[6:31] So a full song with full vocals.
[6:33] If you haven't tried this
out, I really encourage it.
[6:35] It's super cool.
[6:36] You can find it in the
Gemini consumer app,
[6:39] but we think that this
is going to support
[6:42] a range of different
types of applications
[6:44] where music is
really central to it.
[6:49] Just yesterday we announced
something called the Gemini Deep
[6:51] Research agent.
[6:52] And so one of the most popular
things in use cases for Gemini
[6:55] we've seen has been using it
for deep exploratory research.
[6:58] This could be for understanding
the market landscape
[7:01] around something.
[7:02] With our Deep Research agent.
[7:04] We now allow you with
a single API call,
[7:07] you can call Deep Research
to not only access
[7:10] the web and the broader
public information,
[7:12] but you can also ground it
on your own data sources
[7:14] to come back.
[7:16] And it not only will
come back with text,
[7:18] but also can generate charts
and graphics, infographics.
[7:21] And so we see a lot of workflows
where maybe the first step is
[7:24] you want to first
generate some research
[7:27] and understand the context
of what you're dealing with.
[7:29] And then from
there on might have
[7:31] subsequent steps in a workflow.
[7:33] But this is something that's
been really popular for us
[7:35] and we've been
really excited about.
[7:39] We're also investing in
World models at DeepMind.
[7:42] So we have a world
model called genie 3,
[7:43] which is our latest model.
[7:45] And again, you can if you're AI
Ultra subscriber in the Gemini
[7:48] app, you can try this out, but
starting with a text prompt
[7:51] or an image prompt, you can.
[7:54] That's some great music.
[7:55] Sounds like jump by Van Halen.
[7:58] Genie three allows you to create
a fully interactive 3D world
[8:02] based on a simple text
prompt or an image,
[8:05] and you can navigate this
world with your keyboard.
[8:07] You can decide, can.
[8:08] You can basically prompt the
model what type of character
[8:10] you want to be.
[8:11] And your world models are
very important for a number
[8:15] of different applications in AI.
[8:16] And we think here
you'll see applications
[8:18] in entertainment and
education, productivity,
[8:22] content generation, gaming.
[8:25] But we also think it's very
important for robotics,
[8:27] the ability to
simulate the real world
[8:29] and have agents learn how
to interact with that world.
[8:32] This is one of the things
we're most excited about.
[8:35] And so finally, I wanted to
mention Gemini robotics, which
[8:37] is our robotics model.
[8:38] So robotics is really where
physical world meets AI.
[8:42] And we've been investing heavily
in Gemini as a platform to power
[8:48] the next generation of robots.
[8:49] General purpose robots that can
reason through the world, that
[8:52] can use vision to understand
what they're seeing.
[8:56] And we've built the
ability to actuate
[8:58] different types of robots.
[9:00] So we launched our latest
robotics model last week.
[9:02] It's Gemini robotics.
[9:03] The ER stands for
embodied reasoning,
[9:05] but it allows you to
plug Gemini into a robot.
[9:08] And it can leverage all of
Gemini's reasoning capabilities
[9:11] and vision capabilities to power
different types of experiences.
[9:14] So we have worked with a
few different partners.
[9:17] And Boston Dynamics, which is
one of the leading hardware
[9:19] robotics companies,
is now using ER 1.6
[9:22] to power its spot robot, which
is a quadruped robot that
[9:26] can go around a facility
and investigate things.
[9:28] So with ER it's able
to count objects,
[9:31] it's able to read gauges
and things like that weren't
[9:35] possible before with robotics.
[9:37] So again, we're very excited
about the future of robotics.
[9:40] And this is just our latest
model in that offering.
[9:48] OK my last slide here.
[9:49] I just want to talk a
little bit about how we work
[9:51] with Google Cloud at DeepMind.
[9:53] We really want to see our
AI transform every industry
[9:55] and every geography.
[9:56] And we're bringing the same
models, same Google models
[9:59] that we use the
power Google products
[10:00] into our Cloud platform.
[10:02] We work very closely with
the Google Cloud team
[10:05] because we love to get feedback
from developers and customers
[10:07] and to see how these
models are deployed.
[10:09] We're trying to build
artificial general intelligence,
[10:11] and so having a wide variety
of diverse use cases for how
[10:14] these models are being used
is really helpful for us
[10:16] as we improve our models.
