Page 78 - AI Standards for Global Impact: From Governance to Action

AI Standards for Global Impact: From Governance to Action

The main points of discussions were:

1) Data access:

a) The importance of data collection and making data available online for solving
problems in the "AI for network" and "network for AI" domains. It was pointed out that
model benchmarks along with data access are critical for developing future networks
with AI.
b) Data generation could potentially be promoted with the co-generation of data from
academia and industry, with ITU forming a neutral entity to share datasets.
c) Validation and publication of trustworthy datasets along with evaluation metrics.

2) Compute access:
a) The importance of widespread availability of GPU resources for compute.

3) Importance of collaboration

a) Potential relevance and collaboration with ITU standardization groups working on trust.
b) Potential collaboration with bodies such as AI-RAN Alliance.
c) Academic partnerships and papers which analyse relevant datasets.
d) Open source as a mechanism to accelerate the standards with potential feedback
helping to refine specifications.
e) Continued open-source Build-a-Thons and knowledge base generation.
f) Providing space for innovation, especially when releases of standards place constraints
on innovations which are more than incremental.
4) Toolkits

a) The importance of agents and related API toolkits and potential for new standards in
this domain.
b) Inference as a service, especially at the edge, and collaborative inference with network
capabilities and hosted models.

11�2 Innovations in AI models

The session explored recent innovations in AI models for telecom networks. Speakers covered
topics including federated learning, telecom-specific LLMs, agent-based systems, and AI-native
architectures. A common theme was making AI more efficient, adaptable, and suitable for real-
time deployment. The session emphasized practical tools and frameworks that move AI closer
to deployment in next-generation communication networks.

The main points discussed are summarised below:

• The journey from 5G to 6G was used as an example within the context of AI integration. It
was highlighted that AI is no longer a supportive tool but a central enabler of intelligent,
autonomous, and adaptive mobile networks. Moreover, the convergence of AI with
telecom at the architectural level calls for scalable and standardized approaches to
embed intelligence throughout the RAN, core, and edge layers. Some of the key issues
of integrating AI in 6G were also highlighted, such as using AI as an add-on rather than
a full integration across the network, the need for a unified architecture supporting both
AI for Network and Network for AI, current limited support for advanced AI technologies
such as generative AI and federated learning, cross-layer agent deployments, and efficient
data pipeline management.
• A framework for federated learning using tiny language models to predict cellular features
and the use of the NNCodec (Fraunhofer Neural Network Encoder/Decoder) for neural

66