The recognition of "correct formula but calculation error" in math wrong answer explanations has become more refined
On the 6th (local time), Google unveiled 'Gemini,' a next-generation artificial intelligence (AI) large language model (LLM) capable of solving math problems and analyzing incorrect reasoning processes. The top-tier version is said to surpass OpenAI's GPT-4 and is regarded as the highest-performing AI model to date, rivaling human-level capabilities.
Gemini 1.0, introduced by Google, is an AI model similar to ChatGPT's LLM called 'GPT,' optimized in three sizes: Ultra, Pro, and Nano. It is a 'multimodal AI' that can simultaneously recognize and understand text, images, and audio, as well as possess coding abilities. It can solve math problems and point out and analyze incorrect reasoning processes. The development was led by Google DeepMind, the creators of AlphaGo.
Among the three types, the mid-sized and general-purpose 'Gemini Pro' will be integrated into Google's AI chatbot service 'Bard' starting today. It serves as a kind of rival to ChatGPT. Bard, equipped with Gemini Pro, will be available in English across more than 170 countries and regions, with plans to gradually expand service areas and languages. Additionally, Gemini Nano, designed to enable immediate AI use on the device itself without cloud connection, will be installed in Google's latest smartphone, the 'Pixel 8 Pro,' unveiled last October. The largest and most complex Gemini Ultra will be deployed early next year under the name 'Bard Advanced.'
In particular, Gemini Ultra is regarded as the most powerful LLM model released so far. Gemini Ultra scored 90.04 on the 'Massive Multitask Language Understanding (MMLU)' test, which evaluates knowledge and problem-solving skills across 57 subjects including math, physics, history, law, medicine, and ethics. This score surpasses that of human experts (89.3) and GPT-4 (86.4). Google emphasized, "It is the first model to exceed human expert scores," highlighting its strength especially in mathematical and physical reasoning. Gemini Ultra outperformed GPT-4 in 30 out of 32 academic benchmark categories.
A demonstration video Google previewed for local media the day before shows Gemini exhibiting human-like object recognition and judgment. When a person drew a duck on paper with a pen, Gemini immediately described the process. When the duck's body was colored blue, it introduced, "Although rare, blue ducks do exist," and when shown a duck toy and asked about its material, it replied, "It could be rubber or plastic. If it makes a squeaking sound, it's rubber." Also, when shown a scene of a person dodging bullets like in the movie 'The Matrix,' it explained, "This is a famous scene from the movie 'The Matrix.'"
Its understanding of math and physics has become more sophisticated. When asked which vehicle would be faster between one with a square front and one with a triangular front, it answered, "The triangular car, which applies aerodynamics, is faster." When shown a math problem along with an incorrect solution process, it pointed out, "The formula is correct, but there is a calculation error." It carefully identifies which parts of the solution are wrong and even provides customized practice problems related to the mistakes.
Sundar Pichai, CEO of Google, said, "It has been eight years since Google declared itself an AI-first company, and while we have achieved remarkable results, this is just the beginning," adding, "'Gemini 1.0' is the first realization of the vision we had when Google DeepMind was established earlier this year. This new era model represents one of the largest scientific and engineering efforts undertaken by Google."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