[10:18] And so Michael here, we
probably talk maybe once a day
[10:21] at least at our teams,
much more talking through.
[10:25] And we've been
really excited to see
[10:26] as soon as we launch a new model
in the Google Cloud platform,
[10:29] it gets adoption right away.
[10:31] And it's really been amazing to
see the different types of use
[10:33] cases.
[10:33] So now I'm going to
turn it over to Michael
[10:35] to talk a little bit about
how Google Cloud is deploying
[10:37] some of these capabilities.
[10:38] Thanks, David.
[10:42] So at Google we've been doing
this for a very long time.
[10:46] DeepMind pioneered
the technology,
[10:49] but we also have in Cloud.
[10:51] We have a global network
that was originally
[10:54] intended to get a cat video
from New York to Tokyo
[10:57] in 75 milliseconds, but that
accrues to the current moment,
[11:00] because if you are a
technician and you are working
[11:03] on an airplane and
you are asking a field
[11:05] manual for advice, you
need that answer very, very
[11:08] quickly as well.
[11:09] And you need to be
able to use your hands,
[11:11] and you need to be able
to talk to your agents
[11:13] rather than read that manual.
[11:16] We've also been working as a
hyperscaler on reliability,
[11:20] obviously, but there are
many other Google properties
[11:24] that accrue to this moment,
like Google Maps, like Street
[11:28] View, that allow us to
do multimodal reasoning
[11:31] over different kinds
of images as well.
[11:34] So the advantage of having 20
years of experience actually
[11:38] does accrue to the current
moment in surprising ways.
[11:46] So there are several
problems that I'm hearing
[11:51] in general in agent building.
[11:53] I've been launching
models for about two years
[11:56] since I was actually at one
of our partners, Anthropic,
[12:00] where we launched Cloud three.
[12:02] My first model at
Google was Gemini 3.
[12:06] And what we're hearing
is that even Gemini 2.5,
[12:09] Gemini 3 are far more
powerful than necessary to do
[12:13] a lot of enterprise workflows.
[12:15] It's still necessary
to use 3.1 to write
[12:17] the best code in the world.
[12:19] There are many
things that 3.1 is
[12:21] leading on the frontier
of intelligence.
[12:24] But for a bank's
KYC process, they've
[12:26] been using some of our models
for the past several months
[12:31] in order to accomplish
that goal, or some
[12:33] of our clinical research
customers in biotech, health,
[12:37] life sciences have been
doing statistical analysis
[12:39] on data sets, things like that.
[12:42] But not all of those customers
are doing those things.
[12:44] And we're seeing uptake
day by day by day.
[12:47] And that's because they needed
a platform that helped them
[12:50] implement common practices.
[12:51] The agents come with
intelligence built in.
[12:55] But that doesn't mean
they know everything
[12:56] about your organizations.
[12:58] They don't know the background
checks about your customers.
[13:00] If you're running
a KYC process, they
[13:02] don't know the locked database.
[13:04] After your clinical
study, if you
[13:06] are trying to submit to the
FDA and show drug efficacy.
[13:10] So the platform allows access
in scalable and flexible ways
[13:15] to that data.
[13:16] But contemplates credentialing
and data access and information
[13:21] retrieval as well.
[13:22] And finally, it has to be
built on trust in order
[13:25] to take the human
out of the loop
[13:27] and really take
advantage of Cloud scale,
[13:30] you also need to be able to
observe what the agent has done.
[13:33] You have to be able to audit
the decisions that it has made,
[13:37] the data it has
accessed, and what kind
[13:39] of synthetic generated data it
has generated and at what level
[13:42] it has classified that data for
future storage and retrieval.
[13:46] So it is only with
trusts in front and trust
[13:50] at the end of that pipeline
that the middle is going
[13:53] to scale without a
human in the loop,
[13:55] because if it's scaling
with a human in the loop,
[13:57] it's really only scaling
to the size of your teams
[13:59] and not the size of the
Cloud you see it especially
[14:02] in the coding realm,
the SDLC really
[14:06] does contemplate a lot
of those checks already,
[14:09] which is why you see such
an enormous amount of code
[14:12] authorship done in this point.
