Generative AI launched into the mainstream with the equivalent of a technology “Big Bang.”
ChatGPT, developed by OpenAI in the U.S., exploded into the world in 2022. It showcased for everyone, from business and tech leaders to governments and citizens worldwide, the generative large language models’ power to generate realistic responses from a prompt.
ChatGPT, a natural language processing foundation model, is the definition of large. The GPT-3.5 model, with training data of 570GB from books, web texts, Wikipedia, articles and other writing — or about 300 billion words — built a model with 175 billion parameters.
This was surpassed by GPT-4, the latest model available to paying enterprise users. Described as multimodal, meaning it can respond to text and image prompts, it is rumoured to operate using 1.7 trillion parameters, which would make it nearly 1,000 times larger than its predecessor.
The move from ‘bigger is better’ to LLM of best-fit
The world was acquainted with more LLMs in 2023. These included Google’s PaLM 2, powering the search giant’s chatbot BARD (now called Gemini), and Anthropic’s Claude, the latest version of which is Claude 2.
However, the growth in proprietary LLMs has been matched by a boom in open-source models.
Meta’s open-source offerings Llama and Llama 2, which are free for research and commercial use, have been joined by other popular models like BERT, released by Google in 2018, and image generator Stable Diffusion from Stability AI. The diversity of open-source models has grown quickly as AI innovation and investment have taken off around the world and in Asia. One GitHub estimate suggests there were over 8,000 models to choose from in October 2023.
The growth in open-source AI models is a recognition that, while large is indeed powerful in the world of LLMs, one size does not necessarily fit all. Open-source models have proven to offer a universe of choice and customisability for the likes of SMEs and startups. Businesses and developers in Asia now know they have the option to choose and fine-tune open-source models to support unique use cases, supporting creativity and innovation.
For example, high-performance computing provider Bitdeer Technologies Group, an Asia-based NVIDIA cloud services provider, offers AI Cloud platform. Its services include open-source model training and deployment and AI computing to “foster creativity, streamline expertise and accelerate ingenuity.”
In Asia, embracing a diversity of LLM training models could mean creating tools better-suited to the culture or languages of the region, as Singapore is doing with its National Multimodal LLM Programme, or training models for use in specific industries or use cases, like the legal profession. Open-source options may be particularly useful for niche business use cases, as it has been demonstrated that these models can be fine-tuned to outperform proprietary models on specific tasks when trained on specific sets of data, such as with private LLMs.
Bitdeer is maximising AI model choice and customisation
Bitdeer’s Al Cloud platform is opening up the potential of AI model training and fine-tuning in Asia. Underpinned by world-leading NVIDIA SuperPOD H100 infrastructure, it can support businesses in the utilisation of AI models to innovate and grow in 2024 and beyond.
A playground of open-source AI models
Bitdeer offers the choice of a range of open-source generative AI models with which to train, fine-tune and deploy AI applications. These include a selection of top multimodal, natural language processing, computer vision, audio and other open-source foundation models.
Bitdeer’s platform for open-source foundation models means businesses have the means to select and adapt appropriate models for specific use cases. This could provide the ability to better cultivate production-ready AI applications to meet evolving customer needs in Asia.
Faster AI model training power
Bitdeer’s NVIDIA DGX infrastructure with H100 systems puts the pinnacle of AI compute performance to work in training bespoke AI models. Utilising high-performance computing can lead to faster return on investment in AI and lower the total cost of development operations.
World-leading AI infrastructure
Bitdeer AI Cloud is built on NVIDIA’s state-of-the-art SuperPOD H100 infrastructure, offering businesses in Asia performance, scalability and security. A turnkey solution that eliminates design complexity, it is a full-stack offering backed by NVIDIA, with an NVIDIA-powered deployment and ramp-up service, along with support for NVIDIA AI Enterprise.
NVIDIA’s AI infrastructure enables the orchestration of AI workloads and optimization of machine uptime while deploying software that continuously improves over time. This can help with everything from delivering more AI prototypes into production to meeting growing demands and user counts.
Opening the doorway to AI innovation
The proliferation and adaptation of AI foundation models will continue throughout 2024 and beyond. This will include the continual improvement of open-source models, which will be adopted by businesses to unlock creative applications of AI in a fast-moving market for digital products and services.
The utilisation of regional AI cloud offerings like Bitdeer is one way for organisations to access and deploy AI affordably with the best in global infrastructure. This can only open the doorway to growth in regional AI deployments as local businesses use a variety of models to win in the new AI race.