Victor Botev, CTO and co-founder at Iris.ai
In recent years, Artificial Intelligence (AI) models have undergone remarkable advancements, with smart models tailored to specific use cases emerging as strong contenders against proprietary Large Language Models (LLMs) developed by big tech companies. As our understanding of these models deepens, the performance gap between smaller, open-source AIs and LLMs is rapidly closing. The move towards smarter, customisable models emphasises the power of community-driven innovation and highlights the need to prioritise speed, efficiency, and data quality in AI development.
With the increasing speed at which the complexity of these models grows, questions around pausing development have inevitably arisen. Putting a halt to AI development altogether, however, could disproportionally affect small start-up developers. Should such policies come into effect they may well expand existing moats, which are the advantages enjoyed by incumbents that hinder new entrants into the market.
Moreover, even as start-ups have made gains in narrowing the gap between themselves and larger developers, they still face a unique challenge in the AI landscape, where the scale and quality of data is crucial for training LLMs. Hyperscaler firms have already accumulated vast amounts of data through their existing services, granting them a significant advantage. As the number of users utilising their platforms increases, so does the volume of data available for improving their LLMs.
In contrast, organisations with limited customer bases struggle to access such extensive data, hindering their ability to train and improve their AI models. Halting AI developments would only widen the gap, further impeding start-ups from catching up to big tech’s advancements.
With competition being essential for innovation, there are a variety of ways to foster the development of AI and avoid the moats of Big Tech from stifling the industry.
Niche Domain-Specific Language Models:
Instead of directly competing with big tech’s general-purpose LLMs, many AI developers are differentiating themselves by developing smarter language models tailored to specific domains. By specialising in verticals such as academia, healthcare, finance, or law, start-ups can deliver superior AI solutions that better understand the intricacies and nuances of specific industries. Specialised models can address domain-specific challenges and provide more accurate and relevant insights, effectively staving off competition from hyperscalers.
Many are prioritising acquiring domain expertise and collaboration with industry professionals to develop their AI models aligned with specific verticals. An iterative feedback loop between experts and model developers works to refine tools more efficiently over time, ensuring continued improvement and reliability. Moreover, by integrating industry-specific knowledge into their models, developers can provide tailored solutions that better meet the needs of customers in those sectors. This integration creates models that are more equipped to provide results for the specifics of a given field, better navigating the challenges of potential regulation, technical information and vocabulary. This targeted approach enhances their competitive advantage, as domain-specific language models can offer specialised insights and perform complex tasks more effectively than generalised models.
Start-ups specialising in vertical-specific language models are also exploring the potential of transfer learning. By using pre-trained general-purpose language models as a foundation and then fine-tuning them with domain-specific data, start-ups can achieve faster development cycles as developers are able to capitalise on the knowledge already embedded in the general models while working to tailor them to specific verticals.
Open-Source Community
Open-source initiatives have already played a pivotal role in democratising AI technologies and narrowing the gap between smaller competitors and hyperscalers. These initiatives promote collaboration, knowledge sharing, and code transparency, enabling developers to collectively advance AI capabilities. Open-source projects are added to by a wide variety of contributors, harnessing the power of collective expertise and expanding the availability of advanced AI tools and frameworks.
By engaging with the open-source community, those building AI models can access pre-existing libraries, frameworks, and resources, saving valuable development time and costs. Through these collaborations, many start-ups are enhancing their models, fostering innovation, and gaining recognition for their contributions.
The AI landscape is experiencing a paradigm shift towards smaller, customisable models that prioritise efficiency and effectiveness over sheer scale. The performance equivalence demonstrated by smarter models compared to their larger counterparts challenges the notion that bigger is always better. By leveraging their advantages, smaller companies can offer AI solutions that are more transparent, accessible, and capable of meeting specific industry needs.
This positive trend should not be taken for granted, however, and the expansion of moats should not only be a concern for the wider industry but also policymakers. As AI continues to advance, it is crucial for a balance to be struck between fostering innovation and ensuring fair competition. Any new legislation or regulations should consider industry best practices and the potential differences in impact across all of those involved in AI. A thriving ecosystem is one that aims to promote healthy competition, prevents monopolisation, and encourages the growth of diverse AI solutions.
Institutional bodies must focus on educating public authorities about the inner workings of AI and LLMs. By funding academics and critical bodies, granting them the resources to study and hold the big players accountable, institutions can help, not hinder, the open-source movement. This will do away with unclear, complex, and unhelpful regulation – allowing the community to publish and distribute their findings – naturally regulating the ecosystem and improving AI for the many, not the few.
 models have undergone remarkable advancements, with smart models tailored to specific use cases emerging as strong contenders against proprietary Large Language Models (LLMs) developed by big tech companies. As our understanding of these models deepens, the performance gap between smaller, open-source AIs and LLMs is rapidly closing. The move towards smarter, customisable models emphasises the power of community-driven innovation and highlights the need to prioritise speed, efficiency, and data quality in AI development.){kind=link}