General details
EDIHs involved
Challenges
Miðeind is a leader in the field of language technology and artificial intelligence in Iceland. The company works on language technology and artificial intelligence for Icelandic. Their technology makes it possible to work with Icelandic text and speech on computers, phones and other devices. Among other things, it can be used to extract information from text, read over spelling and grammar, translate text between Icelandic and other languages, answer questions, create summaries, etc.
Since EDIH-IS began, and in cooperation with EuroCC2 National Competence Center for HPC/AI, Miðeind was supported with HPC services to enable Icelandic LLMs. Based on that, EDIH-IS was instrumental in bringing Miðeind together with other European experts to obtain a Horizon Europe grant: TrustLLM.
The TrustLLM project will develop European large language models (LLMs) on an unprecedented scale, trained on the largest amount of text so far in European AI, covering a range of underrepresented languages, and pushing the limits of European exascale computing. The main objective is the development of an open, trustworthy, and sustainable LLM initially targeting the Germanic languages.
Solutions
Under the leadership of the University of Iceland's Work Package 4 (WP4), the specific needs of SME Mideind were thoroughly analyzed:
-
Provisioning of HPC Resources: Miðeind needed high-performance computing (HPC) resources to enhance their Icelandic Large Language Models (LLMs).
-
Increase in Manpower: To fully capitalize on the potential of advanced LLMs, Miðeind required more skilled personnel to engage in the development and improvement of additional LLMs.
To address these needs, the following actions were taken:
-
Provision of HPC Resources: HPC resources were directly provided and utilized by Miðeind to improve their Icelandic LLMs.
-
Manpower Augmentation: The issue of insufficient manpower was strategically resolved by jointly engaging in European grants. This led to the creation of two new positions at Miðeind.
These strategic interventions ensured that Miðeind could meet its digital transformation needs effectively, leveraging both advanced computational resources and increased human capital to drive innovation and growth.
See Also:
https://www.ihpc.is/news-resources/sme---mideind---recent-success-stories
Results and Benefits
The benefits for SME Miðeind are substantial, thanks to the support from EDIH-IS. This support includes:
-
Access to HPC Resources: EDIH-IS enabled Miðeind to utilize high-performance computing (HPC) resources, which are essential for developing advanced Icelandic Large Language Models (LLMs).
-
European Grant TrustLLM: Through the TrustLLM grant, Miðeind gained the opportunity for extensive collaboration with European LLM experts, such as those at AI Sweden. This collaboration provides Mideind with valuable knowledge and expertise.
-
"Test Before Invest" Environment: The combination of HPC resources and expert collaboration offers Miðeind a significant "Test before Invest" environment. This allows Mideind to refine and test their LLM-derived products and services, such as chatbots that communicate in Icelandic more effectively than ever before.
EDIH-IS plays a crucial role in the TrustLLM project by:
-
Supporting Outreach in Europe: Actively participating in outreach efforts within the European LLM communities to foster collaboration and knowledge sharing.
-
Providing Computing Time Access: Facilitating access to computing time through the EuroHPC Joint Undertaking, in collaboration with the National Competence Center (NCC) for HPC/AI Iceland.
Perceived social/economic impact
While popular language models like ChatGPT and Google Translate do not yet speak Icelandic fluently, even with the training steps taken for GPT-4 on Icelandic data, the challenge remains significant. The current focus has been on Reinforcement Learning with Human Feedback (RLHF), but incorporating low-resource languages like Icelandic into large language models (LLMs) continues to be a pressing research question that TrustLLM aims to address.
TrustLLM Project Goals:
-
Addressing the scarcity of Icelandic text corpora compared to other languages.
-
Improving LLM fluency in Icelandic and other low-resource languages.
-
Enhancing the competitiveness of companies like Mideind through better LLMs.
Impact:
-
Social and Economic Impact: The significance of this initiative is highlighted by numerous press releases in Iceland, celebrating the acquisition of this major EU TrustLLM grant:
-
RUV Icelandic TV Channel: 200 Million Grant to Develop AI Models
-
University of Iceland Press Release: Mideind and University of Iceland Receive Large European Grant for AI Project
-
EDIH-IS brought together leading European LLM initiatives within TrustLLM to support the development of LLMs for Icelandic and other low-resource languages, ultimately making companies like Miðeind more competitive.
DMA score and results - Stage 0
The Digital Maturity Assessment (DMA) of the customer reveals a diverse landscape of strengths and weaknesses across various components of digital maturity. Notably, strengths are evident in Automation & Artificial Intelligence, where a robust score of 64% showcases significant progress and competence in leveraging advanced technologies.
Additionally, Digital Readiness emerges as another area of strength, with a score of 53%, indicating a strong preparedness to adopt digital technologies. However, several weaknesses persist, particularly in Human-Centric Digitalisation and Green Digitalisation, where scores of 28% and 25% respectively underscore critical areas for improvement in prioritizing user-centric approaches and eco-friendly digital practices.
Further weaknesses are identified in Digital Business Strategy and Data Governance, reflecting the need for more comprehensive and well-defined strategies, as well as stronger policies and processes to manage data effectively.
-
Digital maturity level: 40%
-
Digital Business Strategy: 33%
-
Digital Readiness: 53%
-
Human-Centric Digitalisation: 28%
-
Data Governance: 38%
-
Automation & Artificial Intelligence: 64%
-
Green Digitalisation: 25%
Lessons learned
Working with SMEs like Miðeind reveals key insights into integrating HPC and AI technologies:
Do's:
Provide Tactical Consulting and Ad-Hoc Support:
-
Offer consulting HPC services and AI support to address immediate needs.
-
Ensure support aligns with the SME’s long-term goals.
Invest in Manpower Development:
-
Encourage SMEs to build up their own knowledge over time.
-
Facilitate the acquisition of additional manpower through grants and funding opportunities.
-
Example: Miðeind secured two new positions via the Horizon Europe TrustLLM grant, enhancing their expertise in using LLMs with advanced HPC resources.
Don'ts:
Rely Solely on External Support:
-
Avoid creating dependency on external consultants for critical HPC and AI functions.
-
SMEs should develop self-sufficiency in these areas over time.
Overlook Long-Term Knowledge Building:
-
Don’t neglect continuous learning and internal capacity building.
-
Ensure that SMEs invest in ongoing training and development for their teams.
Providing tactical consulting combined with manpower investment is crucial. Miðeind’s success with the Horizon Europe TrustLLM grant allowed them to build in-house knowledge and leverage cutting-edge HPC resources, demonstrating the effectiveness of this approach.
Other Information
EDIH-IS, in cooperation with the Icelandic National Competence Center (NCC) for HPC/AI (that is part of EDIH-IS), provided access to an HPC resource called DEEP where Miðeind researchers could experiment with larger HPC machines as done before. Regular follow-ups between EDIH-IS and Miðeind have been performed to understand HPC challenges and issues (e.g., i-node scalability problems on HPC systems, etc.). In parallel, the manpower issue of the SME Miðeind was also addressed by exploring different options for European grants. EDIH-IS, with its great contacts in Europe, enabled the cooperation between different European experts that, together, applied in a Horizon Europe call on LLMs. EDIH-IS and the Icelandic NCC for HPC/AI have significantly supported the consortium meetings. The project is called TrustLLM: https://trustllm.eu/
Need support?
Consult our catalogue to locate the Eupopean Digital Innovation Hub nearest to you and accelerate your company's digital transformation.