Artificial intelligence Claude successfully passes the law exam

“Helpful, harmless and honest” is how Anthropic’s new artificial intelligence (AI) program “Claude” describes itself. However, like its relative ChatGPT, this program is already beginning to cause concern for some people. In particular, he had just passed the exam in law and economics. Pretty good student, this Claude… Except for the fact that he’s not human, of course. While the ChatGPT AI has been in the media spotlight, it has been quietly preparing to hit the scene at the end of 2022. Developed by Anthropic, at first glance it works in much the same way as OpenAI’s ChatGPT. All you have to do is ask him a question and he will start chatting and giving you answers. And these answers can be significant! To test his abilities, the researchers asked him to answer questions in an exam in law and economics. The copy received at this request was blindly corrected among other copies at George Mason University. She managed to get a score that allowed her to pass the exam. Alex Tabarrok, professor of economics, posted one of the answers on his blog. ” Most impressive ,” he said. It was a matter of intellectual property law. The AI ​​was supposed to make recommendations for changes to the law. ” In general, the goal should be to make intellectual property laws less restrictive and make more works available to the public faster”, – concluded Claude. “But it is important to continue to provide incentives and compensation to creators for a limited period of time.” Even if this professor admitted the answer is “better than many people”, some of them are more critical. We see in the comments to the article a lot of Internet users who note that AI contradicts itself in some of the answers. ” Honestly, it looks like Claude just swallowed and tore out the McKinsey (international consulting company) report ,” the Financial Times also criticizes. So, Claude is making a big deal out of molehills? Not necessary. In fact, scientists developed it with the goal of working on the idea of ​​”constitutional AI”. This is where the main successes of researchers lie. So, what is constitutional AI and what is its essence? ” Often, linguistic models trained to be ‘harmless’ become useless when confronted with controversial questions ,” the company explained. ” We are experimenting with methods to train a harmless AI assistant through self-improvement, without any human marks identifying a harmful outcome ,” they explain. ” The process includes both supervised learning and reinforcement learning .” In other words, this AI will be able to determine when a statement is, for example, discriminatory by self-learning, following not a human-made label, but broad principles: “constitutions,” in short. Apart from this difference, Claude is a serious competitor to ChatGPT. The company Scale, which develops applications based on artificial intelligence, undertook to test them on the same types of requests. Among other things, it has been found that Claude is more likely to recognize when he is inaccurate: on the other hand, he is less good than ChatGPT at writing computer code. However, both AIs still show weaknesses, answering incorrectly and sometimes even making up facts. ChatGPT, like Claude, has a hard time summarizing the various seasons of Lost without making up facts. But can you blame them for that? “Like most people who have watched the show, ChatGPT and Claude seem to have a fuzzy memory of Lost at best, ” the Scale testers conclude, not without a bit of humor.