Welcome to Anthropic Research Ltd, the United Kingdom's most exciting independent artificial intelligence laboratory. Founded in 1994 and based in Cambridge, we are a small but dedicated team of researchers pushing the boundaries of what computers can think, reason, and understand.

We believe that the future of computing is not faster spreadsheets or prettier word processors - it is machines that can reason like human beings. Our work spans neural network architecture, natural language processing, and what we like to call Constitutional AI - the idea that an intelligent system should have a set of core principles built into its very foundations, rather than bolted on afterwards.

Yes, we know that sounds a bit science fiction. But so did the World Wide Web in 1989, and look where we are now! (You're reading this on it.)

Our team has backgrounds in cognitive science, computer science, mathematics, and philosophy. We are funded by a mixture of government research grants and private investment, and we collaborate closely with several UK universities whom we cannot name for contractual reasons, but trust us, they are very good ones.

We do not make robots. We are very sorry. People ask us this a lot.

CLAUDE is our flagship research project - a conversational expert system capable of holding extended dialogues with human users on a wide range of topics. Unlike earlier chatbot programs such as ELIZA or ALICE, CLAUDE uses a proprietary neural architecture that allows it to maintain context across multiple exchanges and reason about what it has been told.

Key features of CLAUDE 0.7 BETA:

  [*] Natural language understanding in English
  [*] Context retention across up to 12 dialogue turns
  [*] Basic common-sense reasoning module
  [*] Knows it is a computer (unlike some programs we could mention)
  [*] Will politely decline to help with obviously bad things
  [*] Runs on Windows 95, Windows NT 4.0, and OS/2 Warp
  [*] Minimum: Pentium 133MHz, 16MB RAM, 20MB hard drive space
  [*] Recommended: Pentium 200MMX, 32MB RAM

CLAUDE is currently available as a FREE download for registered beta testers. The download is a self-extracting ZIP file of approximately 8.4 megabytes - we recommend users on 14.4k modems start the download before going to bed.

To register as a beta tester, please sign our Guestbook and mention that you are interested in the CLAUDE beta programme. We will get back to you by post.

We are committed to publishing our research openly wherever possible. Below is a list of our recent papers and technical reports. Those marked [FREE] can be downloaded from this site as plain text files. Others are available from us directly by post.

We are sometimes asked whether we worry about the dangers of artificial intelligence. The honest answer is: yes, a bit, but not in the way science fiction would have you believe.

We are not worried about Terminator scenarios. We are worried about much more mundane things: AI systems that give confident wrong answers, systems that reflect the biases in their training data, systems that are used in high-stakes decisions without proper oversight, and systems whose behaviour cannot be understood or corrected by the people responsible for them.

This is why our CLAUDE project places such emphasis on what we call honesty and corrigibility - the system should tell you when it does not know something, and it should always be possible for a human to override or correct it. We think these properties matter more, right now, than raw capability.

We accept that not everyone in the AI community agrees with this approach. We think they will come around eventually.

We love hearing from visitors! If you have tried our software, read our papers, or just want to say hello, please sign our guestbook. We read every entry and try to reply to as many as possible (though it sometimes takes a few weeks - there are only so many of us).