Image: Google.
Last week, Google made several announcements regarding its artificial intelligence products: Google Bard changed its name to Gemini – named after the LLM that powers it – and a paid subscription for the AI chatbot saw the light of day , named after Gemini Advanced. Continuing its momentum, Google has just announced a new AI model.
A model comparable to Gemini Ultra
This Thursday, Google unveiled its new generation model, Gemini 1.5. Even though Gemini 1.0 is quite new, having been launched in December, the new model presents considerable improvements over its predecessor, among which an enlarged pop-up window and an overall improvement in performance and understanding of the AI.
The model is so advanced that Google CEO Sundar Pichai compared it to the company’s most advanced model to date: Gemini 1.0 Ultra. According to him, the new model, just launched in preview, achieves quality comparable to the Ultra model announced last week, while using less calculation.
“Longer pop-ups show us the promise of what’s possible,” adds the Google CEO. “They will offer entirely new capabilities and help developers create much more useful models and applications. »
Gemini 1.5 can absorb much more information than its predecessor
To achieve this increased performance, Gemini 1.5 was built on a new version of the Mixture-of-Experts (MoE) architecture, which allows the model to selectively learn and activate the most relevant pathways in its neural network, thereby increasing its effectiveness, according to the press release.
Google claims that Gemini 1.5 Pro can produce up to one million tokens, which is a huge increase from Gemini 1.0’s 32,000 tokens. This increase is notable because the model’s pop-up window – that is, how much information it can absorb – is made up of tokens. Therefore, the more tokens a model can absorb, the better and more informed its responses are likely to be.
Image: Google.
Google says Gemini 1.5 Pro can process large amounts of information in one go, “including up to an hour of video, 11 hours of audio, and bases of over 30,000 lines of code or more of 700,000 words.
Impressive results in benchmark tests
During its demonstration, Google provided Gemini 1.5 Pro with a 44-minute silent film of Buster Keaton. The AI model was able to quickly process this information and then answer all kinds of questions, including multimodal queries, as can be seen in the video below:
The model also achieved impressive results in benchmark tests. It outperformed version 1.0 Pro in 87% of the benchmarks used by Google to develop its LLMs. Gemini 1.5 Pro also achieved notable results in the Needle In A Haystack (NIAH) assessment and Machine Translation from One Book (MTOB) tests, which test the model’s acuity and learning capabilities.
Ensuring safe and responsible deployment of the model
To reassure users of the new Gemini 1.5 model, Google also clarified that extensive assessments had been carried out to ensure the safe and responsible deployment of this advanced model.
Gemini 1.5 Pro was made available free to developers and enterprise customers as part of a limited preview, through AI Studio and Vertex AI. The model comes with a million token pop-up.
When the model is ready for wider distribution, Google plans to introduce it with multiple subscription tiers, which is expected to start at 128,000 tokens and go up to one million tokens.
Source: ZDNet.com