A new future-oriented intelligent conversation experience—Claude2

Claude Assistant Development History

  Anthropic was founded in 2021 by 10 former OpenAI employees led by the former OpenAI Research Vice President. There are both Tom Brown, chief engineer of GPT-3, and Daniela Amodei, vice president of security and policy at OpenAI, which can be said to have taken away a considerable number of core talents. One of the reasons for leaving to set up a new company is naturally dissatisfied with the status quo of OpenAI.  Since the past few years, Microsoft has frequently injected funds into OpenAI, and then asked them to use Azure supercomputing for research, and licensed the technology to Microsoft, and even raised funds for Microsoft's own investment activities. This is contrary to the original intention of OpenAI, and a group of employees thought of leaving to start a business. Of course, part of the reason is that this group of people wants to make controllable and explainable AI. To put it bluntly, it is to understand the principles behind the AI ​​model first, so as to design more explainable AI models while providing tools. So, after OpenAI completely became a "Microsoft money-taking machine", they left the company and founded Anthropic. Claude's goal is to be an AI system that is safe, close to human values, and ethical. At present (April 2023), Anthropic currently has a team of about 80 people, with a financing amount of more than US$1.3 billion and a valuation of US$4.1 billion. Claude is designed and built on the basis of Create, a large-scale language model independently developed by Anthropic. Create is trained using a huge and diverse training set collected by Anthropic itself, and employs an original self-supervised learning method. Compared with the supervised learning method adopted by the GPT model behind ChatGPT, this method can better promote the model's general understanding of various scenarios, strengthen common sense reasoning ability, and better learn human interaction patterns. After continuous iterative optimization, Claude Assistant has released multiple versions. The latest version integrates the latest capabilities of Create, which can provide a smooth, knowledge-rich, and context-sensitive English voice interaction experience. In terms of models, Anthropic claims that its scale has exceeded 17.5 billion parameters, equivalent to 1.5 times that of GPT-3.

Claude feature update

Experience URL: https://claude.ai/

• Memorize 100,000 tokens at a time, equivalent to 75,000 words

Claude's context window has been extended from 9K tokens to 100K tokens (Claude 2 has been extended to 200K tokens, but the current release only supports 100K tokens). The upgraded Claude-100k version has greatly improved dialogue and task processing capabilities. On the one hand, it is the increase in the "amount of text that can be processed at one time", which directly broadens the types of positions that Claude can work on. Previously, large models were used to process documents of dozens of pages at most. Now, Claude has been able to speed-read company financial reports, technical development documents, identify risks in legal documents, read hundreds of pages of research papers, and even process data in the entire code base. The most important thing is that it can not only read through the full text and summarize the main points, but also complete specific tasks, such as writing code and organizing tables. Claude can be your "code companion", and you can make a demo in minutes. For example, upload a 240-page Langchain API document, and let it use Anthropic's language model to make a simple demonstration of Langchain based on this document.

In addition, the Claude100k can handle about 6 hours of audio volume. For example, the content of a Musk podcast was transcribed into a text of 58k tokens, and then Claude was used to summarize and answer questions. On the other hand, the increase in "memory" has brought about an improvement in the control of topics and an improvement in chatting ability. Previously, large models often "forgot the topic while chatting". After the total number of words in the dialogue window exceeded a few thousand words, they began to speak nonsense. But now, Claude, who has a memory of 100,000+ tokens at a time, is unlikely to have such a situation. Instead, he can firmly remember the topics he talked with you, and chat for several days in a row.

• Claude's training data is mainly English, but the proportion of non-English data in Claude 2's training data has increased significantly. After testing, it is found that claude's ability to understand Chinese is much better than that of chatgpt.
•  Claude 2's training data includes updated data for 2022 and early 2023. That means it knows more about things like internet news.

Performance Testing

Evaluated Claude 2, Claude Instant 1.1, and Claude 1.3 on standard benchmarks, including Codex HumanEval for python function synthesis, GSM8k for solving elementary school math problems, MMLU for multidisciplinary question answering, and QuALITY for long story question answering , ARC-Challenge for scientific questions, TriviaQA for reading comprehension and RACE-H for reading comprehension and reasoning at the middle school level. The specific evaluation results are shown in the following table:

It is worth noting that Claude 2's ability to generate code has improved significantly, and its score on Codex HumanEval has increased from 56% to 71.2%. The study also tested Claude 2's practical ability with several common qualifying-level exam questions. First, Claude 2 scored 76.5 percent on the Bar Exam multiple-choice test, higher than Claude 1.3's 73.0 percent.

Second, the research team also tested Claude 2's proficiency level on the Graduate Record Examination (GRE). Claude 2 scored above 90 percent on the GRE reading and writing tests, and reached the level of candidates who took the GRE test in quantitative reasoning. Median level.

Finally, the study also tested Claude 2 on United States Medical Licensing Examination (USMLE) questions

Anthropic said that companies such as AI writing platform Jasper and code navigation tool Sourcegraph have begun incorporating Claude 2 into their operations.

Summarize

To sum up, we can see that in the field of artificial intelligence, Claude 2 has become a powerful new competitor that cannot be ignored with its excellent natural language understanding ability, rich knowledge question and answer and friendly interaction. ChatGPT. The emergence of Claude 2 brings new thinking and possibilities to human-computer interaction. I believe that with the continuous improvement of its capabilities in the future, it will surely bring us a more intelligent and humanized voice interaction experience. A new future-oriented intelligent conversation experience—Claude2