[14:14] But many of these enterprise
processes benefit as well.
[14:21] So I want to sorry, not
know how to use this thing.
[14:25] I want to move to a little
bit of a conversation rather
[14:28] than slideware.
[14:29] And I want to start with
something David said earlier.
[14:31] Our teams are talking
hundreds of times a day.
[14:33] We ship almost every model
that Cloud application
[14:38] is going to use, maybe with the
exception of embodied robotics,
[14:43] but I want to ask
you, how do you
[14:45] interact with Cloud customers.
[14:47] How does our relationship
not necessarily
[14:50] make Cloud better, but how
does it make our models better.
[14:52] How does it make them smarter,
faster, and more efficient.
[14:57] Yeah, it's pretty amazing
to see that there's now
[14:59] millions of developers that
are building on Gemini.
[15:02] And so with the
Google Cloud team,
[15:03] when we're launching new models,
when we're building new models,
[15:07] we'll oftentimes, run a public
preview or a private preview,
[15:11] sometimes with a set of Cloud
customers to get feedback.
[15:14] And we've selected a diversity
of different types of customers,
[15:17] different types of use cases,
everything from startups
[15:20] to large enterprises.
[15:21] We'll use a lot of that
feedback to understand
[15:23] what's working in the model,
what's not working in the model.
[15:26] And we use that to select which
models we're going to launch.
[15:28] We also use it to improve
specific capabilities
[15:31] of the model.
[15:32] So it's been super helpful
to have that feedback.
[15:34] I think it's great to have
all the Google products using
[15:38] Gemini, but the Google teams
are kind of a captive audience.
[15:42] They have to use our models.
[15:44] And so what's great
about Cloud customers
[15:45] is you all have choices.
[15:46] You're looking at a variety
of different solutions.
[15:48] And so we for you to
choose our models,
[15:51] we really have to be the
best at what you're trying
[15:53] to accomplish with the model.
[15:54] So it really pushes us to
make the models better.
[15:57] And we've gotten
some great feedback
[15:59] from some of these
deep partnerships
[16:01] we've had with Cloud.
[16:03] That's great.
[16:07] So you really recently
released 3.1 Pro.
[16:10] We get into discussions
about Hill climbing,
[16:12] about making numbers go up.
[16:13] There are areas where we
know that good enough is only
[16:17] good enough.
[16:18] But there are also areas where
we focus uniquely and think
[16:23] of us as having specific
advantages in our models.
[16:27] How do you think about some
of those for the audience,
[16:30] where do you focus.
[16:30] Where do you think
other labs might not.
[16:34] Yeah, I think in
terms of our models,
[16:36] one of the advantages
we have at Google
[16:37] is we're a completely
full stack.
[16:39] So we have everything
from our own TPUs,
[16:41] which you heard
about earlier today.
[16:42] And all the innovations
we're doing there.
[16:44] And the DeepMind team works
very closely with our.
[16:47] Team working and
building our chips.
[16:49] We also work very closely
with the Cloud teams,
[16:51] building our next
generation AI data centers
[16:53] to make sure those things
are really well suited
[16:55] to not only train
our models, but also
[16:57] run our models at scale.
[17:00] The other thing that we have
is with Google products,
[17:02] our consumer products, some of
our products like Google Search,
[17:06] which has been completely
reinvented with Gemini.
[17:08] I mean, if you see
Google Search now,
[17:10] it's got AI Overviews,
AI Mode, and we're
[17:12] running that at over 2
billion plus user scale.
[17:16] So it's the most popular
generative AI product
[17:18] in the world right now.
[17:19] And to do that and to be able
to serve at very low latency
[17:22] to billions of
consumers, we have
[17:24] to make these models
and the entire stack
[17:26] incredibly efficient.
[17:27] So that's one of the
things we've done.
[17:28] And you'll see our model,
our flashlight model, which
[17:32] is our smallest Gemini model.
[17:33] That's the reason it's
incredibly efficient.
[17:35] Because we've really
tuned it over the years
[17:37] to work well at massive Google
scale, and we can pass on those
[17:40] benefits to you.
[17:41] I think a couple of other
areas where we really
[17:44] focus multimodality.
[17:45] Again, when DeepMind started and
when the Gemini project started,
[17:51] we wanted it to be natively
multimodal from the outset.
