General purpose AI models

Understanding the EU AI Act

Dan Nechita

Lead Technical Negotiator, EU AI Act

Systems versus models

AI systems

Finite AI products
Deployment ready
Specific inputs and outputs

image 201.png

AI models

Brains behind the systems
GPT-3.5, GPT-4, and GPT-4o

Untitled design.png

General purpose AI models

The AI Act concerns most AI systems, but it also has a distinct set of rules for powerful general purpose AI models.

Potential to power millions of AI systems
Risk of significant impact if models malfunction
Defined as "systemic risks" by the AI Act

broken chain.png

Size and scope

Classification metrics in AI:

Number of Model Parameters, approximates size
Number of Floating-Point Operations (FLOPs), approximates performance

EU AI Act's use of metrics:

Determines obligation levels for GPAI model developers

¹ Data source: Epoch (2024)

Obligations regarding all GPAI models

General Purpose AI Models:

Models with over one billion parameters
Simplified as 'low-risk GPAI models'

Provider obligations:

Light-touch obligations
Ensure downstream safety
Compliance with European law

Obligations regarding all GPAI models

Prepare and maintain technical documentation on model training and evaluation
Provide integrators with documentation for safe model integration
Respect copyright and summarize training content
Summarize content used for model training

Systemic risk

CleanShot 2024-07-30 at 16.43.33.png

Some models may become so powerful that they can cause widespread issues and create systemic risk.

The threshold for systemic risk and GPAI is defined as 10^25 floating point operations, roughly 10x more powerful than GPT4.

GPAI obligations

Obligations:

Immediately notify the European Commission
Perform Model Evaluations
Report any serious incidents
Ensure an adequate level of cybersecurity

CleanShot 2024-07-30 at 16.49.27.png

Let's practice!

Understanding the EU AI Act

Preparing Video For Download...