Artificial Intelligence
1
2
3
4
5
Download this chapter (PDF)
From large to small language models

After the revolutionary introduction of LLMs, there is growing interest in Small Language Models (SLMs). These are models with up to 10 billion parameters, in contrast with LLMs, which can have hundreds of billions or even trillions of parameters.

‍
Training and using LLMs requires enormous amounts of computational resources. In contrast, SLMs are significantly smaller. Therefore, they are less intensive in terms of data processing, hardware, and training time requirements. SLMs also consume less energy, making them more suitable for applications on smaller devices.

‍
SLMs are more accessible to users who want to train and run these models on consumer hardware at the edge of a network, especially for single-purpose devices (e.g. sensors). In addition, SLMs are particularly useful for specific tasks rather than for use as general-purpose tools.

Maturity

WATCH
PLAN
ACT
Show the other PLAN trends
Changing dynamics in Responsible AI From large to small language models AR and smart glasses become more prominent Increasing AI-enhanced user experience Growing big tech competition for the XR stack Increased application of FAIR principles to enable digital ecosystems in Europe Growing relevance of TRUST principles for data repositories to secure data More decentralisation of data ownership enabled by SSI solutions Emergence of organisational wallets Push for transparent supply chains with digital product passports Bringing computing resources closer to the ‘edge’ Growing need for automated cloud and edge resource management Growing emphasis on neuromorphic (brain-inspired) computing Integrating photonics for accelerated information processing The battle for airwaves AI/Ops, digital twinning and the emergence of the LLM in network operations Power usage and ecological footprint From ad hoc experiments to applications Towards Fault-Tolerant Quantum Computing (FTQC) Cyber-physical technology calls for resilient minds A growing urge for chip sovereignty in Europe

Drivers

Impact

education

Education

  • SLMs and edge AI can enhance personalised learning experiences directly on students’ devices, ensuring privacy and accessibility. However, institutions must manage infrastructure upgrades and maintain equitable access to avoid technology gaps. New skills in model compression and hardware optimisation will be needed.
  • The affordability of SLM devices and enabling technology may increase the digital divide.
Research

Research

  • Researchers can benefit from local, real-time analytics, without extensive computational resources. This facilitates studies in resource-constrained environments.
  • Local devices acting as sensors will be able to process data on the fly, offering greater possibilities for location independent research.
Operations

Operations

  • Institutions can deploy SLMs and edge AI to enhance operational efficiency, such as automating administrative tasks, while reducing dependency on external cloud services and potentially lowering costs. However, device-level infrastructure and skill development investment are necessary to avoid vendor lock-in and maintain long-term flexibility.
  • Institutions should proactively plan for necessary infrastructure upgrades and training programs to effectively leverage the benefits of edge AI and SLMs.
More info about AI?
Visit surf.nl
Link SURF icoon