Thinking through the Bull Case for OpenAI’s threat to Google’s dominance

ChatGPT’s interface is to say the least, bold. A chat interface is unbounded in the requests you can make, and the nature of its singular answer delivered with an “English professor’s writing skills [and] an encyclopedia’s knowledge base”, can leave you in the trance of believing its omniscience.

Of course, there are some categories of queries for which ChatGPT is magic, and many where it is plagued by hallucinations or inaccuracies. It’s easy then to wave off any speculation on OpenAI’s threat to Google as blinded by hype. But doing so would be short-sighted. What could the bull case be? I can’t help but think about the latest battle with a seemingly dominant incumbent: food delivery.

With the benefit of hindsight, GrubHub didn’t have a chance when DoorDash et al came onto the scene almost ten years after its founding. By that point, GrubHub, long past the days of its startup swagger, had clung on to a vision of food delivery that constrained its results to restaurants that were able to do their own delivery. In a way they had no choice — as a public company, the measure of GrubHub’s success every quarter was how much cash they generated. When DoorDash et al came onto the scene, they dramatically increased the selection of restaurants to consumers by providing delivery themselves (therefore redefining the atomic unit to a larger superset), and thanks to the benefit of being a private company in a very generous funding environment, could ignore cash and worship instead at the altar of growth.

Back to Google and OpenAI.

Google is to say the least, long in the tooth. The competitive threats that ended up eclipsing GrubHub were founded almost a decade later. Google meanwhile was founded almost 25 years ago!

I love Can Duruk’s description of what’s happened to search results in the meantime. If Google became the portal to the web, humans molded the web to Google. And with our wily ways of Search Engine Optimization and Google’s every increasing need to generate more cash, the quality of the experience decayed.

Enter Large Language Models (LLM) like ChatGPT.

Much like DoorDash, ChatGPT dramatically expands the atomic unit of supply. Whereas Google is constrained to serving results of content already written by a human, ChatGPT doesn’t have that limitation. Its model creates content on demand. Moving forward, it’s easy to imagine a world where ChatGPT or another LLM-driven interface could even include links to other webpages as results, thus forming a superset of LLM + human generated results.

And much like DoorDash, OpenAI has been able to access an absolute firehose of private investment, and is seemingly unfazed by the cost of every query. If DoorDash was chasing scale to tip a market, OpenAI is chasing scale to tune its models to AGI. No cost is spared to achieve the ultimate goal. As many have written, the cost would add up very quickly across Google’s billions of queries per day. Clearly Google wouldn’t use GPT-3-like technology across every query from the beginning, but even across the queries for which you’d think it’d make sense, Google’s tremendous scale still works against them.

So is OpenAI’s or even Bing’s trouncing of Google a done deal? Of course not. First, while OpenAI seems to have no problem accessing private capital, the demand and opportunity is so large here that you’ve just got to imagine OpenAI won’t be able to subsidize the costs forever. They’ve obviously just made the first move of having a freemium tier. Second, I can’t help but think about is the double-edged sword of OpenAI’s ambition, which I believe creates an opportunity for a LLM-driven startup.

I’ve written a lot about DoorDash’s strategy to grow and the importance of finding a white hot center. For DoorDash, that was the suburbs, where competition was weak (if not nonexistent!). Despite being subscale, DoorDash could pick and choose its battles so it could be in a position of strength as it expanded.

OpenAI has taken a different approach with ChatGPT. As I mentioned earlier, ChatGPT’s interface is unbounded in the requests you can make — literally, it lets a person use language to construct any query you can type.

To me, this is a bit like DoorDash launching across all US cities and categories of goods, all at the same time, and when you order something, does its best to deliver it. It shouldn’t be surprising then that while ChatGPT can feel like magic for some results, it gets a lot of queries wrong for a myriad of reasons. You can see what’s possible and seemingly eventual, but the interface doesn’t guide the user to what’s best now.

All this makes me wonder whether there is an opportunity for a new chat interface that doesn’t try to boil the ocean in the way ChatGPT does, but instead picks off a vertical or modality where this interface is strongest? I’d imagine you’d also need some kind of human-driven feedback loop, must like Google benefited from with its search interface, to train the results over time using human engagement data.

As always, curious what you think (and if anyone I should be meeting!).