[17:54] And part of the
reason for that is
[17:55] when you're trying
to build AGI, we're
[17:57] trying to build an
AI system that's
[17:58] going to pattern on how the
human brain works, and is
[18:00] able to do the types
of cognitive tasks
[18:02] that a human can do.
[18:03] And so it's really important
for it to be multimodal.
[18:06] So not only on understanding.
[18:07] And if you haven't tried this
you should certainly try.
[18:10] You can upload a
video to Gemini.
[18:12] We could upload a video of
us speaking on this panel.
[18:14] And we could ask
it to critique us,
[18:16] and it would do a
pretty incredible job
[18:18] and hopefully be
pretty flattering.
[18:19] But those are the
types of things
[18:21] you can do with the multimodal
understanding capabilities.
[18:24] But we've also invested
quite heavily in generation.
[18:26] So to generate images and
videos and we talked about music
[18:29] earlier.
[18:30] So some of these are some
of the areas where I think
[18:33] we think really Gemini shines.
[18:35] Yeah I love certain
areas where we're
[18:37] able to combine
these things, where
[18:39] we can do a grounding
with Google Search
[18:41] and determine where you are.
[18:42] Or we can look up the
weather for that location,
[18:44] but then we can
produce a video of you
[18:48] in the rain to
show you what it's
[18:49] going to be like for that day.
[18:51] And so pipelining and
combining these models
[18:53] is something that we
see very commonly.
[18:56] Now I get to ask you
a couple questions.
[18:58] Then we're going to
hear from Replit.
[19:00] So one of the things with
these models are amazing.
[19:04] And helping the software
engineering process
[19:06] and the software
development lifecycle,
[19:08] it's really speeding up things.
[19:09] So how do you when you're
working with Cloud customers,
[19:11] how do you help them understand
not only what to build
[19:15] but what not to build.
[19:16] Because in an era where it's
becoming much easier with AI
[19:19] code generation to build
pretty much anything, how do
[19:22] you think about that.
[19:24] That's a really good question.
[19:25] That's a really
fun one, actually.
[19:27] I was at a company called
Datadog for seven years,
[19:30] and some of you probably
have used my product.
[19:33] One of the things I liked
there was that monitoring
[19:36] Kubernetes and most
of my customers
[19:38] looked like
monitoring Kubernetes,
[19:39] whether it was Capital
One or Airbnb or whatever.
[19:42] It's pods and
clusters and that sort
[19:44] of thing working
with customers on AI
[19:47] is extremely mission oriented.
[19:49] When you work with somebody
like Thomson Reuters,
[19:52] you're talking
about points of law,
[19:53] you're talking about
clauses in a contract
[19:55] and whether or
not they're better
[19:57] for the advocate
or the adversary.
[19:59] When you're working
with a large bank,
[20:01] you're talking about their
KYC process and anti-money
[20:03] laundering.
[20:04] When I talk to most of
you, most of the customers,
[20:09] there's no easy answer for
what we should build with AI.
[20:13] A lot of where I
do focus is what
[20:16] do you know better than
everybody else in the world.
[20:19] And how can we help
you trust the model.
[20:22] What I have seen.
[20:23] Well, there are three
boundary conditions
[20:25] that I've seen on the models.
[20:27] One is this frontier
of intelligence.
[20:29] The best code in the world
that is being written
[20:32] is really the least bad it's
ever going to be at this point.
[20:34] But we all know
that we try to write
[20:36] better and better
and better code when
[20:38] new and new models come out.
[20:40] But if you are
looking at a catalog
[20:42] and you want to determine if
the green shares are green
[20:44] and you're extracting metadata
that only gets so good,
[20:47] the green shares
don't get any greener.
[20:49] And so that doesn't require
the frontier of intelligence.
[20:52] But if you're a 911 operator,
if you are a DoorDash delivery
[20:56] person and you are
trying to figure out
[20:58] how to apply a policy in real
time, in an urgent situation,
[21:02] you become very latency bound
very, very, very quickly.
[21:05] And this is where
Gemini Flash shines,
[21:08] because it brings the most
intelligence to bear possible
[21:10] within a fixed time budget.
[21:13] And then lastly, we were talking
about flashlight a moment ago.
