{"id":378406,"date":"2021-12-09T00:10:58","date_gmt":"2021-12-08T21:10:58","guid":{"rendered":"https:\/\/en.buradabiliyorum.com\/deepminds-new-language-model-kicks-gpt-3s-butt\/"},"modified":"2021-12-09T00:10:58","modified_gmt":"2021-12-08T21:10:58","slug":"deepminds-new-language-model-kicks-gpt-3s-butt","status":"publish","type":"post","link":"https:\/\/buradabiliyorum.com\/en\/deepminds-new-language-model-kicks-gpt-3s-butt\/","title":{"rendered":"#DeepMind’s new language model kicks GPT-3’s butt"},"content":{"rendered":"

“#DeepMind’s new language model kicks GPT-3’s butt<\/strong>”<\/p>\n

Move over GPT-3, there\u2019s a scrapp<\/a>y new contender for the crown of <\/span>world\u2019s greatest language model<\/i> and it\u2019s from our old pals over at DeepMind.<\/span><\/p>\n

Up front:<\/b><\/span> The Alphabet-owned UK outfit that answered the question of whether humans or computers are better at chess once and for all \u2013 the machines won<\/a> \u2013 has now set its sights on the world of large language models (LLM). <\/span><\/p>\n

To that end, today it announced \u201cGopher,\u201d a language model that\u2019s about 60% larger, parameter-wise, than GPT-3 and a little over a quarter of the size of Google\u2019s massive trillion-parameter LLM. <\/span><\/p>\n

Per a press release<\/a> on the DeepMind blog:<\/span><\/p>\n

In our research, we found the capabilities of Gopher exceed existing language models for a number of key tasks. This includes the Massive Multitask Language Understanding (MMLU) benchmark, where Gopher demonstrates a significant advancement towards human expert performance over prior work.<\/p>\n<\/blockquote>\n

Background: <\/b>DeepMind accomplished the improvements by focusing in on areas where expanding the size of an AI model made sense.<\/p>\n

The more power you can shove into a model for, say, reading comprehension, the better. But the team found that other areas of LLM architecture didn\u2019t benefit as much from brute force.<\/p>\n

By prioritizing how the system utilizes and distributes resources, the team was able to tweak their algorithms to outperform state-of-the-art models in 80% of the benchmarks used.<\/p>\n

\"A
Credit: DeepMind<\/a><\/figcaption>