Before writing this article, I entered the search term “generative AI” into Google, my search engine of choice. Quickly, links to 77,200,000 articles were listed. The available access to OpenAI’s ChatGPT app has received much recent press and musings from the Twitter, LinkedIn, and blog communities (I’ve written four articles about ChatGPT since January). However, there was only a single article about ChatGPT that surfaced on the first page of my “generative AI” search.
While ChatGPT may be receiving the bulk of the press about generative AI products, it’s not the only application that has been developed or is in development in the large language models (LLM) generative AI field (Stephen Wolfram’s blog has the best technical explanation of how ChatGPT works for a non-AI practitioner).
A recent Economist article titled The race of the AI labs heats up describes the initiatives in the race for AI supremacy. There are a couple of charts from this article that demonstrate how large tech companies have an advantage in this race. Chart 1 below shows the computing power being used to train AI systems by year and by creator. OpenAI’s GPT-3, Google’s LaMDA, and Amazon’s Alexa are at the top in terms of the consumption of computing power to power their apps.
Large corporate research departments have dominated the development of AI systems over the past seven years, but research consortiums are rapidly gaining ground (see chart 2 below).
During that same period, AI-related research papers presented at major AI conferences were led by many of the same corporate research teams (see chart 3 below).
According to The Economist, Amazon and Meta respectively produced two-thirds and four-fifths as much AI research as Stanford University. Alphabet (Google’s parent company) and Microsoft produce more research than Stanford, and that doesn’t include research from DeepMind, another Alphabet subsidiary, and the sister lab to Google Research.
When it comes to the large language model (LLM) field, The Economist maintains that the big battle is between Microsoft and Google. With the help of an engineer from Google, The Economist tested ChatGPT3.5 versus LaMDA. Neither AI emerged as clearly superior. Google’s was slightly better at solving math problems, and ChatGPT answered more SAT questions correctly than Google.
There is a reason why these AI models do not significantly outperform each other. The Economist maintains it is because the AI researchers all “hang out with each other” and those recruited to corporations from academia made their transition conditional on the ability to continue to publish their research and present their results at conferences. Meta’s top AI expert estimated that no one is ahead of anyone else by more than two to six months.
At the same time, there’s an early mover advantage for apps like OpenAI’s ChatGPT (with 100 million users a month) generating tons of data that makes its models better. Stephen Wolfram (cited earlier in the technical article) writes that size matters, and the ChatGPT model utilizes 195 billion weights in its LLM neural network model. Wolfram also notes that the boundaries of the AI dataset get a little fuzzier than the mainstream data, so I imagine that as the datasets continue to ingest data from the internet and other sources, there is a size for LLMs datasets where continued data ingestion doesn’t make a notable difference in performance.
Nonetheless, the money continues to flow into generative AI startups. According to The Economist, there were 110 deals for generative AI startups that received $2.7 billion in investments just in 2022.
In a subsequent article published in the February 11, 2023 Economist, the headline Is Google’s 20-year dominance of search in peril? makes it sound like ChatGPT could make a difference in search. According to The Economist, every second of every day, Google processes 100,000 web searches. The accuracy of its search algorithms keeps users coming back and reduces competitors. All other search competitors to Google barely account for 10 percent of all daily internet searches.
ChatGPT is the first visible threat to Google’s search dominance. The $10 billion investment from Microsoft that allows it to license ChatGPT for its Bing search engine is viewed as a strategic advantage to increasing its market share of search traffic because it provides a fluently written answer instead of the list of links provided by Google. Microsoft has an advantage on other search providers (other than Google) in that it already has infrastructure that includes unlimited computing power, storage systems for large datasets, and many web-crawling programs that constantly scrape information from the internet. According to the Competition Markets Authority, the British trustbusting agency, it would take between $10 billion and $30 billion to set this type of infrastructure up from scratch. Microsoft’s Bing currently owns a 5 percent market share of search. Senior management at Microsoft estimate that for every 1 percent increase in search market share, its advertising revenue will increase by $2 billion. If they could increase their market share to 15 percent, the $10 billion investment in ChatGPT would be worth it.
Google’s dominant position will take a while to erode. It’s the default search engine in Chrome which is used by two out of three people on the internet, and it’s also the standard search app on more than 95 percent of smartphones sold in the U.S. At the same time, Google’s size and market share pose potential regulatory issues for its actions that could increase its market share at the expense of its much smaller search competitors.
Those of us in academia have been spoiled by Google’s search format. Google’s co-founders were doctoral students at Stanford and their philosophy of search in terms of listing links and ranking them based on sophisticated algorithms stemmed from their desire to build a product that would produce a more relevant search outcome than the limited databases in academic libraries. While Google has increased its search sophistication, it’s never moved from its list display. Because of the smoothly written responses to queries, ChatGPT and other LLMs have the potential to be the preferred search engine unless Google incorporates its LaMDA product quickly. Regardless of your preference or prediction of who the financial winners are for this type of enhanced search, the real winners are the users who may not have the ability to interpret the search results as well as highly educated users. This family refrain from the Indy 500 racetrack, “Gentlemen, start your engines!”, may apply to the enhanced search engine capabilities of the future.