[21:15] If you want to moderate
the entire internet,
[21:17] you're one of these platforms
that I won't necessarily name.
[21:20] You don't know how many
posts there will be
[21:22] don't know how bad
they are going to be.
[21:24] And no matter how
big your budget
[21:25] is have to contemplate an
infinite volume and flashlight
[21:28] because it was designed
for Google Search
[21:30] contemplates an infinite volume.
[21:32] So we usually take that
lens with our customers
[21:34] and understand what
their domains are,
[21:37] and help them select the
right tool for the job,
[21:39] whether or not they are doing
something that is incredibly
[21:42] hard, like validating a
contract or something that
[21:45] happens at scale, like
moderating the internet.
[21:47] We work hand in hand with them.
[21:49] The other thing I will say is
that the models are smarter than
[21:53] me in everything that I'm not
like 18 years of experience
[21:57] done.
[21:58] And I think that
we'll most of us will
[22:00] find that to be the case.
[22:01] So the other way
we help customers
[22:04] is encouraging them not to
tell the models what to do,
[22:07] how you think they should
solve their problem,
[22:09] but really to give
them the context
[22:11] that is necessary to solve the
problem and trust that they
[22:14] will.
[22:14] As long as you tell
them what your goal is,
[22:18] your strategy, what the context
of your business might be just
[22:21] like hiring somebody.
[22:22] When you hire a brilliant intern
or not even an intern, a new
[22:26] hire from Princeton with one
year of experience or something,
[22:29] they're going to
pick up what you're
[22:31] putting down pretty quickly.
[22:33] But that doesn't mean that
if they came on to Datadog
[22:35] on day one, they know how to
do time series arithmetic.
[22:37] So they're not going to do a
very good job if you put them
[22:39] in front of a customer.
[22:40] Without that
training, you're also
[22:42] not telling them exactly
how to help a customer
[22:44] in every single situation.
[22:45] That's almost impossible.
[22:46] You're trusting that they will
figure that out in the field
[22:49] and get better at it as
they get more experienced.
[22:51] And the models and
especially agents.
[22:53] The models are
mutable, obviously,
[22:54] but the agents are entirely
capable of that kind
[22:57] of self-improvement.
[22:58] And so the platform
helps with that as well.
[23:01] And then related
to that, I mean,
[23:02] earlier today, Google Cloud
announced the Google Enterprise
[23:06] Agent platform in the keynote.
[23:08] And so can you tell
us a little bit
[23:10] about that, how you're
seeing customers
[23:12] deploy agents, what are
the most common use cases,
[23:14] and any advice for building
and scaling with that platform.
[23:17] So choosing a model is just
one part of building an agent.
[23:20] And that part is very important.
[23:21] But the agent platform
contemplates much more
[23:24] than the model itself.
[23:27] We separate this into build,
scale, govern and optimize.
[23:32] Build and scale.
[23:33] We've talked about a
lot for many years.
[23:35] We offer we offer an engine.
[23:40] These are places that allow
you to build your harnesses
[23:43] to deploy your harnesses
and give you an easy place
[23:46] to plumb observability
into those harnesses.
[23:48] The very important parts
that we're adding today
[23:52] especially are
govern and optimize.
[23:54] And those two things
go hand in hand.
[23:56] We are able to watch what your
agents are doing in real time.
[24:00] And like I was
saying, your agents
[24:01] should self-improve even
though the models themselves
[24:05] are immutable, you should be
able to see in your stack traces
[24:10] how they are behaving
in real time.
[24:12] And that might
mean stopping them
[24:14] from doing something
because they've
[24:15] breached a policy that you
don't want them to breach,
[24:17] or more commonly, giving them
examples of what they have done
[24:21] well or poorly in the
past, and putting that
[24:24] back into the context window.
[24:25] Context engineering
isn't a process
[24:28] of writing a lot of
English with your hands.
[24:31] It's really just throwing a
trace into the context window
[24:34] and say, don't do
that ever again.
[24:36] And it will
extrapolate from there,
[24:38] and it'll get better and
better and better from there.
[24:40] And so that flywheel will both
protect you, your customers
[24:44] and your businesses,
but it will also
[24:46] increase the productivity
of those agents
[24:48] for your businesses.
