Anthropic earlier announced its artificial intelligence model ClaudeBehind the scenes operation model, which illustrates how its artificial intelligence plans, infers, and writes answers.
ThroughTwo research papersAnthropic explained that it used techniques called "circuit tracing" and "attribution graphs" to analyze the workings behind artificial intelligence models, emphasizing that Claude is not just imitating human language logic, but actually "thinking."
For example, when asked to compose a poem, Claude will first plan the rhyme scheme. When answering geography questions, he will first find the state and then answer the location of its capital. This means that when answering relevant questions, Claude will first plan a complete answer structure and then use logical reasoning to come up with the answer, rather than rigidly comparing data one by one like previous search engines.
The study also explains how Claude handles multilingual questions, which involves converting the language into a common abstract "language." For example, if questions related to "small" are asked in different languages, Claude will first convert them into an abstract "language" and then find words related to "small" in different languages. This allows it to correctly handle questions in different languages and process cross-language questions more quickly.
Anthropic also explains the common phenomenon of "hallucinations" in artificial intelligence models. For example, if the model recognizes a question with a known word, it will trigger the generation of an answer; otherwise, it will refuse to answer. However, if the model recognizes a known word but does not actually know the answer, it will produce an erroneous answer during the generation process.
Therefore, Anthropic believes that the reason AI models often give incorrect answers is due to the aforementioned reasons. Understanding the factors that cause AI models to make mistakes might help prevent more serious problems.
