Google with Gemma will challenge Meta’s Llama (and also ChatGpt)
Google is also launching itself into open source artificial intelligence with two models that are free and available for researchers and developers. But what are the pros and cons of open and closed models? All the details
One of the best known to offer an open source generative artificial intelligence (AIg) model, meaning it can be used for free by developers and researchers, was Mark Zuckerberg with his Llama launched a year ago . Now Google is also entering the field by launching Gemma 2B and Gemma 7B, two open models, which also compete with OpenAI's ChatGpt which, unlike them, is closed, meaning it can only be modified by the company that owns it.
GOOGLE'S OPEN SOURCE AI TEMPLATES
About ten days after the announcement to put all its AI products under the Gemini name, Google presented two new open source models of generative artificial intelligence, Gemma 2B and Gemma 7B, which it makes available to any developer , which can modify and optimize them as needed.
Both templates are available under commercial license regardless of company size, number of users and project type. Furthermore, Google has promised the release of new variants in the coming weeks.
“Made with the same technical, infrastructure and research components used to create Google's Gemini models, Gemma offers best-in-class performance compared to similarly sized open models,” the company notes. Despite their small size, in fact, according to Punto Informatico , "they offer performances comparable or superior to those of larger models".
The launch also follows last week's unveiling of Gemini 1.5 Pro, available to enterprises via the Vertex AI platform, where the total number of first- and third-party models available in the Vertex AI Model Garden comes to over 130 , “one of the most greats in the sector".
GEMMA VS OTHER OPEN SOURCE MODELS (ACCORDING TO GOOGLE)
“Based on the benchmarks published by the Mountain View company – writes Punto Informatico -, the Gemma 7B model offers similar or superior performance to the Llama 2 models […] by Meta and Mistral […] by Mistral AI in four activities: response to questions, reasoning, mathematical calculation and writing code”.
Google has also added filters to prevent the use of models for dangerous activities, such as weapons development.
OPEN OR CLOSED? THIS IS THE DILEMMA OF AI MODELS
Both Google and Meta – but also the French Mistra AI – are approaching the 'open' world in order to undermine OpenAI and its closed model. Open source models allow those who use them to create software, or chatbots, based on artificial intelligence and capable of simulating a human conversation by answering user questions. “In this way Google – and previously Meta – hopes to convince developers to use and test its technologies and to rely on its cloud services, i.e. online data management,” writes Il Post .
Although the heart of the debate in the debate between open and closed source is the democratization of technology, security and profit take precedence. While open software encourages creativity, innovation and democratization, on the other hand, according to supporters of closed models, it threatens our security.
OpenAI itself, Euronews recalls , was originally founded to produce open-source artificial intelligence systems, but in 2019 the company declared that it was too dangerous to continue releasing its Gpt language model to the public, as it was too similar to speech human and, if put in the wrong hands, could have generated high-quality fake news or given dangerous indications. This is what happened with a modified version of Llama who was asked how to make a nuclear bomb.
In fact, the thesis is that closed source models are safer because they have usage policies that protect against dangerous use of AI. Even in these cases it is possible to jailbreak policies, i.e. violate them to bypass protections, these vulnerabilities are often fixed quickly.
These critical issues also affect investments; in fact, open source companies, inspiring greater confidence in security matters, raise funds more easily and have greater capital to innovate.
This is a machine translation from Italian language of a post published on Start Magazine at the URL https://www.startmag.it/innovazione/google-con-gemma-sfidera-llama-di-meta-e-pure-chatgpt/ on Thu, 22 Feb 2024 14:48:14 +0000.