Google DeepMind Discusses Open Models & AI Ownership

Google DeepMind researchers Gus Martins and Ian Ballantyne recently explored the critical topic of AI ownership and the role of open models in achieving it. In their presentation, titled "Sovereign Escape Velocity: Ownership w Open Models," they detailed Google DeepMind's latest advancements with the Gemma family of models, emphasizing how these open-source solutions empower developers and enterprises to maintain greater control over their AI deployments.

Understanding Gemma models

Martins and Ballantyne introduced the Gemma family, which comprises models of varying sizes, including the 2B, 4B, 26B A4B, and 31B Dense variants. They explained that these models are designed to offer a balance of performance and efficiency, with the smaller versions being suitable for personal devices and NPU/mobile hardware, while larger models can be deployed on desktop or single-GPU setups. The concept of "effective parameter efficiency" was highlighted, suggesting that these models deliver strong performance relative to their size, often outperforming larger proprietary models on targeted tasks.

The Drive for AI Ownership

A central theme of the discussion was the burgeoning need for AI ownership, especially among governments and enterprises. Ballantyne elaborated on how relying solely on proprietary cloud APIs for AI services can create significant operational liabilities, particularly concerning data privacy and the potential for service disruptions or changes in terms of use. Open models, conversely, offer a path to greater control, allowing users to deploy, customize, and fine-tune models on their own infrastructure. This "sovereign escape velocity" refers to the ability of organizations to achieve a degree of autonomy in their AI capabilities, independent of external service providers.

Feasibility and Cost Realities

The presenters touched upon the practical considerations for adopting open models, outlining feasibility criteria such as task accuracy, hardware feasibility, and acceptable performance. They acknowledged the cost realities involved, which include upfront hardware investment and ongoing running costs, but emphasized that these can be balanced against the long-term benefits of ownership and the avoidance of pay-per-token models. The ability to run models on local hardware also addresses concerns about data security and latency, crucial for many sensitive applications.

Personal and Edge AI Deployments

The discussion then shifted to the practical applications of these models, particularly in personal and edge computing scenarios. Martins demonstrated how models like Gemma can be run on mobile devices and desktops, enabling local data processing and reducing reliance on cloud connectivity. The concept of "battery priority" was raised, noting that for mobile devices, power utilization is a critical cost factor, often more so than raw token generation costs. The ability to deploy efficient models on these constrained devices is therefore paramount.

Enterprise Consolidation and Fine-Tuning

For enterprise use cases, the trend is moving towards consolidating AI workloads from large cluster-scale setups to more manageable, consolidated nodes. This approach, typically involving 1-2 GPUs, not only reduces serving costs but also simplifies the physical network overhead. Furthermore, specialized tasks can benefit from targeted fine-tuning, such as the development of domain-specific variants like MedGemma for medical applications. The availability of open models under permissive licenses, like Apache 2.0, is key to enabling this enterprise adoption and customization.

Best Practices and Next Steps

Martins and Ballantyne concluded by outlining best practices for leveraging open models. These include:

• Drop-in Evaluation: Using local engines or lightweight serving with standard API interfaces to quickly assess model capabilities.
• Custom Evaluation: Building robust, cross-platform evaluation suites tailored to specific domain tasks.
• Serving Management: Factoring in the operational realities of serving, including downtime, and managing the underlying hardware (GPUs, NPUs).
• Enterprise Scaling: Scaling from local prototypes to production-grade serving, often utilizing managed platforms.

The overall message conveyed was that open models like Gemma are democratizing access to powerful AI capabilities, offering a flexible and controllable path for innovation across a wide range of applications and industries.