Pointed out that interactive AI such as ChatGPT should be considered a `` word calculator '' rather than a `` search engine to find the correct answer ''
In recent years, high-performance interactive AI such as OpenAI's 'ChatGPT' and Google's 'Bard' have emerged and are attracting attention. Many people may think of AI using these language models as 'something like a search engine that will answer any question you ask', but software developer and blogger
Think of language models like ChatGPT as a “calculator for words”
https://simonwillison.net/2023/Apr/2/calculator-for-words/
Many people enjoy asking conversational AIs such as ChatGPT various questions and seeing how accurate or bogus the answers are. Conversational AI often gives responses that appear to be very accurate, and because Microsoft has AI in its search engine Bing, people think of conversational AI as ``something like a search engine.'' is also natural.
However, although interactive AI certainly returns a plausible explanation for the user's question, it cannot be said that the content is necessarily accurate. About this phenomenon, science fiction writer Ted Chan compared it to 'ChatGPT is a blurry JPEG image of the web' in an essay published in The New Yorker.
'Language models hold much of the information on the web in the same way that JPEG holds much of the information in high-resolution images,' Chan says. , it won't be found, it's only an approximation, but that approximation is usually acceptable because it's expressed in grammatically correct sentences, which ChatGPT is good at.'
“ChatGPT’s model is huge, but not big enough to hold all the exact facts encountered in the training set,” Willison said. 'That's not the case,' he said, pointing out that users should be skeptical of ChatGPT's answers and check the facts.
It is called 'Hallucination' that interactive AI such as ChatGPT generates answers that look like it but are not true. Of course, when humans use search engines to look up something, they can come across statements that are not factual.
However, humans can compare and compare multiple sources of information and consider what is true. It is dangerous to use conversational AI such as ChatGPT like a search engine because it may not be possible to compare information and arrive at the correct facts just by looking at one answer generated by conversational AI.
Mr. Willison answered the question, 'If it cannot be used as a search engine, what is conversational AI using language models suitable for?' I like it,' he said.
As the name suggests, the language model is a tool for handling 'language', and what is really good is language manipulation, not accurate answers to questions. Mr. Willison enumerates the following examples as utilization methods suitable for interactive AI using language models.
- Give an essay and summarize its content.
• Given a specific sentence, ask about the information contained in the sentence.
• Ask them to itemize the facts presented in the article.
• Rewrite sentences to make them punchier or more professional.
・Use it as an excellent thesaurus by asking the question, 'Tell me a word that suggests XX.'
・Use it for fun and creative play with text, such as ``Rewrite this sentence like a 17th century pirate'' or ``Make this sentence a rap battle''.
In addition, some interactive AI also have a function to perform a search on the question and include the search result in the answer. You can think of these search tools as an extended language model that helps you answer your questions more precisely.
However, even with interactive AI, including search tools, there is a risk of hallucinations, and there are cases where incorrect sentences generated by the language model are added to sentences that should have been derived from the search results. It is said that there is
Also, the analogy that ``language models are like word calculators'' has the flaw that ``although calculators always give the same answer, language models give slightly different answers to the same prompts''. I have.
Willison points out that many of the challenges with interactive AI are that it makes it look better than it actually is. “To get the most out of your language model and avoid the many traps it sets for unwary users, you need to spend some time interacting with it and learning how it works and how it works. We need to build an accurate mental model of what is possible and where it is easy to go wrong, and I hope this 'calculator for words' framing will help us do that.'
Related Posts:
in Software, Web Service, Posted by log1h_ik