Partager:

JAKARTA - The artificial intelligence chip startup Cerebras Systems announced on Tuesday 28 March that it was releasing an open source model similar to ChatGPT for free use by researchers and businesses to encourage further collaboration.

Silicon Valley-based celebrations released seven models, all trained on its supercomputer artificial intelligence called Andromeda, ranging from smaller 111 million-parameter language models to larger 13 billion-parameter models.

"There is a big move to close what the open source has done in artificial intelligence... it's no surprise that there is now big money in it," Andrew Feldman, founder and CEO of Cerebras said. "The excitement in the community, the progress we've made, is largely because it's very open."

Models with more parameters can perform more complex generating functions.

For example, OpenAI's chatbot ChatGPT, which was launched late last year, has 175 billion parameters and can generate poetry and research, which has helped attract great interest and funding for artificial intelligence more broadly.

Cerebras says smaller models can be applied to mobile phones or larger transient smart speakers run on PCs or servers, although complex tasks such as summarization require a larger model.

However, Karl Freund, a chip consultant at Cambrian AI, said that big doesn't necessarily mean better.

"There have been several interesting papers showing that (smaller models) can be accurate if you train them more," Freund said. "So there's trade off between bigger and better trained."

Feldman said his biggest model took a little more than a week to train, work that usually takes several months, thanks to the Cerebras system architecture, which includes chips as big as food plates built for artificial intelligence training.

Most of the current artificial intelligence models are trained on Nvidia Corp (NVDA.O) chips, but more and more startups like Cerebras are trying to take part in the market.

Models trained on Cerebras machines can also be used on the Nvidia system for further training or customization, said Feldman.


The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)