Google launches Gemini – a powerful AI model it says can surpass GPT-4
Google unveiled Gemini on Wednesday, aiming to compete with OpenAI’s GPT-4 multimodal AI model family, which powers ChatGPT’s premium edition. The largest version of Gemini, according to Google, outperforms “current state-of-the-art results on 30 of the 32 widely used academic benchmarks used in large language model (LLM) research and development.” It’s a successor to PaLM 2, an earlier AI model that Google intended to be on par with GPT-4 in terms of capabilities.
The Google Bard chatbot now operates in more than 170 countries with a specially tuned English version of its mid-level Gemini model; however, due to possible regulatory concerns, it is not available in the EU or the UK.
Gemini is multimodal, just like GPT-4, meaning that it can process various input kinds, or “modes.” It can therefore process audio, text, code, and even images. The objective is to create a kind of artificial intelligence that can precisely resolve issues, offer guidance, and respond to inquiries in a variety of domains, from the technical to the scientific. Google hopes to tightly integrate this technology into its products and claims it will power a new era in computing.
“Gemini 1.0’s sophisticated multimodal reasoning capabilities can help make sense of complex written and visual information,” Google states. “Its remarkable ability to extract insights from hundreds of thousands of documents through reading, filtering, and understanding information will help deliver new breakthroughs at digital speeds in many fields from science to finance.”
Also learn about other AI tools…..
According to Google, Gemini will come in three sizes: Gemini Nano (“for on device tasks” like Google’s Pixel 8 Pro smartphone), Gemini Pro (“for scaling across a wide range of tasks”), and Gemini Ultra (“for highly complex tasks”). The complexity of each is probably divided by the number of parameters. A larger neural network with more parameters can generally perform more complex tasks, but it also needs more processing power to operate. This indicates that while Ultra can only run on data centre hardware, Nano, the smallest, is intended to run locally on consumer devices.
“These are the first models of the Gemini era and the first realization of the vision we had when we formed Google DeepMind earlier this year,” wrote Google CEO Sundar Pichai in a statement. “This new era of models represents one of the biggest science and engineering efforts we’ve undertaken as a company. I’m genuinely excited for what’s ahead and for the opportunities Gemini will unlock for people everywhere.”
While Gemini will be offered in three sizes, the public can only utilise the mid-level model. As previously stated, Google Bard is currently using a customised version of Gemini Pro. Based on our preliminary informal testing, Gemini Pro seems to perform significantly better than Bard’s previous iteration, which used Google’s PaLM 2 language model.
Additionally, Google asserts that when Gemini is used with its proprietary Tensor Processing Units (TPU), it performs more effectively and is more scalable than its prior AI models. “On TPUs,” according to Google, “Gemini runs significantly faster than earlier, smaller and less-capable models.”
Google Gemini Vs. GPT-4,
Google has made previous attempts, such as with Gemini, to overtake OpenAI’s constantly improving GPT-4 model (now called “GPT-4 Turbo”). That was originally intended to be accomplished by the aforementioned PaLM 2, which was released in May. On paper, Gemini Ultra does perform better than GPT-4, according to Google, but not everyone is convinced. In its article about Gemini, MIT Technology Review expresses scepticism about Google DeepMind’s assertion that Gemini outperforms GPT-4 on 30 of 32 common performance metrics. However, there are not much differences between them. Demos indicate that it does a lot of things really well, but not many things that we haven’t seen before.”
How slender are the edges? The chart of eight machine learning benchmarks (MMLU, Big-Bench Hard, DROP, HellaSwag, GSM8K, MATH, HumanEval, and Natural2Code) that Google provides in its press materials aims to measure skills such as reading comprehension, multi-step reasoning, basic arithmetic, general knowledge in 57 subjects, and Python coding. With scores like 83.6 percent vs. 83.1 percent or 74.4 percent vs. 67.0 percent, Gemini Ultra outperformed GPT-4 in every metric with the exception of one (the wonderfully named “HellaSwag”).
Specifically, Google claims that Gemini Ultra is the first AI model to outperform human experts on the MMLU (massive multitask language understanding) benchmark, having scored 90 percent on the test that measures knowledge of 57 subjects, including math, physics, history, law, medicine, and ethics.
What does it all mean, though? Maybe not much to the typical person asking a question of Bard or ChatGPT-4. Google anticipates that these benchmark results will result in more accurate and helpful responses. Suppose you want to show Bard (through Gemini) a picture of your damaged bike in the hopes that it will be able to provide you with repair instructions. Will it ever be able to do that in reality? If not, is there any real significance to the benchmarks of 2 percent above GPT-4? There is currently a value dilemma in the AI field.
The effectiveness of machine learning benchmarks is still being investigated and debated, even among machine learning researchers. Their use can also be contentious because it may be possible to test an AI model on data that may be present in the benchmark’s data set. Therefore, it’s crucial to interpret any metrics similar to these extremely cautiously.
For the time being, Google anticipates that Gemini will serve as the first shot in a fresh conflict between companies like Anthropic, Meta, and Microsoft and OpenAI over control of AI assistants in the future. Further details about Gemini’s operation and its potential in science domains can be found on the Google DeepMind website.