This week OpenAI had a fantastic keynote at their Dev Day event. They announced new products and enhancements to ChatGPT. Interestingly, some parts of the internet described the keynote as an "Extinction Event" for many AI startups. Many of the startups reportedly facing extinction share the common trait of being thin wrappers over OpenAI's APIs. They usually offer knowledge management systems built using on some variant of Retrieval-Augmented Generation or RAG (In other words, chat with a PDF or some arbitrary data). While these services certainly have some value they bring to their customers, they have ultimately proven to be fragile products that fell victim to integration with the base platform.
This is not a new phenomenon. In fact, its history repeating itself when compared with the App Store and Apps famously "Sherlocked" by Apple.
Developers have come to accept that, without warning, Apple can make their work obsolete by announcing a new app or feature that uses or incorporates their ideas. Some apps have simply buckled under the pressure, in some cases shutting down. They generally don’t sue Apple because of the difficulty and expense in fighting the tech giant—and the consequences they might face from being dependent on the platform.
Because OpenAI offers a product in ChatGPT and makes GPT APIs available to developers they are competing in the same space they are enabling. Under these circumstances, platform vendors have numerous advantages over 3rd party developers.
- They have access to more APIs and can directly modify the platform for better integration.
- They have access to usage data across all of their 3rd party developers.
- They usually have more money to spend enhancing their products
What app developers learned from their experiences with the App Store included assessing the viability of products entirely dependent on another company's platform and APIs, particularly those that aim to fill gaps in the underlying platform (such as ChatGPT but with PDF functionality). This doesn't account for the real possibility of APIs being directly or indirectly taken away as seen with Twitter and Reddit
What do I do about it?
So what are companies interested in deriving value from AI to do? The solution is to look inward beyond the hype and "Chat with your data" demos. Specifically, assessing existing value streams if you are an already established company or your overall value proposition if you are a startup. LLMs can be a great tool for arbitrary text data transformation and data inference but they are not a product in and of themselves. They have to fit into a more cohesive product vision. If you are sitting on a lot of data, could LLMs be a part of your data pipeline that transforms or makes inferences on the data? In my experience, this was previously an expensive manual data transformation and extraction that can now be automated. This transformed data can be used to either train new models (within the ToS of your LLM) or enhance your product experience.
Effectively using AI should enhance the core value proposition to the user. It should do so in a way that does not compromise the user experience by introducing poor accuracy and nondeterminism where it's unnecessary.
Amazon is a great case study of this. Their customer reviews summarization feature uses Generative AI on data that they would otherwise just be sitting on. It's arguably unintrusive, can be computed ahead of time so users don't have to wait, and is positioned right beside the source so users may dive deeper if they choose to do so.
In conclusion, it's unfortunate that many of those startups were unsuccessful. However, there are many lessons we can learn from their attempt. Looking beyond the hype and looking deeply at your value proposition is key. Build products not features.
What are your top learnings from these events? I'd love to know down in the comments.
New models and developer products announced at DevDay (2023) Openai.com. Available at: https://openai.com/blog/new-models-and-developer-products-announced-at-devday (Accessed: November 11, 2023). ↩︎
Palazzolo, S. (2023) OpenAI’s ‘Extinction Event’ For Other AI Startups, The Information. Available at: https://www.theinformation.com/articles/openais-extinction-event-for-other-ai-startups (Accessed: November 11, 2023). ↩︎
Martineau, K. (2023) What is retrieval-augmented generation?, IBM Research Blog. IBM. Available at: https://research.ibm.com/blog/retrieval-augmented-generation-RAG (Accessed: November 11, 2023). ↩︎
The term "Sherlocked" is a reference to Apple's Sherlock 3 search product integrating features of a 3rd party app known as Watson.
I quickly formed a real company (I had chosen the name Karelia because I had the domain already, for my personal website), got some help in polishing up the User Interface and Website from Robb Beal (who later moved on to create the innovative Spring), and watched Watson take off. It seemed that the sky was the limit, until I was called in for a meeting with Apple's Phil Schiller. I listened to him tell me that Apple was going to announce Sherlock 3, and it was very similar to Watson. I watched a demo of their program: all but one of their modules connected to the same service that Watson did and looked almost the same. I was too stunned to be upset. - Dan Wood
Karelia Software (2006) The long story behind karelia’s new logo, Karelia.com. Available at: http://www.karelia.com/blog/the-long-story-behind-karel.html (Accessed: November 11, 2023). ↩︎
Washington Post (Washington, D.C.: 1974) (2019) “How Apple uses its App Store to copy the best ideas,” 5 September. Available at: https://www.washingtonpost.com/technology/2019/09/05/how-apple-uses-its-app-store-copy-best-ideas/ (Accessed: November 11, 2023). ↩︎
Chowdhury, H. and Kanetkar, R. (2023) “OpenAI just taught little-guy developers an ugly lesson,” Business Insider, 3 November. Available at: https://www.businessinsider.com/chatgpt-wrapper-startup-founders-see-risks-of-openai-2023-11 (Accessed: November 11, 2023). ↩︎
Stokel-Walker, C. (2023) “Twitter’s $42,000-per-month API prices out nearly everyone,” Wired, 10 March. Available at: https://www.wired.com/story/twitter-data-api-prices-out-nearly-everyone/ (Accessed: November 11, 2023). ↩︎
Binder, M. (2023) Twitter and Reddit’s high-priced APIs are bad news for the internet’s future, Mashable. Available at: https://mashable.com/article/social-media-paid-api-internet-future (Accessed: November 11, 2023). ↩︎
New Scientist (1971) (2023) “Fluent answers from AI search engines are more likely to be wrong.” Available at: https://www.newscientist.com/article/2371097-fluent-answers-from-ai-search-engines-are-more-likely-to-be-wrong/ (Accessed: November 11, 2023). ↩︎
Kamps, H. J. (2023) “Build a company, not a feature,” TechCrunch, 17 January. Available at: https://techcrunch.com/2023/01/17/build-a-company-not-a-feature/ (Accessed: November 11, 2023). ↩︎