OpenAI could be about to take on Google at its own game with the launch of a new search engine based on ChatGPT technology.
Google has dominated the search market for more than two decades, accounting for 90% of all global searches but will its caution over generative AI be its downfall? This caution could leave space for a new player that mixes the understanding of AI with the breadth of data found online.
Very little is known about the rumored search engine beyond the fact it will be based on asking questions and interacting with an AI model using natural language. This is similar to the way Perplexity and Google Gemini work — bringing live data to AI.
The rumors stem from the registration of search.chatgpt.com. Reports suggest OpenAI is planning a big announcement on May 9 and that could well be the new search engine — although OpenAI can be even more tight-lipped than Apple on new products.
Why is OpenAI looking at search?
There is some live search functionality built into ChatGPT through its partnership with Microsoft Bing but its always been more of an “add-on” rather than a deep integration and as a result this has always been somewhat spotty and unreliable.
The intersection of LLMs plus search I don’t think anyone has cracked the code on yet. I would love to go and do that, I think it would be cool.
Sam Altman
During an interview with Lex Fridman last month, OpenAI CEO Sam Altman suggested that ChatGPT could be the future of search, with a deep integration between LLM and search.
He said: "If we can build a better search engine than Google then we should," but added it would understate what AI can actually be. "Google shows you 13 ads and ten blue links, which is one way to find information, but the thing that is exciting to me is that maybe there is a much better way to help people find and act on information."
“Maybe it is interesting to say ‘How do we help you find the information you need, how do we help create that in some cases or point you to it in others’,” Altman added.
He said integrating search into ChatGPT is an example of a cool thing to do. “The intersection of LLMs plus search I don’t think anyone has cracked the code on yet. I would love to go and do that, I think it would be cool.”
How might an OpenAI search engine work?
The new ChatGPT search is killing the "simply call conventional search and summarize with LLM" wrappers.Why? It takes actions. Try searching with "what did the most popular post in hacker news say today?" and you'll see the difference.If you simply translate that to a google… pic.twitter.com/Ynqs44cP2jMay 3, 2024
There are a few ways OpenAI could go with search. Perplexity takes a chatbot approach, allowing users to ask questions and then display AI-generated plus search result data.
Google Gemini is more of a pure chat result, offering up analysis from the AI and providing links to sources for that information, as well as a Google search link for the same terms. Google also has the Search Generative Experience experiment that creates AI-summarized results.
Rumors I’ve seen suggest OpenAI’s approach will be closer to that of Google Gemini, built into ChatGPT but with more live data. For example, a user could ask a question about the top song on the billboard chart at this moment and it will just show the result without a lag or caching.
What are the benefits of search to AI?
One of the biggest downfalls of large language models is the cut-off date for information. This is tied to when the training process finished and no new information could be added.
There are a few ways to tackle this problem including fine-tuning and expanding the context window so users can input their own datasets — but the best is live web access.
Google, Bing and Perplexity have their own spiders scouring the web constantly and updating databases on new results and pages. Some perform better than others.
This helps overall AI performance as this data can be fed back into the chatbot, allowing it to retrieve information in real-time, and then produce analysis based on its training data combined with the search results. But to work properly it requires deep integration.
