SearchGPT – Non Starter

SearchGPT unlikely to go anywhere.

• OpenAI’s latest attempt to take on Google is probably more about generating revenue and reducing losses but given Bing’s failure to make a dent in Google, it looks doomed to failure.
• Google has an enviable position in search with a 90% market share and an advertising business that generates well over $200bn in annual revenues but Google occupies this space for a reason.
• This is because Google has been categorising the Internet for over 20 years which, combined with its early leadership in AI, has meant that its knowledge graph is far better than anyone else’s.
• This appears in two areas.
  • First, the long tail: where Google’s ability to surface rarely viewed webpages or information is second to none.
  • All of its competitors can do the most commonly searched items but for something off the beaten track, Google continues to outperform.
  • Second, understanding: Humans often do not input their requests clearly or they often make mistakes.
  • Google has an uncanny ability to work out what is being asked for and provide the right answer even if the query was entered unclearly or incorrectly.
• This is what keeps users coming back to Google despite over 18 months of endless hype and excitement about LLMs not one of the myriad solutions available has made a dent in Google’s market share or its revenues.
• Its most recent Q2 2024 results demonstrated a continuation of this trend (see here).
• Furthermore, Microsoft has made a big effort with its already established search engine and now Copilot to use generative AI to take traffic and market share away from Google which has resulted in abject failure.
• The problem as I see it is that the knowledge graph of an LLM is fixed in time while a knowledge graph used for search has to be current all of the time.
• Most LLM providers have chosen to fix this problem with an algorithm called retrieval augmentation generation (RAG) which conducts an internet search to pull in the knowledge that the LLM is missing as a result of being frozen in time.
• These are far from perfect and in most cases are outperformed by a search on any of the search engines available.
• Nvidia has also claimed that a model built for a specific task is 10 times more accurate than an LLM whose knowledge base has been enhanced by RAG.
• OpenAI will be using a combination of its GPT-4 foundation model and a RAG or RAGs and as a result, I am not optimistic that this is going to be a successful product launch.
• The net result will be no impact for Google and very few if any advertising revenues for OpenAI which means that it is still going to have to raise money again fairly soon.
• Hence, if Alphabet’s shares decline substantially when this product launches, there will be a short-term opportunity in Alphabet.
• Without a correction, I remain fairly indifferent to Alphabet although it is one of the cheaper big technology companies.