What GPT? Its a TaaP!

ChatGPT has overtaken popular imagination in ways unfathomable for even the OpenAI research team. No 2023 trends’ research is complete without a discourse on the transformative potential of generative artificial intelligence. And ChatGPT is only a precursor to many other foundational models waiting in their beta avatars to go mainstream. Foundational models have quietly been ingesting troves of open-sourced and proprietary data, consuming many petaflops of computing power for over five years while the world obsessed over Sophia, an anthropomorphized bot from Hanson Robotics.

The Cambrian explosion we are witnessing now is way beyond the Alphabet vs. Microsoft debate. This is about shifting our priors about artificial intelligence as a species. The heuristics around what makes a technology general purpose need a more inclusive mainstream discussion. No offense to OpenAI, but I don’t think ChatGPT is general enough. Yet.

What is a General Purpose Technology? - TaaP!

Technology-as-a-platform (TaaP) for others to build on. Arguably, the first general-purpose technology we homosapien-sapiens encountered was the fire. It both altered our diet as a species and our history. The latter by giving the sapiens an edge over other species in conflict and survival from the elements. Fast forward to something in our living memory. Personal computing.

Chat - GPT?

Greg Brockman, Chairman of OpenAI, believes ChatGPT is GPT enough, “Our mission is to really build a platform that others are able to build businesses on top of.” However, most of the so-called “AI applications” currently being built-on ChatGPT are a little more than white-labeled versions of ChatGPT itself. OpenAI is offering the ChatGPT API for $0.002 for about 750 words of output, and any company can resell ChatGPT in their own app.

However, the real under-the-hood issue with large language models such as ChatGPT is the sheer computation required. The real winner here is Nvidia, the maker of the powerful GPUs needed to run some of these large language models. Everyone is talking about ChatGPT but rarely does anyone bring up CUDA, the programming model required to run Nvidia chips that can run ChatGPT-based applications. Using ChatGPT in search engines like Bing may eventually be unviable as it increases the cost of providing search results by orders of magnitude.

I am bullish on NotionAI and Snap, though, as they will expand the scope of ChatGPT’s potential use cases and will help onboard the following hundred million users beyond the initial hype curves and FOMO (Fear of missing out). The key is to focus on the CX/ UX as both NotionAI and Snap do, making the interface with AI intuitive, simple, and fun for users.

Another promising application is Tome, with its recent series B funding round led by Lightspeed of $45 million. Tome is an AI-first productivity company that clocked the fastest run to a million users since its beta launch last September. Tome is yet to make its first dollar like most of the 539 generative-ai startups Pete Flint at nfx, a venture capital, lists. These have collectively raised over $11 billion. Impressive arguably, but a fraction of the $75.2 billion in capital raised by fintech companies last year. And that when funding in fintech is down 45% from the 2021 high of almost $140 billion. The conceptual frameworks around foundational generative computation have been around for over four decades. Still, the computation and datasets needed to train such models have only been around for the last five.

Digital property rights in the world of Generative AI

Generative AI models have an attribution problem, generally speaking. Our mental models around search have specific priors wherein we expect the new Bing to be categorical about the source of the search results. Getty recently filed its second lawsuit against Stability AI for copyright infringement. According to the filing, “Stability AI has copied more than 12 million photographs from Getty Images’ collection, along with the associated captions and metadata, without permission from or compensation to Getty Images, as part of its efforts to build a competing business.” Large language models must ingest large datasets to build a competitive moat. And ownership of digital assets and intellectual capital is still evolving, with regulations worldwide playing catch-up.

Hype curves vs. Maturity curves

In 2021, among executives of the world’s 2,000 largest companies (by market capitalization), those who discussed AI on their earnings calls were 40% more likely to see their firms' share prices increase - up from 23% in 2018, according to an analysis by Accenture. However, when it comes to making the most of AI’s full potential and their own investments, most organizations are barely scratching the surface. Only 12% of firms surveyed as part of Accenture’s study have advanced their AI maturity enough to achieve superior growth and business transformation. However, Accenture expects the shift from here to be faster than it took for organizations to embrace “digital transformation.

While most organizations scramble to articulate their real “AI strategy” to their stakeholders both within and without, it is important for us to start with the fundamental premise, “What problem are we solving for?”.

Embedding new technologies in existing business models goes beyond optics and FOMO. It calls for going back to the basics. Keep TaaP-in!