Microsoft's open-source 1.5 billion parameter small model Phi-1.5 updated to support multimodal capabilities

Microsoft's research department has added an important new feature to its smaller-scale language model, indicating that low-cost AI technology can have the same capabilities as OpenAI's GPT-4.

In an interview, Microsoft researchers stated that the Phi 1.5 model is now "multimodal," meaning it can view and interpret images. They said this new skill only slightly increases the model's size, providing a pathway for democratized access to AI technology and helping alleviate the shortage of graphics processors used to run software like ChatGPT.

GPT-4, the power source behind ChatGPT, has recently become multimodal as well, but requires exponential energy and processing power. Phi 1.5 is open source, meaning anyone can run it for free.

"This is one of the major updates OpenAI has made to ChatGPT," said Sebastien Bubeck, head of Microsoft Research's Machine Learning Foundations group. "When we saw this, there was a question: is this ability only reserved for the largest models, or can our small Phi 1.5 also achieve it? Surprisingly, we can."

GPT-4 has approximately 1.7 trillion parameters, which are used for prediction, software switches, and controls. The more parameters, the more computation is required for each token or group of letters generated by the model. In comparison, Phi 1.5 has 1.3 billion parameters. If parameters were represented as distance, GPT-4's size would be equivalent to the Empire State Building, while Phi 1.5 would be the size of a 12-inch sandwich.

The effort to develop small AI models with the same capabilities as larger models is not just an academic exploration. While impressive, models like OpenAI's GPT-4 and other large-scale models are also expensive.

Individuals often request ChatGPT to draft emails, while companies typically want it to handle large amounts of enterprise data for prompt responses. These requests can be costly. The highest price for a single prompt with GPT-4 is $5, with similar prices from other providers. Companies typically pay around $100 for 1,000 prompts. "When you apply LLM to large datasets or allow multiple people to run prompts in parallel... you want to make sure pricing is taken into account," he said.

Small models require less computation to run, thus they need lower-power processors and shorter response times. Meanwhile, smaller-scale models with slightly weaker functionality can handle tasks proposed by many companies and individuals.

Lower-power models also have the advantage of reducing greenhouse gas emissions and minimizing the generation of illusions.

"We are considering how to build these systems in the real world so that they can function properly. All the work we have done on small models provides interesting puzzle pieces for constructing this ecosystem," said Ece Kamar, head of Microsoft's research department.

Researchers state that while these small models are powerful, they can never replace larger foundational models like GPT-4, which will always remain ahead. However, these two models are complementary. For certain applications, large models will be necessary. For very specific tasks, small models may be more cost-effective.

Ahmed Awadallah, senior principal researcher at Microsoft Research, suggests that the future may involve using both small and large models simultaneously to handle tasks. "You can also imagine deploying small models in different ways. When it lacks confidence in performing a task, it can fall back to the large model," he said.

Microsoft researchers are also experimenting with using multiple small models together as "agents" to handle different aspects of tasks. For example, one model can write software while another model can check for errors.