English | EN 
Introduction to LLMs
Lately, there has been much talk about large language models, or LLMs. They are widely used, and almost everyone has heard of them. These AI models are incredibly proficient with words, assisting humans with anything from email composition to information retrieval. Our usage of technology is evolving, becoming more engaging and simpler as a result. But as our needs evolve, the way we use them will as well.
Currently, LLMs are predominantly hosted on the cloud and offered as web services. Examples include OpenAI’s ChatGPT, Google’s Bard, Microsoft’s BingChat, and Anthropic’s Claude. The cloud-based model allows these complex and resource-intensive systems to be accessed from anywhere with an internet connection.
Running an LLM such as GPT-4 requires substantial computational resources. This typically involves high-end GPUs, large amounts of RAM, and considerable storage space for the model and its training data. The exact specifications can vary, but they are often beyond the reach of most individual users or small organizations.
Cloud-based large language models offer several advantages along with certain limitations. One of the primary benefits is accessibility. Users can easily access these models from any device, making them highly versatile and user-friendly. Additionally, the responsibility for maintenance and updates lies with the service provider, ensuring that the LLMs are always up-to-date. Furthermore, cloud-based LLMs provide virtually unlimited scalability and resource management.
On the flip side, there are also significant drawbacks to using cloud-based LLMs. One major disadvantage is the reliance on internet access, which can be problematic in places with inadequate connectivity. Imagine driving your car in rural areas and not being able to use the voice assistant. Privacy issues also come up, especially when sensitive data is involved. Due to external handling and storage, sending data to a cloud service for processing may involve privacy hazards. These benefits and drawbacks draw attention to the compromises made when deploying LLMs on the cloud.
LLMs on our local machines
Open-Source Large Language Models (LLMs) represent a significant shift in the landscape of artificial intelligence. These models can be used, modified, and distributed by anyone. This openness fosters a collaborative environment where developers and researchers can take part and build on each other’s work, accelerating innovation and accessibility in the field of AI.
Thanks to various technological advancements, such as more efficient model architectures, optimization of algorithms, and the increased availability of powerful computing hardware at a consumer level some people have managed to run LLMs on local machines.
The gap between proprietary and open-source models is continually narrowing, thanks in part to the collaborative nature of the open-source community and the rapid pace of advancements in AI research. This makes open-source LLMs not only a viable alternative for many applications but also a field of active innovation and development. In addition, it’s important to keep in mind that these open-sourced models could be improved with custom (and private) knowledge, using a technique called fine-tuning.
The Future of LLMs
I believe that in the near future, we will start seeing some LLMs applications that make use of local resources. Also, I am convinced that some companies will start developing their models for internal use, and they will be hosted on-premises. I think there will be hybrid deployment models that combine the strengths of cloud-based and local processing. This approach might involve initial heavy processing in the cloud and then local refinements for enhanced privacy and customization. The move towards local deployment will likely enhance customization and data control. Enabling models to offer user experiences that are more tailored to individual preferences.
In summary, the future of LLMs is likely to be significantly influenced by the trend towards local deployment. This evolution is driven by the growing need for privacy, security, and offline functionality, paving the way for more personalized, efficient, and versatile AI solutions across various industries.
However, it should be noted that these assertions regarding local deployment as the definitive direction for LLMs represent only my personal predictions and should not be construed as reflecting the official position of Sogeti.
