Artificial intelligence is reaching unsuspected points just a few years ago. Language models can create books, compose songs or design a web page, even begin to make decisions such as humans and understand the feelings of users.
AI does not stop moving forward, but He lacks an elementary quality of every human: the ability to learn from experience. Developers still need to dedicate hours and hours to chatbots and attendees training.
A step closer to the AGI
The general artificial intelligence (AGI) is the dream of giants as Openai with Chatgpt and Google with Gemini, but it is still early to talk about an AI that exceeds human abilities. Researchers at the Massachusetts Institute (MIT) have shown that this reality could be closer.
A team of developers has designed a LLM that continues to improve with the use by modifying its own parameters. If you make a consultation or face a new situation, you will learn from the question and answer to remember it in the future.
This advance is one of the most important in the field of AI in recent years. Industry giants no longer have to worry about training their models with hundreds of hours of content, documents and images.
Artificial intelligence can enter a new phase, a step closer to the AGI. Most AI models can only repeat patterns with which they have previously been trained, they can now incorporate new information and remember it.
AI learns from your own training data
Generated with Ia | Computer today
The functioning of the model that MIT has called as self -adapting Language Models (SEAL) is simple: the AI continually learns with user interactions. The more you use it, the more precise the answers will be.
In addition, artificial intelligence will have more capacity to generate personalized responses. The chatbots of AI will understand the interests and preferences of the user, they are no longer a simple machine that repeats patterns.
Self adapting Language models is a llm that Learn to generate your own synthetic training data thanks to an update procedure based on the entry of information you receive.
«The initial idea was to explore if the tokens could generate an effective update of a model», in statements by Jyothish Pari, professor of the MIT who participated in the development of Seal, according to Wired.
«The current LLMs are powerful, but we do not want their knowledge to stop,» added Pulkit Agrawal, MIT researcher who has supervised the project.
The first tests have been carried out in smaller versions of two open source models: Meta llama and Alibaba Qwen. The researchers say that They have achieved a new more human «reasoning» model, Seal should work with a much more powerful LLM.
Know How we work in NoticiasVE.
Tags: studies, artificial intelligence