The race to build ChatGPT-based search engine has begun, you can try them now
- Tram Ho
Jiang Chen, a machine learning expert who used to work at Google, was mesmerized when he first tried ChatGPT. This chatbot seems coherent in its responses and contains enough information to answer any question.
But the halo of the technology’s power has dimmed as Chen tries to use the same underlying technology that made ChatGPT to build a better search engine for the startup he co-founded, Moveworks. The company wants to use AI to help employees sift through information like technical support documents and HR pages.
The results show that Chen’s new AI search engine is excellent at gathering all kinds of useful information from the included documents, including providing addresses and phone numbers. But the problem is that some of them are not real. “Its self-manufacturing capabilities are amazing,” said Chen.
Meanwhile on the internet, extreme excitement around ChatGPT is running rampant and it’s not too confusing as many popular views say it could reinvent search engines. Obviously, this chatbot can provide complex answers to questions by aggregating information it finds in billions of words on the internet.
But, the way it works is essentially the opposite of the idea of a search engine that can reliably retrieve information found online. There’s a lot of misinformation on the web, but ChatGPT doesn’t recognize it, even then not only uses it, but uses it to create new misinformation. The underlying AI algorithms do not pull information directly from a database of events or associations, but instead just generate sequences of words for statistical purposes that resemble those of what is. seen in its training data. And of course, it doesn’t care about the truth.
Despite that challenge, and perhaps fueled by the excitement surrounding ChatGPT, the web search giants, as well as a number of startups, are defiantly moving forward. Microsoft, which has invested around $10 billion in OpenAI, the creator of ChatGPT, is expected to somehow add the technology to its search engine, Bing.
Google, which has been quietly working on a similar chatbot called LaMDA, is said to be trying to give Microsoft feedback. The company plans to release an AI chatbot soon and could introduce up to 20 similar products this year.
On the other side of the globe, Chinese companies also want their own smart and natural language systems for reasons ranging from language to politics. And the burden of creating a native ChatGPT will inevitably fall on the shoulders of this country’s tech giants.
After launching its Stable Diffusion-style art creation tool, Baidu recently announced that it is working on an AI chatbot called Ernie, which is expected to launch in March. As for the e-commerce conglomerate. Alibaba later also revealed that its “answer” to ChatGPT is being tested internally. WeChat’s parent company, Tencent, said it was conducting “relevant studies”.
Meanwhile, some startups have been quick to launch search engines with chat interfaces similar to ChatGPT. These include You.com, Perplexity AI, and Neeva.
These tools illustrated both the potential and the challenge of adapting ChatGPT-style technology for search purposes. You.com, founded by Richard Socher, a language and AI expert, can provide answers through a chat interface. Responses will be accompanied by citations, which can help users trace the source of a piece of information.
But this model sometimes combines sources that do not belong together. For example, when you ask about a specific person, an answer can be generated by combining information from the biographies of many people with the same name.
Another problem with a system like ChatGPT is that its responses are based solely on the data it has been trained on. Retraining the entire model can cost millions of dollars because of the size and scale of its data. That’s why YouChat is confused when asked about the latest sports scores but knows what the weather is like in New York right now. Socher did not want to disclose how the information was combined, seeing it as a competitive advantage.
“I think now a lot of these chat interfaces are far superior to the search engine experience in some respects, but in other respects they are clearly still much worse,” admitted Socher. receive . “We are working to mitigate all of these problems.”
Aravind Srinivas, founder and CEO of startup Perplexity AI, who previously worked at OpenAI, said the challenge to create a ChatGPT-like system with recent data credentials meant that will need to be combined with something else. “They alone will never be a good search engine,” he said.
Saam Motamedi, a venture capitalist at Greylock Partners who has invested in artificial intelligence-based search firm Neeva, said it remains unclear how compatible chat interfaces will be with the model. main revenue for search engines, namely advertising. Google and Bing use search queries to select ads that appear on top of the list of links provided in response. But Motamedi suspects that new forms of advertising may need to emerge for chat-style search interfaces to become possible, but he is not entirely clear what those will be. Neeva currently charges a subscription fee for unlimited ad-free searches.
The cost of running a model like ChatGPT on Google’s scale can also cause problems. Luis Ceze, co-founder and CEO of OctoML, a company dedicated to helping companies reduce the cost of implementing machine learning algorithms, estimates that running a search on ChatGPT can be 10 times more expensive than a search Google. Because each answer will require running on a large scale model and complex AI system.
The ChatGPT craze has also surprised some AI programmers and researchers. The bot’s core algorithm, called GPT, was first developed by OpenAI in 2018 and a more powerful version, GPT-2, was revealed in 2019. It’s a machine learning model designed to take text and then predict what will happen next. However, OpenAI has shown it can perform impressively if trained with huge volumes of text. The first commercial version of the technology, GPT-3, has been available for developers to use since June 2020 and can do many of the things that ChatGPT recently demonstrated.
Accordingly, ChatGPT used an improved version of the underlying algorithm, but the biggest leap in its capabilities came from the fact that OpenAI had human test subjects who would provide feedback to the system. about what produces enough answers to satisfy them. But like previous text generation systems, ChatGPT still has a tendency to reproduce misinformation from its training data, as well as produce results that look reasonable but are in fact incorrect.
Gary Marcus, professor emeritus at New York University and a vocal critic of AI hype, believes ChatGPT isn’t suitable for search because it doesn’t really understand what it’s saying. He added that tools like ChatGPT can cause other problems for search companies by flooding the internet with search engine-optimized text, generated by the AI itself. “All the search engines are going to crash,” he said.
Alex Ratner, an assistant professor at the University of Washington and co-founder of Snorkel AI, which trains AI models to make them more efficient, calls ChatGPT “a legit turning point” for what’s partly soft can do. But he also said it could take some time to figure out how to prevent language models like GPT from fabricating information. He believes finding ways to keep it up-to-date to keep the search engine fresh will most likely involve new methods for training basic AI models.
But how long those fixes will take to appear is still unknown. It may be a long time before this technology can completely change the way people search for answers.
“I told my team that people will see the difference between before and after ChatGPT,” said Chen from Moveworks. “But whether it will replace search engines is another question.”
Refer to Wired
Source : Genk