The release of ChatGPT has triggered a new wave of discussions about the end of the era of traditional search engines. Many people believe that AI-powered chatbots have a real potential to replace traditional search engines. In this article, I want to discuss whether such conversations are speculation.
What is ChatGPT
ChatGPT is an advanced chatbot designed to generate human-like responses to text input. ChatGPT is based on the large language model (LLM) called GPT, created by OpenAI. LLM is a system that is trained on string prediction tasks. It tries to predict the likelihood of a character/word/string given either its preceding context or its surrounding context.
What makes ChatGPT so good for users? Why do people think it will replace Google Search?
The critical difference between ChatGPT and Google Search is how ChatGPT approaches the search. When a user types a search query into the search box, Google Search provides a list of resources that are supposed to answer the question. The responses ranged from the most relevant (first few search results) to less relevant. But the user has to manually go through the search results and find the answer to their question. ChatGPT, on the other hand, does the work for the user; it mimics the process of data analysis and offers a ready-to-use answer to the user.
Why didn't ChatGPT replace Google Search already?
ChatGPT has many limitations preventing it from replacing a search engine. We won't dive into many details about the model itself; if you're interested in learning more about ChatGPT, consider reading this article.
In this section, I want to focus on the five limitations that directly impact the quality of the search.
1. ChatGPT is based on the language model GPT that requires training to perform well. Training is an expensive procedure, so OpenAI trained the model only once, and despite all errors that this model still has, they didn't retrain it.
2. ChatGPT operates using a snapshot of data from 2021. It doesn't know anything about things that happened after 2021. If you ask the system which country won the World Cup in 2022, it won't tell you that it was Argentina.
For the same reason, there are many categories of questions that ChatGPT cannot help with. For example, ChatGPT cannot answer questions that require location awareness (questions like “What is the weather today?” or “What are the best places in my area?”) or access to specific types of information (shopping reviews).
3. ChatGPT has problems with factual correctness. It can easily provide incorrect information, so it cannot be used for anything that requires factual actuality. When responding to user questions, GPT might even invent things, and it can be dangerous because the output it provides looks very trustworthy (it feels like the tool knows what it's talking about).
4. ChatGPT requires a lot of computational power and has problems with scalability. A single AI answer costs more than ten regular Google search requests. The daily cost of running ChatGPT is $100k. It’s possible to make it work for a few million daily users, but scaling it up to billions of people will be extremely expensive.
5. It's impossible to verify the accuracy of information. ChatGPT doesn't provide any sources that are used to generate the response. For people to trust the system, they need to understand where the information comes from and see the references.
OpenAI might be able to resolve many limitations that ChatGPT has right now. For example, if the company finds a less expensive way to train the language model, it will be able to release a model that continuously learns and improves itself.
Why didn’t Google create a rival?
When Sundar Pichai, Alphabet CEO, was asked about Google's plans to release a rival of the ChatGPT, he mentioned that Google won't launch it because of 'reputational risk.' Many assume that Google doesn't have anything like GPT in its arsenal. But it's far from the truth.
Google integrated many LLMs into their products. For example, MUM (Multitask Unified Model) and LaMDA (Language Model for Dialogue Applications) are used in Google Search and YouTube comments to improve search results and the range of user comments. But LLMs that Google uses do their work behind the curtains so that the results of their work might not be noticeable to the mass audience.
So, why was Google not the first company to release something like ChatGPT? And why does Google avoid doing that right now? I think that what Sundar Pichai means by 'reputational risk' is that large enterprise companies cannot afford to release half-baked products. If Google releases a product that performs like ChatGPT with all its limitations, the company will face many legal penalties. Startups like OpenAI, on the other hand, have more freedom to release products like ChatGPT that are basically 'research products' rather than 'consumer-ready products.'
Predictions for the future of search
OpenAI spends around $100k daily to make ChatGPT publicly available. If OpenAI doesn't close public access to ChatGPT, it will coexist with Google Search for some time. ChatGPT will be used for smaller tasks that don't require access to timely updated information and are less prone to fact-checking. But it won't kill search engines because LLMs cannot help with many categories of queries, such as weather forecasting.
The hype around LLM will change how companies think about search. Users' expectations will evolve, and they will start to ask new things. It might motivate Google to implement a GPT-like language model in their system. I think Google Search will look very different in a few years. Most likely, Google will release a hybrid version of its search engine. The next generation of search will be a mix of a traditional search and AI assistant that will be based on one of Google's LLMs (such as the LaMDA model). Google did plenty of research in this area and even had a product called Google Assistant, which is a chatbot (surprisingly!).
Another important thing worth mentioning is how the mass audience perceives AI. Even in the tech world, many people believe that ChatGPT is a true AI engine, but in reality, it is a large language model (LLN). GPT might be only the first step towards what we call AGI (Artificial General Intelligence), a true AI that can have common-sense knowledge. Right now, language models don't have a true understanding of the world and they are prone to hallucinating. So it makes me wonder whether LLN will be relevant in the long-term perspective.