[24:51] So I think we should
turn it over to Michele
[24:54] at Replit to talk
about how they're
[24:56] using the Gemini Enterprise
Agent platform and your business
[25:00] in general.
[25:01] Yeah so as I mentioned
before, rabbit
[25:04] is a coding platform that
really democratizes access
[25:07] to software creation
for everyone.
[25:09] And we've been using
Gemini basically
[25:11] since the inception
of the project.
[25:14] And I think I want to give you
a public praise about really
[25:18] understanding that intelligence
is not going to be just one
[25:21] size fits all package.
[25:23] You understood from
day one that there
[25:25] should be a powerful model
that really pushes the frontier
[25:28] and at the same time.
[25:29] Flash and Flash light are
like Swiss army knives
[25:32] that you can deploy
everywhere in product
[25:34] and add that touch
of intelligence that
[25:36] makes the user experience
way more delightful,
[25:38] but really doesn't
break the bank
[25:40] and doesn't impact the
budget for the users.
[25:43] And the other advantage that
we experience on a daily basis
[25:46] is that rapid is the fact
that Gemini is often Vertex,
[25:50] and there is beauty in having
everything in one single place,
[25:53] and our entire
infrastructure runs on GCP.
[25:55] We get access to intelligence
on Vertex as well.
[25:58] So that makes the
custom maintenance.
[26:00] And in general,
the work of our SRE
[26:02] is much easier compared to
spreading workloads across
[26:05] to many providers.
[26:07] One of the I mean, what
I love about Replit
[26:10] is that you really created an
entire market of engineers.
[26:16] It's not just engineers who
know the word Kubernetes, who
[26:18] hold a pager that wake up
at o'clock o'clock o'clock
[26:20] o'clock in the morning
that are engineers.
[26:21] Now, we're all engineers now.
[26:22] Yeah they don't even
notice is exactly.
[26:25] Exactly you don't.
[26:25] You don't need to
know what Kubernetes
[26:27] is to be an engineer anymore.
[26:28] Yeah it's a new
market, basically.
[26:30] We're charting together.
[26:32] And if you think
about it, for decades,
[26:35] we focused on making $30, $40,
50 million software developers
[26:38] in the world more productive.
[26:40] And now suddenly, the
aperture is much wider.
[26:42] We're targeting basically
every knowledge worker
[26:44] in the world,
probably 1.5 billion,
[26:46] two billion people in the world.
[26:48] So very exciting
journey ahead for us.
[26:51] Yeah, that's very exciting.
[26:52] And how do you think
about how it should evolve
[26:55] as the nature of the software
development role evolves,
[26:58] right from writing code
to more managing agents,
[27:01] and then maybe
managing agents that
[27:02] are managing other subagents.
[27:04] How are you thinking
about that at replit?
[27:06] We are really at this stage.
[27:08] We launch Agentforce back
in early March this year.
[27:12] And what we try to
do is rethink what
[27:16] is the driver shift
of the product.
[27:18] You're not there anymore.
[27:20] Just thinking as a builder.
[27:21] We actually empower you to
become a project manager.
[27:24] So you spend quite a
lot of time planning,
[27:27] and then we spin up
agents in the ground that
[27:29] accomplish the job.
[27:30] So even though the product
hides most of the complexity,
[27:34] in reality, our users as of
today are managers of agents
[27:37] already.
[27:38] And I think that's
the key of entrance
[27:41] for non-technical users.
[27:43] They need to be able to unleash
the power of several agents
[27:47] running in parallel,
but they don't
[27:48] want to have the
cognitive burden
[27:50] of having to manage
tens or hundreds of them
[27:52] is way too much.
[27:53] Even see developers
running several shells
[27:56] in parallel, several ideas.
[27:58] That is something that I
think will be left only
[28:00] for the top coders in the
world, and everyone else
[28:02] will be just running
agents in the ground
[28:04] without even realizing that.
[28:06] Yeah, great.
[28:07] So dive a little
bit deeper, I guess
[28:09] on how Replit layers on the
underlying LLM or LLMs as you
[28:15] choose them and provides
dramatically more value
[28:18] to these engineers.
[28:21] Yeah, I think it's
very important
[28:25] to put focus on the application
layer on top of models.
[28:30] It's very exciting.
