Google, Amazon, Nvidia, and others put $235 million into Hugging Face • Hugging Face, which acts like GitHub for machine learning and other AI models, codes, and datasets, raised $235 million in a Series D fundraising round, reported CNBC. Investors in this round included Google, Amazon, AMD, Intel, IBM, Nvidia, and Salesforce, all of whom have invested significantly into generative AI foundation models or processors running these models. • (The Verge, Emilia David) / August 24

Llama 2 is about as factually accurate as GPT-4 for summaries and is 30X cheaper • Summarization is one of the top immediate practical applications of LLMs (the other ones in our experience so far being retrieval augmented generation, talking to your data and long-document question answering). One of the biggest challenges with summarization, however, is factuality: does the summary reflect accurately what the original document said? There are other characteristics, such as fluency and relevance that are also important, but LLMs are actually pretty good at both of those. Factuality (or its evil twin: hallucination) on the other hand is a known issue with LLMs. And it’s no use being fluent if you’re wrong. In this experiment, we found Llama-2-70b is almost as strong at factuality as gpt-4, and considerably better than gpt-3.5-turbo. We also ran cost comparisons for the summarization and found that Llama 2 tokenization is longer than ChatGPT tokenization by 19% and this needs to be taken into account for cost. Despite this, Llama 2 is 30 times cheaper for GPT-4 for equivalent levels of factuality in summarization • (Anyscale, Waleed Kadous) / August 23