Skip to content

2023

Responsible AI needs distributed ownership

Microsoft's chief responsible artificial intelligence officer has published a reflection on the company's program, emphasizing leadership commitment, inclusive governance, and actionable standards. The timing is useful. Generative systems are moving into products rapidly, and many organizations are discovering that an ethics statement does not tell a product team what to do on Tuesday afternoon.

Responsible artificial intelligence (AI) needs a central function. It cannot remain the central function's job alone.

Guardrails are architecture, not decoration

NVIDIA has released NeMo Guardrails, an open-source toolkit for adding programmable controls to applications built with large language models. The toolkit is intended to help developers keep conversations on topic, reduce harmful or inaccurate responses, and restrict unsafe connections to external applications.

The word guardrail can sound like a thin layer around an otherwise complete product. In a consequential artificial intelligence system, it should be treated as architecture.

Research integration is an organizational-design problem

Google has combined DeepMind and the Google Brain team into a new unit called Google DeepMind. The stated ambition is to bring together talent, computing resources, infrastructure, and research advances to accelerate progress in artificial intelligence.

Mergers of technical groups are often described as exercises in scale. Their success depends just as much on whether distinct communities can combine knowledge without losing the differences that made each one valuable.

Model choice creates a portfolio to govern

Amazon Web Services (AWS) has announced Amazon Bedrock, a managed service intended to give customers access to foundation models from several providers through a common cloud environment. The appeal is obvious: teams can experiment with different models and choose the capability that fits the application without building the underlying infrastructure themselves.

Choice reduces dependence on a single model. It also creates a portfolio that somebody has to govern.

A framework needs a community of practice

The National Institute of Standards and Technology (NIST) has launched its Trustworthy and Responsible Artificial Intelligence Resource Center, known as the Artificial Intelligence Resource Center (AIRC). The new site gathers the Artificial Intelligence Risk Management Framework (AI RMF), its playbook, crosswalks, and implementation resources in one place.

Central access is useful. The larger opportunity is to create a place where organizations learn how the framework behaves in practice.

Tool-using AI needs explicit authority

OpenAI has begun introducing plugins that let ChatGPT retrieve current information, run computations, and interact with external services. The early examples include browsing, code execution, travel, shopping, and other application connections.

This changes the nature of the system. A model that produces text can mislead. A model connected to tools can also act.

GPT-4 raises the standard for deployment evidence

OpenAI has released Generative Pre-trained Transformer 4 (GPT-4), a multimodal model that accepts image and text inputs and produces text. The accompanying technical report and system card describe strong performance across professional and academic benchmarks, alongside familiar limitations: unreliable facts, reasoning errors, bias, and behavior that can be difficult to characterize completely.

The release offers more than a new capability. It offers a useful distinction between evidence about a model and assurance about a deployed system.

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed