In 1998, Sergey Brin and Larry Page, two students who had just graduated from Stanford University, presented a first version of Google, a new type of search engine. At the same time they presented their PageRank algorithm. The operation is quite easy to understand. Specifically, it is possible to quantitatively measure the popularity of a website. The score is higher when it is referred to by a significant number of other sites.
We know the rest. Of course, the PageRank model relies on many other metrics and has evolved tremendously over time. However, this classification principle continues and remains the basis for most of the many tools available.
A radical change in our research habits
Google engineers want to change the way search engines work. The idea is to use a language model like GPT-3, but it’s much more advanced. Users would be prompted to ask a question and the tool could answer them directly.
So the change would be significant as Google would not point to a page with the information it was looking for, but could cross multiple sources to satisfy you. A bit of a human expert, this tool would generate real synthetic responses from multiple websites with references and sources, a bit like the way articles are written on Wikipedia.
Currently, the researchers believe that BERT or GPT-3 AIs are not entirely satisfactory. You are therefore working on future versions that might respond to more technical and specific problems in particular. The current tools are also English-oriented, which makes them inoperable for non-English-speaking target groups. They hope to change that.
Ultimately, and if scientists can overcome these pitfalls, it is indeed a radical change in our research experience that could see the light of day.