Artificial Intelligence (AI) companies, after investing billions in creating large language models (LLM), are now turning their focus to smaller language models as potential revenue boosters, reports the Financial Times. The creation and training of these AI models carry much lesser costs than LLMs.
Tech giants like Apple, Microsoft, Meta, and Google have recently launched smaller AI models with fewer parameters but with impressive capabilities. The prohibitive cost of LLMs, along with concerns regarding data usage for training due to potential copyright infringement, prompted the shift.
Companies such as Meta and Google are offering smaller language models with only a few billion parameters as a cheaper, energy-efficient, customizable alternative to larger language models. Trained and launched, these models require less energy and can guarantee the confidentiality of data.
“By offering such high quality for lower costs, you’re effectively giving clients the opportunity to use a lot more applications and carry out those actions which, in their opinion, would not have given them sufficient ROI (on LLM) investments to justify their actual use,” noted Eric Boyd, corporate vice president of Microsoft Azure AI Platform.
In addition to Google and Meta, Microsoft and French startup Mistral have also released smaller language models with enhanced capabilities and a stronger focus on specific tasks.
The advantage of smaller language models also lies in their ability to perform tasks locally on the device, without sending information to the cloud. This is beneficial for clients concerned about confidentiality and those who do not want to send information beyond internal networks. These models can also be utilized on smartphones. For instance, Google’s Gemini Nano model is installed on Pixel and Samsung S24 smartphones. Apple has hinted at developing AI models for operation on iPhone devices.
However, Sam Altman, Head of OpenAI, stated that his company would continue to work on creating larger AI models with advanced capabilities. These models will be capable of reasoning, planning, and task execution and will eventually achieve a human-level intellect.
