Tech giants and startups are shifting to more compact and efficient artificial intelligence models, aiming to reduce costs and improve performance. These models, unlike their "big brothers" such as GPT-4, can be trained on smaller datasets and specialize in solving specific tasks, The Wall Street Journal reports․

Microsoft, Google, Apple, and startups like Mistral, Anthropic, and Cohere are increasingly turning to small and medium-sized AI language models. Unlike large language models (LLMs) like OpenAI's GPT-4, which use more than one trillion parameters and cost well over $100 million to develop, compact models train on narrower datasets and can cost less than $10 million, using fewer than 10 billion parameters.

Microsoft, a leader in AI, has introduced a family of small models called Phi. According to the company's CEO Satya Nadella, these models are 100 times smaller than the free version of ChatGPT, yet handle many tasks almost as effectively. Yusuf Mehdi, Microsoft's commercial chief, noted that the company quickly realized that operating large AI models is more expensive than initially anticipated, prompting Microsoft to seek more economical solutions.

Other tech giants have also followed suit. Google, Apple, as well as Mistral, Anthropic, and Cohere have released their own versions of small and medium-sized models. Apple, in particular, plans to use such models to run AI locally, directly on smartphones, which should enhance speed and security while minimizing resource consumption on the devices.

Experts point out that for many tasks, such as document summarization or image generation, large models can be overkill. Ilya Polosukhin, one of the co-authors of Google's foundational 2017 AI paper, vividly compared using large models for simple tasks to going grocery shopping in a tank. "You shouldn't need quadrillions of operations to compute 2 + 2," he emphasized.

Companies and consumers are also looking for ways to reduce the costs of operating generative AI technologies. According to Yoav Shoham, co-founder of the AI company AI21 Labs from Tel Aviv, small models can answer questions at just one-sixth the cost of large language models when measured in monetary terms.

Interestingly, the key advantage of small models is their ability to be finely tuned to specific tasks and datasets. This allows them to work efficiently in specialized fields at lower costs, for instance, exclusively in the legal sector.

However, experts note that companies are not planning to entirely abandon LLMs. For example, Apple has announced the integration of ChatGPT into Siri for performing complex tasks, and Microsoft plans to use OpenAI's latest model in the new version of Windows. Companies like Experian in Ireland and Salesforce in the USA have already switched to using compact AI models for chatbots, finding that they provide the same performance as large models but at significantly lower costs and with reduced data processing delays.

The shift to small models comes amid a slowdown in the progress of large publicly available AI models. Experts attribute this to a lack of high-quality new data for training and generally point to a new and significant stage in the evolution of the industry.