It seems like Silicon Valley’s golden child isn’t perfect after all. In a rare moment of weakness, large-language model developer OpenAI scrapped a new model, codenamed Arrakis, in mid-2023 that would have helped it run its popular chatbot ChatGPT more cheaply, my colleagues Jon Victor and Aaron Holmes reported on Tuesday.
OpenAI researchers hoped to use a machine-learning concept called sparsity to help the new model run more efficiently. To create an LLM, developers pass enormous amounts of data through an AI model, which helps the model adjust its parameters, or its “settings,” to generate more accurate information. In sparse models, the values of many of the models’ parameters are set to zero. This means that, when the model is asked a question, it runs the query through far fewer parameters than is typically the case. That’s a much faster and less-intense undertaking, which allows developers to spend less on the AI chips powering this process.
Sparsity isn’t a perfect solution to the ballooning costs of AI model training and running: Since many of its “settings” are zeroed out, a sparse LLM may be less capable of answering questions on topics not included in its training data than a standard “dense” model. In the case of Arrakis, the sparse model worked in early tests but didn’t perform well afterwards. We’re not sure why.
0 Commentaires