In recent times, artificial intelligence businesses are spending billions of dollars to create the largest language models to power generative AI products. Well, now they are profiting from a new way to generate revenues: the so-called small language models.
Apple, Microsoft, Meta, Open AI, and Google are the main characters in this run to innovation and recently they introduced new AI models with fewer parameters, intending to deliver exceptionally strong capabilities while at the same time addressing business concerns regarding the high costs and high computational power that is necessary to power large language models such as OpenAI’s ChatGPT. This strategic shift is designed to make Artificial Intelligence more accessible to small/medium businesses that the financial demands of the larger models may have eroded.
At the moment, the amount of parameters in an AI model is strictly correlated with its capabilities and skills to handle articulated tasks. OpenAI’s newest model, GPT-4o, and Google’s Gemini 1.5 Pro, both recently revealed, contain over 1 trillion parameters. On the other hand, Meta is working on a 400-billion-parameter version of its own open-source Llama model. These gargantuan models, even if very powerful, bring with them massive expenses and computational requirements that are causing doubts among many enterprise customers.
On top of that, issues linked to the privacy of the data and copyright liability obstructed the adoption of many large-scale AI-generative models. To further address these problems, companies such as Meta and Google started to promote smaller language models with “only” some billion parameters. The smaller models are much more cost-effective, energy-wise, from a customization point of view, and they require way less computational power to be trained and run. This contributes to making them the practical alternatives that will still deliver a high performance meanwhile they will safeguarding our sensitive data.
Eric Boyd, corporate vice president of Microsoft’s Azure AI Platform, put emphasis on the choice of providing high-quality AI at a much lower cost, enabling many opportunities for clients, and offering a better ROI. As a response to this market demand statement, Google, Meta, Microsoft, and Mistral (French Startup) have all decided to release smaller language models that show a significant improvement in their capabilities and can be personalized for many specific and detailed applications.
Meta’s president of global affairs, Nick Clegg, said that the upcoming 8 billion-parameter Llama 3 model performs similarly to GPT-4 on many metrics. Similarly, Microsoft’s 7 billion-parameter Phi-3-small model is proved to be more capable than OpenAI’s GPT-3.5.
One important advantage of these smaller models is their capability to process operations on a device locally, eliminating, like this, the previous need for cloud computing. This new feature suits perfectly to privacy-conscious customers who would rather keep their information within their networks.
Moreover, these smaller models make the integration of AI features easier into mobile phones and many other ordinary devices. Google’s “Gemini Nano” model, is eventually embedded in its latest Pixel phone and as well Samsung’s S24 smartphone. Apple also declared that it is going to develop AI models for its new iPhone after having recently introduced the OpenELM model.
Microsoft’s Boyd stated that the diffusion of smaller models would bring a period of development of innovative applications across several devices. OpenAI's CEO, Sam Altman, said that the company already offers different-sized AI models intending to meet all the needs of their customers, and would keep developing and marketing their services. Meanwhile, Altman realized the utility of smaller models, he stated that OpenAI still has the goal to build larger models with more advanced possibilities and to reach human-level intelligence.
Finally, we denote the rapid shift towards smaller AI models is reflecting a substantial effort by the big companies to balance their performance with the practicality that is now more than ever crucial. This gigantic development is opening up many new doors for AI progress and integration across many industries, making broader and easier AI assimilation and challenging innovation in the field of technological advancement.