ChatGPT-4 unveiled: beyond the hype

How can we tell the difference between Chat GPT-3.5 and Chat GPT-4? We might not be able to, and that should be fine. However, the bigger question is how we discern the impact on the various application layers using generative AI models such as ChatGPT.

OpenAI has come a long way since the original paper describing the large language model was published in 2018. ChatGPT-4, much like its predecessors, is trained on vast datasets of text, much of it scraped from the internet, which is mined for statistical patterns. These patterns are then used to predict what word follows another. It’s a relatively simple mechanism to describe. Still, the result is flexible systems that can generate, summarize, and rephrase writing and perform other text-based tasks like translation or generating code.

In its announcement of GPT-4, OpenAI stressed that the system had gone through six months of safety training and that in internal tests, it was “82 percent less likely to respond to requests for disallowed content and 40 percent more likely to produce factual responses than GPT-3.5.”

However, that doesn’t mean the system doesn’t make mistakes or output harmful content. For example, Microsoft revealed that its Bing chatbot has been?powered by GPT-4 all along. Many users were able to?break Bing’s guardrails?in all sorts of creative ways, getting the bot to offer dangerous advice, threaten users, and makeup information. GPT-4 also still lacks knowledge about events “that have occurred after the vast majority of its data cuts off” in September 2021.

Building value-additive application layer on ChatGPT

There has been an upswing of start-ups rushing to leverage LLMs to build out application layers to capture value. Before we dive deeper into some areas, let us look at the generative AI tech stack.

It remains to be seen which of these layers beyond the obvious ones (hardware and cloud) will capture the most value. While proprietary models like ChatGPT are the talk of the town, we might find open-source models built by companies like Anthropic, Cohere, and Character.ai come closer to OpenAI levels of performance, training on similar datasets (i.e. the internet) and with similar model architectures. The example of Stable Diffusion suggests that?if open-source models reach a sufficient level of performance and community support, proprietary alternatives may find it hard to compete. Still, early days for us to draw conclusions, the hype around ChatGPT-4 notwithstanding.

Let us look at some start-ups, from writing assistants to generative art, looking to capture value by building on top of the foundation models.

The first lens we should apply when evaluating these start-ups and their respective moats is whether they are building on proprietary or open-source datasets. While there are obvious first-mover advantages around proprietary data sets, open-source datasets may provide more sustainability beyond the short-term outcomes. The second lens potentially is the machine learning operation itself. Consider Shutterstock as a source of training datasets for DALL-E and a distribution app for DALL-E-generated images. I love this long read by Jay Alammar, wherein he elaborates on how the value accrual at every layer might lead to a competitive moat at the application layer. An essential element is to look at moats beyond those based on technology. We may find companies with established advantages around distribution and access to communities outcompeting those solely reliant on building a technology layer on top of the LLMs.?

Jay presents this diagram as his take on the potential moats for AI-based applications. The one that might be as important as access to community or the captive audience as with say, Snap.

Other early adopters include Stripe, which is?using GPT-4 to scan business websites and deliver a summary to customer support staff. Duolingo?built GPT-4 into a new language learning subscription tier. Morgan Stanley is creating a GPT-4-powered system to retrieve information from company documents and serve it to financial analysts. And Khan Academy is leveraging GPT-4 to build an automated tutor.

While our social media feeds are drowned with new feats from ChatGPT-4 and examples of the model providing unique code or other responses, we need a mental model for discerning how the generative AI space is likely to evolve from here.