Artificial intelligence chip startup Cerebras Systems on Tuesday mentioned it launched open supply ChatGPT-like fashions for the analysis and enterprise group to make use of for free in an effort to foster extra collaboration.
Silicon Valley-based Cerebras launched seven fashions all educated on its AI supercomputer referred to as Andromeda, together with smaller 111 million parameter language fashions to a bigger 13 billion parameter mannequin.
“There is a big movement to close what has been open-sourced in AI…it’s not surprising as there’s now huge money in it,” mentioned Andrew Feldman, founder, and CEO of Cerebras. “The excitement in the community, the progress we’ve made, has been in large part because it’s been so open.”
Models with extra parameters are capable of carry out extra complicated generative capabilities.
OpenAI’s chatbot ChatGPT launched late final 12 months, for instance, has 175 billion parameters and may produce poetry and analysis, which has helped draw massive curiosity and funding to AI extra broadly.
Cerebras mentioned the smaller fashions might be deployed on telephones or sensible audio system whereas the larger ones run on PCs or servers, though complicated duties like massive passage summarization require bigger fashions.
However, Karl Freund, a chip guide at Cambrian AI, mentioned larger isn’t all the time higher.
“There’s been some interesting papers published that show that (a smaller model) can be accurate if you train it more,” mentioned Freund. “So there’s a trade off between bigger and better trained.”
Feldman mentioned his largest mannequin took a little bit over per week to coach, work that may sometimes take a number of months, due to the structure of the Cerebras system, which features a chip the scale of a dinner plate constructed for AI coaching.
Most of the AI fashions at present are educated on Nvidia’s chips, however an increasing number of startups like Cerebras try to take share in that market.
The fashions educated on Cerebras machines may also be used on Nvidia techniques for additional coaching or customization, mentioned Feldman.
© Thomson Reuters 2023
