MS launches small language model ‘Pi-3’… Competition between Google and Meta heats up

3 models including mini, small, and medium… “Improved performance, dramatically cheaper”

(San Francisco = Yonhap News) Correspondent Kim Tae-jong = Microsoft (MS), which is leading the artificial intelligence (AI) craze in collaboration with ChatGPT developer OpenAI, is predicting competition with Google and Meta by launching Small Language Model (SLM).

MS announced on the 23rd (local time) that it will release a small language model ‘Phi-3 Mini’ developed by the company. Pi-3 Mini is the lightest model among the three compact models that Microsoft will release in the future.

MS, which launched Pi-2 in December last year, explained that Pi-3 has better performance than the previous version and can provide response at the level of a 10 times larger model.

Compared to large language models (LLMs), SLMs run on personal devices such as mobile phones and laptops, and the cost of machine learning and operation is low.

“The Pi-3 is not just slightly, but dramatically less expensive,” said Sebastien Boubeck, Microsoft’s vice president of generative AI research. “Compared to other models with similar capabilities, it costs one-tenth the cost.” said.

The parameters of Pi-3 Mini are 3.8 billion, which is smaller than the upcoming Pi-3 Small with 7 billion parameters or Pi-3 Medium with 14 billion parameters.

The large language model GPT-4, which is the basis of ChatGPT, is estimated to have 500 billion parameters. The parameter indicates how many complex commands the AI ​​model can understand.

Microsoft explained that SLM is designed to perform simple tasks, making it easier for companies with limited resources to use.

“Pi-3 builds on what we learned from previous versions,” said Eric Boyd, vice president of Microsoft Azure AI Platform. “Pi-1 focused on coding, and Pi-2 started learning inference, but Pi-3 “I’m better at coding and reasoning,” he explained.

Accordingly, competition in SLM is expected to intensify along with competition in LLM.

Google previously released Gemma 2B and 7B, which are useful for simple chatbots and language-related tasks, in February.

When Meta released ‘Rama 3’ on the 18th, it also introduced a large model with 70 billion parameters and a small model with 8 billion parameters used for chatbots and coding support.

AI startup Anthropic also launched Claude 3 Haiku, which allows you to read and quickly summarize research papers with graphs.

[email protected]

Report via KakaoTalk okjebo
< em>
Unauthorized reproduction/redistribution, AI learning and use prohibited>
2024/04/24 04:04 Sent