
The wide possibilities of AI, since its creation, have always raised a multitude of questions. Data is processed and recorded in their black boxes – in reference to when there is something they do not understand – and researchers at the artificial intelligence or AI company Anthropic, creators of the Claude system, claim to have made a breakthrough in understanding and knowing exactly how the language model they use works. Research in which they also seem to have discovered things as disturbing as they are unexpected.
Anthropic’s findings on Claude
An emotional moment! Father gets caught watching his son’s goal in the middle of an interview
Anthropic has analyzed ten different behaviors in Claude and obtained interesting results. One of them was the use of different languages: “Does Claude have a part that speaks French and another that speaks Chinese, and so on?” They asked. A question to which they have already obtained an answer, as they identified that it used components independent of any language to answer a question or solve a problem and then chose a specific language when responding.
Claude was asked: “What is the opposite of small?” in English, French and Chinese, and it was found that it first used the components independently of the language to give an answer. This is a deduction it made before responding in a particular language. This suggests that the AI uses things in one language and then learns and applies them in others. This was also used to solve simple mathematical problems and it was noticed that it developed its own strategies and therefore different from those it has seen in its training data.
Some behaviors are listed in the MIT study and seek to resolve these concerns and transform them into certainties: “These findings are not only scientifically interesting, but also represent a significant step towards our goal of understanding AI systems and ensuring that they are reliable,” the company explained.
It currently takes several hours of human effort to understand the circuits we see
The surprising discoveries about AI
There was a case where they used Claude to write poems where they found that the AI itself always looks ahead and chooses the word at the end of the next line, not just improvising: “We set out to demonstrate that the model did not plan ahead, and we found that it did,” they concluded at Anthropic. And they explained another example: “In a response to a jailbreak example, we found that the model recognized that it had been asked for dangerous information long before it was able to steer the conversation.”
A technology that many see benefits to: “It currently takes many hours of human effort to understand the circuits we see, even in indications with only tens of words. To get to the thousands of words that support the complex chains of thought used by modern models, we will need to improve both the method and (perhaps with the help of AI) the way we make sense of what we see with it.”
Doubts in Google’s AI program
They are not the only ones developing similar technology, in fact, Google is also doing so. And it is here that its own executives acknowledged and explained that it is normal not to understand all the processes by which an AI arrives at a result. An explanation for which they used an example since the company’s AI program adapted itself after being asked in the language of Bangladesh “which it was not trained to know”.
An unsettling glitch where the answer did not leave the CBS journalist at ease when interviewing Google’s CEO, Sundar Pichai: “You don’t fully understand how it works, and yet you’ve made it available to society?” he asked with great concern. And he replied: “It’s not a big deal, I don’t think we fully understand how the human mind works either“.
This news was originally published on this post .
Be the first to leave a comment