Our journey through 2024 and next steps in 2025

The AI landscape of 2024 was defined by a trend rooted in duality – democratization of AI through open source, and the increased sophistication of AI-powered products. Model performance continued to increase, while many innovations yielded lower training and inference costs. Open source models emerged as credible alternatives to closed solutions, democratizing access to the technology. At the same time, enterprises demanded more advanced AI capabilities, requiring significant compute power for training both general purpose and specialized models and running inference at scale.

Translating the remarkable improvements in technical AI development into market evolution, 2024 saw the generative AI upswing of 2023 mature into increased commercialization of AI. Year 2024 saw usage of large language models (LLMs) scale as embedded parts of digital products, ranging from Microsoft Copilot and GitHub Copilot to ChatGPT and Meta AI, just to name a few. These and others exemplify my prediction from last year on AI creating most value by integrating models into digital products. For most use cases, LLMs will generate value as new features to existing digital products and services, enhancing the user experience and ultimately helping build user experiences that end-users love, leading to data flywheels with increased value creation for both customers and software vendors.

Evidence of value creation to end users and commercial traction in a market of significant size was also echoed in the market by significant funding rounds for AI powerhouses such as OpenAI, Mistral AI and Anthropic, and multiple acquisitions of key AI companies and AI talent, such as Microsoft’s acquisition of Inflection AI, Amazon’s acquisition of Adept AI, Google’s acquisition of Character.ai, and likewise AMD’s acquisition of Silo AI.

Fostering advancements in European frontier AI

Silo AI most importantly continued pushing the boundaries of AI development in Europe, beyond growing the company’s business (including headcount, revenue, gross margin etc).

Early in the year, we released the first version of the SiloGen model-as-a-service platform, providing end-to-end capability to develop and deploy models at scale, including guardrails, controls and the ability to optimize models for domain-specific needs. And eventually SiloGen was deployed with several customers such as Tietoevry, Transcom, MySpeaker, Etals and Doctrin.

Likewise, we made progress along multiple lines on AI model development, especially for European languages. We expanded our family of open source language models, launching multiple versions of both Poro and Viking (covering work on Poro 34B, Poro Chat, Viking 7B, 13B, 33B and upcoming models), with the latter now extending to support English and all Nordic and European languages. As part of these, we also contributed to scaling LLM training on AMD compute platforms, including contributions to open source frameworks. These models demonstrate our commitment to developing open AI that preserves and promotes linguistic diversity.

Continuing to foster the broader AI community, we expanded partnerships with industry-leading companies, such as Google, European AI champion Aleph Alpha and a leading frontier AI model developer Mistral AI, as well as leading AI networks like the German Klatten and Schwarz-family led appliedAI, research powerhouses ELLIS Institute and Cyber Valley, and the Wallenberg-led cross-industry collaboration network Combient. As we move into 2025, we're scaling up these efforts, working on even larger and more capable multilingual models as well as expanding partnerships for continued efforts as part of the AI community.

The European AI ecosystem saw significant developments in 2024. The implementation of the EU AI Act established the first comprehensive framework for AI regulation globally, while major investments in compute infrastructure and research institutes strengthened Europe's foundation for AI development. Mario Draghi's landmark report on European competitiveness spurred discussion on both challenges and opportunities ahead, emphasizing the need to translate Europe's research excellence into commercial success. Beyond initiatives like the LUMI supercomputer, EuroHPC’s AI factories and ELLIS Institutes, we've seen an acute need for commercial AI initiatives for Europe to succeed.

As the year comes to a close, we are happy to see in Europe both world-class AI research and the ability to build and scale technology and AI companies for the real world. This has manifested in leading AI companies releasing models and products as well as raising significant funding for further growth, such as ElevenLabs, Mistral AI, Helsing, DeepL, H Company, Stability AI, Aleph Alpha and poolside (see more here).

AMD x Silo AI – Shaping the future of AI computing

Beyond overall roadmap execution, commercial success and market development in AI, this year implied more fundamental development and progress for us at Silo AI. In the summer, we joined forces with AMD, to be on a joint mission to shape the future of computing, while continuing to build a leading AI lab. This strengthened our ability to work on more advanced and larger AI initiatives, expanding our capacity to help customers create value with AI, while maintaining our European roots and strengths as a private AI lab.

Joining forces with AMD has strengthened Silo AI’s ability to support enterprises in this evolving landscape. As AI is becoming an integral part of digital services, products and platforms, the demand for sophisticated AI models continues to grow. Looking ahead, we are uniquely positioned to help companies navigate this transformation, combining our AI expertise with AMD's compute leadership to deliver solutions that drive real business value and competitive edge. This alignment of capabilities enables us to take on more ambitious initiatives and spearhead increasingly challenging AI research and development.

Looking ahead – AI in 2025

Year 2025 promises to be even more exciting. The commercialization of AI will continue, as state-of-the-art AI continues to evolve and models move from research labs to real-world applications.

Many are predicting that 2025 will be the year of AI agents, while plateauing scaling laws have raised concerns, but I believe the evolution of enterprise AI goes far beyond general purpose AI assistants and agents. Majority of the discussion about scaling laws focuses on language, including the original 2020 scaling law paper and also solutions through test-time compute, whereas language is not the only relevant data modality. We expect to not only see industry-specific AI models becoming the norm, but also other data modalities being exploited in biology, pharma, material sciences, robotics, autonomous vehicles and world models, among many other fields. While these modalities have not seen saturated scaling laws, or even been exploited to start with, the rapid advancement in model efficiency means these will become increasingly accessible for companies in different industries and of various sizes. 

AI technology is maturing rapidly, and its impact on business and society continues to grow. As we enter 2025, we remain committed to create value to customers as a leading private AI lab. And with a continued focus on human-centric, trustworthy AI, we are excited to help shape this next phase of the AI revolution. The building blocks are in place at AMD – now it's time to continue expanding this potential into reality and real-world impact.

Wishing you all the best in the new year!

Peter