ChatGPT’s capabilities have captivated organizations across industries, with many looking to leverage it in their operations. The rapid series of announcements from Microsoft and OpenAI have only fanned the flames, prompting speculation about how ChatGPT could disrupt Google’s search superiority.
Organizations will follow this lead, looking to integrate large language models (LLMs) into search engines. The LLMs which underpin AI apps, such as GPT-3 for ChatGPT, have huge potential to interpret and mimic naturalistic human writing.
Based on billions of parameters of data, LLMs can pick up on the nuances of human languages and interpret colloquialisms, subtexts, and nuanced questions. As such, their output can be equally rich and convincing, a huge progression from AI-generated text of years gone by.
Yet there’s a difference between being able to write convincingly and writing accurately. The very size that enables LLMs to create a human-sounding output also hinders their factuality, leading to hallucinations and false answers. In industries where factual accuracy is a priority, this is a concern, and that also applies to search queries.
The size of LLMs
LLMs from well-known providers like OpenAI and Google are trained on large amounts of example text scraped from across the internet. The scope and variety of this training data is intended to cover nearly every conceivable topic, allowing LLMs to develop an idea of how lay people speak.
Problems arise once these LLMs are moved into specialist fields. Industries such as material science, engineering, or manufacturing all have specific terminology that has an individual meaning to each field. This will vary drastically from how it is used in everyday conversation. This doesn’t just include definitions either: one discipline may have very different relationships between concepts and terms compared to the same words in another discipline.
Consequently, LLMs that are posed requests in specialist fields often end up gathering irrelevant information from an unrelated area. They will end up using incorrect definitions when sourcing information, or outright misunderstanding the question being asked of them.
These issues are amplified when it comes to search. Broadly, when a user asks a query of an internet search engine, they are not experts in the area they’re looking to gain an understanding. As such, the potentially inaccurate responses generated by the LLM could be taken as fact by the users. This is especially true if a response isn’t accompanied by citations to substantiate a claim.
LLMs have not been designed to enable search queries in niche fields. But there is an alternative: smart language models. These are models that are trained on high-quality, curated datasets.
Smart language models are designed from the ground up for factual accuracy. Intended to serve a particular field of expertise, the focus of their training data enables these models to understand domain-specific language and concepts. They are also able to cite the sources of information for these responses, assisting in explainability. The focus of smart language models is particularly useful in the context of search, as these models are better able to process the links, rankings, and ads intrinsic to search.
Smart language models provide a strong contrast to many of the ongoing issues found with LLMs in search, where there is no guarantee of factual accuracy and no citing of sources.
LLMs and Smart Language Models both have unique properties that could be the basis for the future of search, enabling personalized, human-like and interactive search capabilities. However, the technology will need to evolve until plausibility and factuality are done to an equal standard.
Today smart language models are more suitable for enterprise applications than LLMs because they focus on factuality, but LLMs plausibility will be important going forward to make the factual results easier to understand and digest by the users.