SearchGPT is a PR stunt

OpenAI made an announcement about its latest temporary test, the SearchGPT Prototype, and, as expected, the press obliged.

- The Verge - OpenAI announces SearchGPT, its AI-powered search engine
- CNN - OpenAI is taking on Google with a new artificial intelligence search engine
- Yahoo Finance - OpenAI announces SearchGPT in challenge to Google's search dominance

Somehow, the press missed the two keywords — 'temporary' and 'test' — and instead focused on the word 'search.' That was enough to declare that we now have 'an AI-powered search engine with real-time access to information across the internet.' It was also labeled as 'a meaningful threat to Google.'

You can feel the excitement in the air. Finally, we have something ready to take on Google. Sadly, that’s not the case. Not even close.

A search engine, in very simple terms, consists of three different systems. The first part is the crawler, which finds pages and downloads the content. Then you have the indexer, which stores the content in the database and appends extra information about every stored page. Finally, you have the retrieval system, which, based on your request/question, identifies the most suitable pages and returns to it as quickly as possible. For the curious ones among my readers, here is an explainer video on how to build your own search engine.

What Google does is try to understand the context and the intent of your request. Among other things, it also corrects for spelling mistakes, wrong word order and much more. All of that in sub-second speed. All of that without adding the word AI in the mix. On top of that, it is also serving the ads that make its billions.

Now, let's see what OpenAI is doing. First, it crawled the internet and scraped content like any other search engine. Then it had to train its model. After hundreds of millions of dollars — or maybe a few billions — it produced ChatGPT, which, to the astonishment of its audience, can return paragraphs of text to a simple question. The information returned is up to date to the moment before the training began. Interestingly, one of the weaknesses of ChatGPT’s technology is that it is very difficult for it to forget. Without retraining the whole model, there is no reliable way to do that.

With Google, you have to find the right sets of words to find the content which most closely matches your question. With OpenAI — which gave birth to a new profession, the prompt engineer — you have to provide detailed instructions on what the returned content should look like. While Google provides a set of different answers for you to further refine the context and intent, ChatGPT confidently returns anything which looks remotely plausible.

In very simple terms, that’s the difference between Google, a search engine, and ChatGPT, which is not a search engine.

I hear the arguments — it is a test, a prototype. OpenAI has deals with content publishers and will be able to get the latest information in real time.

Correct. It is a test and a prototype, but in order to take advantage of information in real time, it has to build its own search engine. The current technology won't support it.

Let me illustrate with a simple example.

I wanted to find information about Veksø Helmets while writing an article about the Bronze Age. While searching for information about the Bronze Age, I came across a reference to the helmets found in Denmark, which were created between 857 and 907 BC. I also initially got the spelling wrong — 'Visko Helmet.' When entering these two words into Google, it a) corrected my spelling, b) provided a link to Wikipedia and c) in case that Bronze Age was not what I was looking for, it offered me links to a company making safety helmets.

Same task submitted to ChatGPT.

Simple question: 'Visko Helmet' resulted in a simple answer 'Visko Helmet appears to be a specific term or product, but there isn't enough context to understand exactly what it refers to. Can you provide more details or specify what you are looking for regarding Visko Helmet? This could help in giving a more accurate and helpful response.' Right — I have to be more specific.

Enhanced question: 'what is a Visko Helmet' resulted in this answer: 'The Visko Helmet is a product from the brand Visko Tools, a well-known manufacturer of a variety of tools and safety equipment. The helmet is typically used for protection in industrial, construction, and other labor-intensive environments. Key features of the Visko Helmet might include:...' actually there was a list of useful features of this safety helmet.

ChatGPT knew about a few possible answers, but chose one without verifying both my intent and context.

Okay, I can hear the objections. Yes, you are correct. That was ChatGPT and not SearchGPT.

Actually I do have access to SearchGPT and asked the same question — 'what is a Visko Helmet' and got the following answer: 'The Viksø helmets, also known as the Visko helmets, are Bronze Age horned helmets discovered in 1942 in Viksø, Denmark. These helmets are made from bronze and are notable for their horned design, which predates Viking-era helmets by about 2,000 years. They are adorned with decorative elements such as bosses, eyes, and beaks, and are thought to have been used in religious or ceremonial contexts rather than for combat.'

You must be surprised how it is possible that this new temporary test of a prototype works so well. Was I so wrong about its capabilities? The one significant feature which I forgot to mention is that SearchGPT not only provides answers but also provides links to the website where it found these answers. I was expecting Wikipedia or other Norway National Museum. Nope, it was bing.com, the search engine, which, despite billions spent by Microsoft, still hasn’t cracked the double-digits in market share.

That's the current state of SearchGPT. It’s a wrapper for Microsoft's search engine. There is nothing wrong with that approach. It is an option to see if you can somehow combine the ChatGPT technology with a 'traditional' search engine and see if that will provide better results.

Just don't declare this test an AI search engine or a threat to Google. OpenAI is still a very young startup — true, with billions poured in by its investors — and it is still trying to create product(s) for production use. But, don't forget, there are others trying to do something similar.

It is fun to cheer for the underdog getting into a fight with the 800-pound gorilla, but, before you do, think about the last 25 years and what Google built (and Microsoft couldn't). We are a long way from an AI search engine (whatever that might be), and, just because you put AI at the beginning, it doesn't make it so. OpenAI is as distracted as ever. It keeps adding new things without a clear path forward.

The recurrent pattern? OpenAI’s strategy is as foggy as it gets.