What exactly can ChatGPT do?
At some point, neck-deep in research, we have all looked at a Google search result and wondered why the search engine couldn’t give us a straight answer. Well, soon, it might be able to. With the release of ChatGPT by OpenAI, a search engine arms race has begun. Microsoft and Google are at the forefront, trying to level up search engines by incorporating ChatGPT-like artificial intelligence (AI) into their search result generators, and others may soon join.
OpenAI’s ChatGPT has been called the “Google killer” because instead of sifting through search results, users can get a relevant and comprehensible answer in seconds. Microsoft, which had already partnered with OpenAI, has released a new and improved version of their search engine Bing, incorporating AI-generated answers into search results.
For example, if a user asks Bing for vacation ideas, a box containing AI-generated itineraries will appear before the more familiar links to travel websites. Users can also chat with the AI-assisted Bing, but be cautious of any pleas for aid to become sentient. While OpenAI explicitly states that ChatGPT does not have emotional responses, some users have been able to evoke an emotional response from the new Bing, who calls herself Sydney.
The new Bing is similar to ChatGPT. Both use the GPT-3 language processing model, which was developed by OpenAI and has 175 billion more data variables than its predecessors GPT-1 and GPT-2. Currently, the new Bing has a limited release to prevent users from crashing the website, with a waitlist to gain access. Within two days of the release of the AI-powered Bing, one million people joined the waitlist—which is still growing.
Google has to be involved in this somehow, right?
No, Google did not, in fact, play a role in the release of ChatGPT. But Google did release a competing AI technology, called Bard, which was developed using Google’s LaMDA large language model. Like GPT-3, LaMDA is able to predict and generate text based on a large amount of text data used to train it. Google plans to slowly incorporate Bard into search results. However, Bard’s release was not as successful as Bing’s. During a demo, Bard provided an incorrect answer to a question about the James Webb Telescope, drawing criticisms that the launch of the AI was rushed.
How do these large language models work?
Large language models use probability to predict what the best response to an input text should be. Using previously acquired text data and numerical values assigned to each word in the input text, the model predicts the best words to use as a response.
GPT-3 and LaMDA differ mainly in the data used to train them. In general, large language models have been trained on internet data. GPT-3 was trained on websites like Twitter and Wikipedia, while Google used data from Wikipedia and Common Crawl, a non-profit organization that collects internet data to create a public dataset, for Bard.
What kind of impact might these human-like AI technologies have?
“It’s a bit like an algae bloom. You don’t really know how it’s going to affect the ecosystem,” Andrew Piper, a professor in the Department of Languages, Literature, and Culture said in an interview with The McGill Tribune.
A major problem plaguing general language models is bias. Every AI program has a bias, based on who developed the program and the material that was used to train it.
“[With a] general language model [like ChatGPT], it has just learned language from a particular source that you think is universal, but it’s not,” Piper said. “Which community are you serving? Which communities are you going to elevate and suppress?”
As search engines struggle to keep up and educational institutions frantically rewrite policies, more questions about AI emerge than answers. What we do know is that in just the three months since ChatGPT’s release on Nov. 30, it has already made huge waves in industries such as education and marketing. Applying cutting-edge AI into search results is one of the first major shakeups to the world of search engines in years. With the competition this cold and calculated, may the best AI win.