"Google wants to set new standards with its artificial intelligence Gemini. But there are doubts as to whether the company will finally catch up with ChatGPT developer Open AI.
If parents need help with their children's math or physics homework, the Internet company Google says they will soon simply ask "Gemini." The company's latest artificial intelligence (AI) recognizes a student's handwritten worksheet with physics problems in a promotional video and then makes suggestions for corrections.
With “Gemini,” Google is releasing what it claims is the most powerful and versatile AI model the company has ever developed. Google announced this in a blog post recently. The American technology group is once again attacking the chat GPT developer Open AI and its partner Microsoft. The pressure is high for Google to keep up with the new model from Open AI. Many experts consider GPT-4 to be the most powerful AI on the market. A successful launch of Gemini would therefore be extremely important for Google. Does that work?
"No serious competitor for GPT-4"
Patrick Bunk is skeptical. Gemini is better than the previous model PaLM 2, but "still not a serious competitor for GPT-4." Bunk is an AI expert and, among other things, founded the AI start-up Ubermetrics, which is now part of the media analysis provider Unicepta. One problem: The best variant of the new AI has not yet been published.
Gemini comes in three sizes. The most powerful version, Gemini Ultra, is intended for highly complex tasks, while the slightly slimmed-down version, Gemini Pro, is intended for broader tasks. The model that runs in an even smaller version is called Nano is available also on smartphones. It's available now on the Google Pixel 8 Pro.
Google advertises that Gemini Ultra, with a score of 90 percent, is the first artificial intelligence to outperform human experts in so-called Massive Multitask Language Understanding (MMLU). A combination of 57 subjects such as mathematics, physics, history, law, medicine and ethics are tested to test general knowledge and problem-solving skills. GPT-4 achieved 86.4 percent in the same query. The methodology was developed by independent scientists from Berkeley, Columbia and Chicago universities.
However, experts sometimes doubt the validity of such tests because a model - by chance or by design - may have been fed more with the corresponding test data. The systems are a result of the data, the training power and the subsequent fine-tuning, says Aljoscha Burchardt, an expert in language technology and AI at the German Research Center for Artificial Intelligence in Berlin, to the F.A.Z. "That defies meaningful comparisons, especially since you usually don't know the details. " Many experts also assume that the models will become very similar in the medium term anyway.
Open AI could quickly counter
Gemini Ultra won't be released until next spring anyway.
The US media "The Information" recently reported delays because Google's developers were dissatisfied with the AI's responses to non-English language queries. Gemini Ultra is still undergoing security checks and will be optimized before its release next year, says Google. Gemini Ultra should not be mixed with the Pro variant, says Bunk.
"With Gemini Pro, Google does not have a product that is comparable to GPT-4 today." According to Google, Gemini Ultra is better in the lab, "but I would be surprised if Open AI didn't also achieve similarly good results in its internal tests in the lab and just didn't communicate that." And by the time Gemini Ultra is released, Open AI could have already released a new version of its AI.
Nevertheless, Gemini Ultra has been “a fantastic model” overall. Gemini is Google's first artificial intelligence designed to be multimodal from the ground up. So it can not only generate and understand text, but also audio files, images, videos and code - and combine them. This allows the AI to better process differentiated information and answer questions on complex topics, according to Google. "This would be a major advance and a development that is eagerly awaited by many in the AI world," says Rasmus Rothe, founder of Merantix, a platform dedicated to artificial intelligence research and investment in building promising AI systems. However, ChatGPT has also been able to process voice or images in the paid version for months.
Gemini is also said to have significantly better programming capabilities than PaLM 2, which was only released in May. Google's chatbot Bard is also now available in English running a customized version of Gemini Pro. This is “the biggest quality improvement since the introduction of Bard,” writes Google. At the beginning of next year, an “advanced” version of the chatbot will offer access to Gemini Ultra.
New functions for Bard not in Europe for the time being
However, the new Bard is initially available in "more than 170 countries and territories", but Europe is not among them. Bard was originally launched much later in the EU than in other parts of the world. The delay was probably caused, among other things, by the strict European data protection regulations. Developers and enterprise customers can access Gemini Pro starting December 13th.
However, Gemini Nano is particularly interesting for consumers. Being able to use current language models on smartphones is "a great advance, with Google's announcement preempting Apple by several months," says Patrick Bunk. One function is, for example, combining WhatsApp group messages. “These language models will save us more time in life than with all other uses combined,” Bunk is convinced. The only downside: So far the model only runs on the latest Google smartphone." [1]
1. So will Google ChatGPT die Stirn bieten. Frankfurter Allgemeine Zeitung (online)Frankfurter Allgemeine Zeitung GmbH. Dec 7, 2023. Von Maximilian Sachse
Komentarų nėra:
Rašyti komentarą