[28:31] The race that we're experiencing
in technology today,
[28:33] I think it's unprecedented how
fast the frontier is moving,
[28:36] but at the same time,
that is raw power.
[28:39] And if you put raw power in
the hands of a power user.
[28:42] They come up, they tailor
their own workflow,
[28:45] they fund the different products
that make their life easier.
[28:48] But when you're targeting
a knowledge worker which
[28:50] just wants to get the job
done, what they're looking for
[28:53] is a platform that allows
you to get anything that you
[28:57] have in mind accomplished.
[28:58] So over time, we're starting to
think even less about a software
[29:03] centric view for rapid is more
a place where you can accomplish
[29:07] a knowledge worker
task, and if that
[29:09] is by means of creating an
application, by all means,
[29:12] we are also creating agents.
[29:14] We are creating slide decks,
we're creating motion videos.
[29:18] Ultimately, we
want to empower you
[29:20] to be creative by means of
creating software in the touring
[29:23] background, but the entire
complexity of writing
[29:27] code, maintaining code,
publishing applications
[29:30] that is completely hidden.
[29:32] And I think what I'd love
to talk about in public,
[29:35] because I don't think many
users understand that it's
[29:38] the same level of quality
you get from a Google Cloud
[29:41] deployment when you
build anything on it.
[29:43] So we're using
Cloud Run and we're
[29:45] using Google Cloud Storage.
[29:46] So even though you're building
an MVP or your prototype,
[29:50] you're already able to scale
to millions of users overnight.
[29:53] That's amazing.
[29:54] So you delegate the
authorship to Replit
[29:58] the maintenance and the DevOps.
[30:00] You also take care
for your customers.
[30:02] What's the software
development life like.
[30:04] What is building software
look like in the future.
[30:07] I don't think we're
going to add any more
[30:10] focus on the SDLC or the
software development lifecycle.
[30:15] And the reason being, if you
spend some time on rabbit,
[30:19] the cycle is literally two hops.
[30:21] You create the software
and then you republish it.
[30:25] Everything else is
happening in the background.
[30:27] We have a prod agent.
[30:29] So it's an agent
that basically pulls
[30:32] all the time your deployments
checks if your queries are fast
[30:35] enough, checks if there
are errors in your logs,
[30:38] and then recommends you how to
fix the application, we just
[30:41] launching today a way to keep
track of all your software
[30:45] dependencies in
terms of security.
[30:47] So if there are vulnerabilities
just published through a CV,
[30:50] we immediately warn you about
it and we give you a task
[30:53] that the agent can accomplish.
[30:55] The only thing you need to do is
to accept it and then republish
[30:57] the application.
[30:58] So the cycle that we
studied in college,
[31:01] several different steps.
[31:03] You write your requirements and
then you create the software
[31:06] and write tests I
think is all being
[31:08] compressed into this
creation experience
[31:11] that you have on Replit.
[31:12] And then when you're ready to
go publish the next version
[31:15] of the app.
[31:15] That's amazing.
[31:17] Can you talk a little
bit more about how
[31:18] you're orchestrating the
different models you mentioned.
[31:21] You pick the right size
model for the right size job.
[31:24] What goes into that.
[31:24] And then also is there are
using any open models at all
[31:28] in terms of your mix.
[31:32] When we put our hands
first on the early versions
[31:34] of Flash and flashlight,
a lot of the testing
[31:38] was based on vibes.
[31:39] So we really try to
understand how far can we
[31:42] push these models.
[31:43] And we've been
constantly impressed
[31:45] by how powerful they were.
[31:47] If you think about
it, a Flash model
[31:49] is usually six months behind
over the frontier was.
[31:53] So if you use
Flash today, you're
[31:55] basically using a way more
expensive model from six months
[31:58] ago in the frontier.
[31:59] That really forces you to
rethink your product roadmap.
[32:03] There will be things today
that maybe are not affordable,
[32:06] especially for our long tail
of users across the world.
[32:09] But six months from
now, we're going
[32:11] to be able to put them
in the hands of everyone.
[32:13] So that is why I'm very excited
about the smaller models.
[32:17] The way we test them today is
far more rigorous than when just
[32:23] flesh came out.
[32:24] We have a lot of
online evaluations.
[32:26] So every single day there
are several versions
[32:28] of our agent in
front of everyone.
