Artificial intelligence (AI) continues to grow. Since the arrival of ChatGPT in the digital world, the technological giants have had to reinvent themselves when it comes to the services and products they offer. We have seen how Microsoft integrated AI into its search engine to make Bing one of the most contextualized search engines. For its part, Google has done the same with Bard. Now, Meta is entering the world of AI with LIMA.
Now, those from Meta are also joining the AI trend. A few hours ago, the company led by Mark Zuckerberg announced a new AI-based language model variant called LIMA (Less Is More for Alingment). This new tool promises much more advanced performance than ChatGPT.
The company explained that the large language tool (LLM) training process is what makes LIMA the “most advanced LLM on the market,” according to Meta. To understand the process, it is necessary to remember that LLLM models such as ChatGPT and Bard are trained in two phases.
In the first phase, the AI is given a number of prompts to understand the “basic” concepts. In the second phase, human intervention enters with reinforcement learning methods with human feedback. In this phase, unnecessary information is purged and training is carried out to guide the operation of the tool to fulfill specific tasks and user requests. It is worth mentioning that this second phase is the most expensive of building an LLM model.
It may interest you: Chinese rival of ChatGPT censors responses: he cannot talk about COVID-19 or President Xi Jinping
Unlike ChatGPT and Bard, LIMA, Meta’s new AI tool, skipped this second phase. According to the company, what they want to verify with LIMA is that an AI model extensively trained in its first phase does not need the second phase. For example, for this new Meta model, only 1000 prompts were selected. These were handpicked from sources like WikiHow, StackExchange, and Reddit.
After completing the first phase and before announcing the new tool, Meta allowed a group of (human) experts to test LIMA against ChatGPT and Bard. According to the report, LIMA was able to answer all the questions correctly and returned 43% answers equal to or better than those of the OpenAI (ChatGPT) and Google (Bard) AI models.
Image: Pexels
