7th December 2023 – (New York) Google has revealed Gemini, its new artificial intelligence model intended to challenge ChatGPT and reestablish the search giant as the leader in AI technology. However, experts say Gemini exhibits similar flaws and limitations as other large language models dominating the field.
Unveiled on Wednesday, Gemini is described by Google as “natively multimodal” because it was trained on text, images and video, unlike most current AI systems focused solely on text. Google states Gemini is its largest and most powerful AI to date.
An initial version of Gemini will power Google’s chatbot Bard starting today for English language users. More advanced versions are slated for release in 2023, pending safety checks.
Google claims Gemini scores higher than ChatGPT-creator OpenAI’s top model GPT-4 on certain benchmarks for assessing AI abilities. However, it has not disclosed full details of Gemini’s architecture, training data or size.
While representing significant progress, experts say Gemini likely cost hundreds of millions of dollars to develop yet still struggles with core challenges of reasoning and common sense. “Anything using multiple modalities is a step forward. But these models likely show the same brittleness as ChatGPT in understanding the complexities of the real world,” said UC Berkeley professor Alexei Efros.
This continues an “arms race” between tech giants to dominate AI through proprietary models, algorithms and data, Efros added. But opaque development processes preclude independent validation of claimed capabilities.
Google is under pressure after OpenAI’s ChatGPT made conversational AI mainstream in late 2022. This prompted big investments from Microsoft and others, threatening Google’s lead in AI research.
Gemini aims to showcase Google’s progress, with multimodal capabilities intended to handle visual information better. For example, Gemini can summarise research papers with graphs or suggest games using a world map. However, analysts say its core flaws remain. “There’s no reason to disbelieve Gemini outpaces GPT-4, but the next version will likely surpass it,” said University of Washington professor Oren Etzioni. “This must-win war sees companies pour hundreds of millions into models with limited real-world reasoning.”
Etzioni said giant models like Gemini ultimately seek billions in cloud revenue, not solutions to AI’s fundamental problems. The lack of transparency around proprietary models also prevents independent testing, a challenge for Gemini and ChatGPT alike.
While negative outcomes from ChatGPT prompted Google’s caution, its response appeared slow until OpenAI sparked a generative AI frenzy in early 2022. Despite inventing key techniques in the field, Google hesitated to release its own chatbot. Gemini, named after NASA’s moon-landing precursor Project Gemini, now aims to retake the initiative through upgraded AI chips and visual capabilities. But true moonshot advances in reasoning and judgment remain elusive.
Flaws in today’s models persist, said ex-Allen Institute CEO Oren Etzioni. Google touts new safety measures for Gemini’s more general skills, but ongoing external research is needed to stress test its trustworthiness.