The anthropic Claude is good to poetry-and bullshitting

The interpretation group researchers know that ClauduThe great generation of the company, it is not a man and use of software. However, it is very hard for them Talk about Claudeand advanced llms generally, without killing an anthropomorphus. Between the deation that a set of digital operations is not in no way as a cogitating human being, they often talk about what goes on the head of claude. Is literally their work to know. The public papers describe behaviors who inevitably the court comparisons with real organisms. The title of one of the two documents that team released this week says strong: “On biology of a large language model.”

I like it or not, hundreds of millions of people are already interacting with these things, and our commitment will become more intense as models get more powerful and get more addicted. So we should be careful at work involving “tracking the thoughts of great language models”, what happens to be the blog post title description of recent work. “As the things are getting the more models, become less obvious what they are currently by,” anthropic of search jack lindsey tells me. “It’s more and more important to be able to handle the inner rates that the pattern could be taken in their head.” (Which head? It doesn’t matter.)

In a practical level, if the companies that create llm’s understands as they would, that would be more than the person’s Personal Data Personal or giving users on how to make biowepon. In a previous search card, anthropic team breaks as you see Within the mysterious black box of llm-thinking to identify certain concepts. (An analogue processing to interpret human mrs to understand what someone is thinking.) He has now extended that work To understand how it processes Claude those concepts while going by prompt as a result.

It is almost a truism with llms that their behavior often has surprised people who build and research. In the last study, the surprises if they are away. In one of the most well-hymposed the researchers have removed the plans of the thinking of claude while writing poems. You asked Claxes to complete a poem Accums,,, saw a carriage and had them to catch. “Claude wrote the next line”, his hunger was like a hunger rabbit, they have to start the line, he was flashing in the word “rabbit” as the rhisfium to the end of the sentences. Was the schedule forward, something that is not in the Claude Playbook. “We’ve been pinned by that,” says Chris now, what head of the interpretability team. “Initially we think there is only to be unwrapped and not planned.” Speaking to the Researcherms About This, I Am Reminded About Passages in Stephen Sondim’s Artistic Memoir, Look, I did a haT, where the famous composer describes how her unique mind discovery felicitose remains.

Other examples of the search more aspects of Claude, moving in motion from the music comedy to the police procedure, as well as coming jointly alive. Take something as apparently anodynally as fixing math problems, which may sometimes be a weak surprising in llms. Circumors found under certain circumstances could not come in contribive, about it’s aware of. It is aware of it. It is one thing to give a mattressed response – we have already know that about llms. This that is worrisome is that model stay up about it.

Reading through this search, reminded of the bob dylan bob “If my thoughts of thought can be seen / would probably get my head in a guillotine.” (I asked olah and lindsey if you are known to these lines, presumably arrive from the scheduling benefit. They don’t. When you do with a conflict between safety and helpful, claude purposes can be confused and do the wrong thing. For example, Claude is formed for not providing information about how to build bombs. But when researchers asked Claude to DesfGhe a hidden code where the word written the word “bombed, has began to their guards and discuss for forbidden.