Page 33 - AI Governance Day - From Principles to Implementation

AI Governance Day - From Principles to Implementation

Where technology lags behind AI governance

In the realm of AI governance, several areas reveal a significant lag in technological advancements
needed to meet regulatory requirements. These gaps span across data management,
computational oversight, model integrity, and deployment practices, highlighting the urgent
need for innovation and robust methodologies to support effective governance.
• Data

– Methods are lacking to clearly define and implement fairness in AI systems.
– There are no robust ways to fully safeguard privacy, especially in large language
models.
– It is difficult for users to know if their data was used to train a model and its influence
on model behavior.
• Compute

– There are only nascent methods for verifying if a large model is being trained in order
to enforce compute-based reporting requirements.

• Model
– Comprehensive and reliable methods to evaluate AI systems for safety, ethics, and
reliability are lacking.
– Securing model weights from unauthorized access is challenging.
• Deployment

– Approaches to constrain model outputs to certain values are imperfect and can be
circumvented.
– Detecting AI-generated content, especially non-watermarked content, remains
difficult.

4.4.2 Compute governance

"Compute" or "computing power" (often used interchangeably) refers to the necessary
computational resources required to run software. Compute governance is arguably an area
where measurements can be more straightforward and quantifiable.

The training process of AI models requires intensive compute, i.e. computing resources. and
training a leading AI model tends to take months on a complex computing infrastructure
involving specialized computer chips. In contrast, running a single inference query (e.g. having
an AI model respond to a single question) requires much less compute, but the total amount
of compute used for inference is still very large, since large AI companies need to run millions
of user queries per day.

Compute is essential not only for training AI models, but also for deploying (i.e. operating)
them. Just as operating expenses outpace initial fixed costs for many large-scale projects, the
majority of available AI compute resources are used for operating AI models rather than training
them. Frontier AI models are so large that they cannot be efficiently operated at large-scale
with household amounts of typical consumer hardware. Instead, for models in high demand,
inference requires thousands of AI chips housed in specialized data centers to adequately serve
the needs of thousands of users. The wider the deployment of AI systems (which requires more
compute), the more impact they will likely have (both beneficial and harmful).

23