DeepMind goes LLM with Gemini | | | | Turtles AI

DeepMind goes LLM with Gemini
  #DeepMind's next #AI, #Gemini, aims to beat #ChatGPT by combining text generation with #AlphaGo #planning and #problemsolving #skills, though few details are revealed about the still-in-development system. Experts argue LLMs lack physical reasoning, so combining techniques offers progress. But risks and ethics must be considered for powerful AI. When the artificial intelligence system AlphaGo first defeated world champion Lee Sedol in the ancient game of Go in 2016, it astounded the AI world. But the creator of AlphaGo, Demis Hassabis, cofounder and CEO of DeepMind acquired by Google in 2014, is now planning another ambitious AI project called Gemini that according to Hassabis will surpass the capabilities of OpenAI’s famous ChatGPT. While ChatGPT uses a large language model, Gemini will merge that talent with techniques from AlphaGo, allowing it to plan better, solve problems and accomplish tasks. However, Hassabis gives few details about Gemini, which is still being developed after months of work. It is estimated that this large multimodal model contains between one and five trillion parameters and was trained on a dataset of between 20 trillion and 100 trillion tokens, equivalent to 40 to 200 terabytes of data. The training data likely accumulated until at least December 2022, and the model will likely converge or finish training around July 2023, with a public release that may occur as soon as October 2023. Many experts argue that dominant language models like ChatGPT lack "common sense" and physical reasoning, showing that combining different AI techniques may offer progress. But creating ever more powerful AI systems raises complex ethical concerns about risks and proper governance that Hassabis himself acknowledges. Don't forget to read our guide to LLMs to better understand the basics about them.