[32:30] We do a A/B testing,
as you can imagine.
[32:33] And we try to understand
which are the tasks, which
[32:37] are the accessory functions
that the agent can accomplish
[32:40] with lower level
of intelligence.
[32:42] And we keep tuning that.
[32:43] And it's a constant cycle.
[32:44] Basically on a nightly
basis we redeploy
[32:46] a new agent for everyone.
[32:48] Got it.
[32:49] You use the word A/B test.
[32:50] And I'm very curious because
you have to support basically
[32:52] any app that could be written.
[32:54] How do you do evals and end
tests and which part do you own
[32:58] and which part do you
delegate to your users
[33:01] to evaluate for themselves.
[33:04] So we definitely
put a lot of effort
[33:05] in creating offline evaluations.
[33:07] And we also collaborate
with the Gemini team.
[33:09] We give them insights of better
models running on our evals.
[33:14] At the same time, a lot of
effort goes on online evals.
[33:17] And you're absolutely right.
[33:19] A lot of the applications are
the first time we actually
[33:23] see them being built on ramp.
[33:24] It's very similar to
the queries on Google
[33:27] Search back in the days.
[33:28] I think half of them are unique
queries that hit the platform.
[33:31] It's similar for us.
[33:33] We can't really evaluate
if a certain app has
[33:36] been built successfully,
but what we can do
[33:38] is we can track
the user sentiment.
[33:40] We can track if
they keep engaging
[33:42] with a certain application.
[33:43] And we've been seeing the
success rate and the sentiment
[33:47] of the user, like
climbing steadily
[33:49] over the last year and a half,
partially thanks to models
[33:52] becoming more powerful,
but also thanks to the fact
[33:55] that our platform is
far more exhaustive.
[33:57] Our agent is way more
advanced than it was before.
[34:00] And I expect these numbers
to start to plateau maybe
[34:04] by next year,
because the truth is,
[34:06] a lot of the software that
is created on rapid or on web
[34:10] coding platforms
is not going to be
[34:12] as advanced as maybe a
database built on Google Cloud.
[34:16] It's relatively simple software.
[34:19] There are play books
that can be followed.
[34:21] Models are becoming
extremely good at them,
[34:23] and I'm very excited about
literally the next 12
[34:26] months, because we're probably
going to see a peak in terms
[34:28] of what coders can accomplish.
[34:30] Very exciting.
[34:33] I think we're are
we out of time.
[34:36] I have so many more questions.
[34:38] One more.
[34:39] Well, I am wondering, I guess,
for everybody building things.
[34:42] You mentioned that about
half of the applications
[34:44] are unique snowflakes.
[34:45] What are the common
patterns that people
[34:47] are building, either
for their enterprises
[34:49] or as weekend Warriors.
[34:54] I think the day one task that
everyone does on our platform
[34:57] is they create a
prototype of an idea
[34:59] that they have in their
mind, and they go,
[35:01] they do that 0 to
one really quickly.
[35:04] What makes it different is
that they're not just creating
[35:06] a front end or a landing page.
[35:08] They can actually go in
depth and create a functional
[35:10] prototype.
[35:11] I think there are
two paths from there.
[35:13] One is you show up at a
meeting with decision makers,
[35:17] and stakeholders.
[35:18] Suddenly you're not there with a
slide deck or a six page essay.
[35:22] You share a URL.
[35:24] They check the prototype, and
then they test your product idea
[35:28] and they give a thumbs
up or thumbs down.
[35:31] The second one that
we're also very excited
[35:33] and I think is really changing
the scenario of how SaaS
[35:36] is going to evolve
over time is the fact
[35:38] that a lot of internal tools
today are built on rabbit.
[35:41] Yeah, of course, we are big,
the founders of our own product.
[35:44] But I can tell if you
shop rabbit today,
[35:47] maybe 80% of the software
that they use on a daily basis
[35:50] is built on rabbit.
[35:51] So all our internal tools
ranging from HR sales finance,
[35:56] everything is built
on our platform
[35:57] and we see this happening
across the entire industry.
[35:59] That's cool.
[36:01] Thank you for being a
customer of our Gemini models
[36:04] and providing really
valuable feedback.
[36:05] It's been super
helpful for our teams.
[36:07] Thank you.