April 05, 2023

Is Google Losing the AI Race?

The buzz around Artificial Intelligence (AI) has increased significantly in recent months after research company OpenAI publicly launched ChatGPT, a generative AI-based chatbot that can do everything from answering mundane queries to more complex tasks like writing code and doing your homework. Microsoft has invested $13 billion in OpenAI, and has integrated ChatGPT’s AI capabilities with Bing’s search engine. Naturally, this has prompted questions about the implications for Google, which dominates the search industry with over 90% market share.

Because of the Microsoft/Bing connection to ChatGPT, some feared that Google might be trailing Microsoft in its AI capabilities. Is it losing the race? And what does that mean for its 90%+ share in search?

The truth is Alphabet (Google’s parent) has been investing in AI for years, which has been helping it deliver the most relevant search results to queries (making for a better user experience) and providing a higher return on ad spend for advertisers. Alphabet has in fact developed several large language models in Google search. These include:

• Bidirectional Encoder Representations from Transformers (BERT): BERT allows the search engine to consider each word in the context of the other words in a sentence, rather than each word on a standalone basis. This context enables the search algorithm to provide more relevant search results. An example of BERT in action can be seen in the example below, where the question is whether you can pick up medicine for someone else, when it’s in their name. The AI-enabled search on the right interprets the question correctly; the other does not. 

• Multitask Unified Model (MUM): MUM is a multimodal AI tool that can process information across different formats to provide insight. For example, in the query below, the AI will source information about Mount Fuji such as altitude, terrain, and weather conditions – and then interpret that data to provide an assessment of whether these particular boots are suitable to hike Mount Fuji.

• Language Model for Dialogue Applications (LaMDA): Similar to BERT, LaMDA is a language model trained to contextualize words next to each other in a sentence or paragraph. The main difference is LaMDA is trained in “dialogue,” so it can hold open-form conversations that take into account previous queries in the exchange in its assessment. Alphabet started rolling out Bard, a LaMDA language model and its answer to ChatGPT, in the US and the UK in late March.
  

If Alphabet had the technology for a conversational AI tool, why were they late to the punch in rolling it out (letting ChatGPT grab the headlines)?

We believe there are two reasons. The first is credibility. As good as they are, the many “Fail” examples on the web of both ChatGPT and Bard highlight that these generative AI tools are still working out their kinks. Google likely did not want to jeopardize its credibility as the leading search engine in the world.

But the bigger reason we believe is likely cost. The compute costs associated with generating responses to queries are significantly higher using generative AI tools. According to Alphabet Chairman John Hennessy, a search query using generative AI costs ~10x as much as a standard keyword search – putting a potential dent in Google search’s estimated 80%+ incremental gross margins. It is worth noting, though, that some of the margin-dilutive aspects of using large language models to power Google searches are already embedded in Alphabet’s financials, since a portion of Google searches have been relying on these for some years. In addition, Alphabet CFO Ruth Porat mentioned at a recent conference that Alphabet has teams dedicated to improving efficiency in this area.

It remains early days in assessing how much traction generative AI ultimately ends up getting, and how well Google defends its market share against Bing and any other competitors. While we don’t expect all search queries to utilize generative AI in the near-term, there are a significant number of Google searches that do not currently end with a single click. These are the ones likely to most utilize AI solution to provide a more relevant answer.

We have focused on the search business in this blog, but there are numerous use cases for AI that have the potential to have a profound effect on our lives, just as the rise of the internet and smartphones have in the past. The addressable market is large and there are several companies across the AI ecosystem that are poised to benefit from this rising tide. But that’s for another time!