- Added Gemini preview report
Update December 5:
Google is reportedly planning a preview of Gemini for journalists and developers this week, The Information reports, citing a person familiar with the matter. The cloud version of Gemini is still scheduled for release next year.
Developer and investor Siqi Chen posted a screenshot on Twitter.com showing the potential versions of Gemini. There will be two different model levels, “Pro” and “Ultra,” each of which will also be available in a Vision version. The Vision version will certainly be able to describe images and possibly generate them as well.
Originally published on December 3, 2023:
Google delays its GPT-4 competitor “Gemini” until next year – Report
Google reportedly postponed the launch of its Gemini multimodal AI model.
This is reported by The Information, citing two sources familiar with the matter. According to the sources, Google has moved the launch events planned for next week in California, New York and Washington to January. The decision was made personally by Google CEO Sundar Pichai.
The delay came after Google engineers noticed that Gemini was less capable of responding to queries in non-English languages. Gemini has reached the level of GPT-4 in some aspects, but apparently still needs to be improved in terms of multilingualism. Smaller versions of Gemini are already being tested, while the large Gemini model is still under development.
It was previously leaked that Google would be delaying the cloud version of Gemini. Now, the AI-powered products, specifically the Bard chatbot, won’t get a Gemini upgrade until next year.
Google CEO Sundar Pichai recently announced a series of next-generation models that Google plans to roll out next year based on Gemini.
Why Gemini is so important to Google – and to the industry
The release of Gemini may be one of the most important product launches in Google’s history. The AI model is meant to prove that Google can keep up with, if not surpass, OpenAI, while paving the way for a new Internet in which the flow of information shifts from traditional search and the WWW to chatbots.
At the same time, Gemini’s success would be a signal to the industry that OpenAI’s GPT-4 is not yet the measure of all things, but that the underlying Transformer technology and scaling principle (more data, more training) still offer room for major advances.
Google’s access to data and compute should give it a leg up on OpenAI, but so far, it hasn’t worked out, in part because Microsoft is partnering with OpenAI.
In fact, since March 2023, no company, big tech or innovative startup, closed source or open source, has managed to release a model that comes close to GPT-4. Instead, the market is full of language models at the level of GPT-3.5, a standard that now seems easily achievable.
The fact that GPT-4 is so much more advanced is probably due to its larger, more complex, and more expensive architecture. GPT-4 uses many networked AI models (Mixture of Experts) rather than a single large model. Google Gemini is likely based on the same concept.
The complex architecture also comes at a high cost for inference, which is likely why OpenAI is trying to lower prices with models like GPT-4 Turbo, even if it means sacrificing some quality.