Microsoft released Phi-3 Mini, a lightweight artificial intelligence model for tiny languages, on Tuesday.
TakeAway Points:
- Microsoft unveiled Phi-3 Mini, a low-complexity AI model designed for small languages.
- The goal of the model is to be less complex and less expensive than existing complex large language models, like OpenAI’s GPT-4 Turbo.
- Additionally, the model can function with chips that power standard computers rather than the pricy and difficult to locate AI processors made by Nvidia.
Microsoft’s Phi-3 Mini
Two more models on the same scale are scheduled for release by the business. The goal of the model is to be less complex and less expensive than existing complex large language models, like OpenAI’s GPT-4 Turbo.
“Phi-3 is not slightly cheaper; it’s dramatically cheaper; we’re talking about a 10x cost difference compared to the other models out there with similar capabilities,” said Sébastien Bubeck, Microsoft’s vice president of GenAI research.
This is Microsoft’s smallest model to date, and since it can operate locally, it can be used on a smartphone for free without a data connection.
Additionally, the model can function with chips that power standard computers rather than the pricy and difficult-to-locate AI processors made by Nvidia.
Phi-3 Small and Phi-3 Medium Going Soon
The two additional models, Phi-3 Small and Phi-3 Medium, which have not yet been revealed, will have 7 billion and 14 billion parameters, respectively, while Phi-3 Mini has 3.8 billion. Hugging Face, Azure, and Ollama will all sell the Phi-3 small.
In December, the business unveiled Phi-2, a model that functioned similarly to larger ones like the Llama 2. According to Microsoft, Phi-3 outperforms this previous iteration and can answer questions in a way that is comparable to a model ten times larger than itself.
Phi-3 Mini is just as competent as ChatGPT, but “in a smaller form factor,” according to Eric Boyd, corporate vice president of Microsoft Azure AI Platform.
Introducing Orca-Math
The information revealed earlier in the year was that Microsoft was assembling a team dedicated to creating AI models with less weight. In addition to the Phi series of models, the business has developed a model named Orca-Math that is limited to solving mathematical issues.
The continuous race of huge language models has shifted from massive models to smaller models that are tailored to specific tasks, because larger AI models require more computation and money. Smaller, more task-specific language models can be trained to be less accurate than specialised models.
While Meta’s Llama 3-8 billion model is adept at helping with coding, Anthropic’s Claude 3 Haiku is good at summarising big documents fast, and Google’s Gemma models are excellent at language-based activities and chatbots. An AI-powered writing assistance called Jurassic-2, developed using a specific AI model, was recently introduced by smaller businesses such as AI21 Labs.
According to Boyd, Phi-3 was designed to be more proficient in both reasoning and coding, whereas Phi-1 was only focused on coding.
Microsoft last week invested $1.5 billion in G42, an artificial intelligence company based in the United Arab Emirates. This gives the American behemoth a minority stake and a board seat, enabling the two to forge closer ties in the global tech race. The partnership coincides with Washington’s efforts to stifle Beijing’s technological advancements, as it has added four Chinese companies to its export blacklist for attempting to obtain AI chips for China’s military.
