One tool to rule them all?

I was introduced to GPT4 through my job in customer support. In the beginning I used it mostly for writing replies to customers and to satisfy random curiosities.

Then one day, doing my taxes, I asked ChatGPT to suggest a website that would allow me to fetch historical exchange rates for a number of dates – it didn’t know of any free website that would do that but offered to write me a python program to get it from an API.

Yes, why not?

I tested it, and it worked. I was floored!

It’s in the air

A lot seems to be happening in the AI space. Every day brings new Twitter / X posts of people doing exceptionally cool things. Due to how incentives are structured in the entrepreneurship space everyone wants to make a product with a ‘moat’ that can serve a sufficiently large user-base to be a successful business, this translates to off-the-shelf products that can do one thing well, or a few things well – think the Rabbit r1. Yet everything coming out seems to be positioned in the world of packaged products designed to work for a large number of people.

I contend there is a significant untapped potential in the custom solution space. Hyper-specific solutions leveraging AI within workflows to automate decision nodes, and do things like transcription, writing, proof-reading, and audio generation. I have found it ridiculously easy to create, with just a little effort, custom solutions that work almost perfectly for my use-cases. I don’t know how soon a commercially available product could come on the scene and make a generic enough solution that swiss army knives a large number of use-cases – think Notion for content management.

Then there is the problem of making any generic solution work with organizational data. This is addressed somewhat by document uploads, which are either processed as embeddings, or become part of somewhat more sophisticated RAG systems.

Generalized AI tools are trained on broad data sets, which means they often miss the nuances and specific needs of individual organizations. This can lead to a mismatch between what the AI tool offers and what an organization actually needs.

The Promise of a ‘Teach and Do’ Model

Addressing this challenge, there’s speculation about a transformative approach in AI – a ‘teach and do’ model. This concept, akin to the yet-to-be-launched ‘rabbit AI’s teach mode’, envisions an AI tool that organizations can teach their unique processes and let it handle tasks autonomously. This model, if realized, could be a game-changer, eliminating the need for custom AI solutions by offering a platform that is both adaptable and intuitive.

The Current Reality: Tailor-Made AI Agents

While the ‘teach and do’ model remains on the horizon, the current reality is that significant gains for organizations lie in tailor-made AI agents. The beauty of these bespoke solutions lies in their customization. Unlike off-the-shelf products, these AI agents can be trained on specific organizational data, ensuring that they align perfectly with the unique needs and workflows of a business.

The approachability and feasibility of these custom solutions have improved dramatically with advancements in AI technology. Organizations can now collaborate with AI developers to create solutions that not just automate tasks, but also bring a level of insight and efficiency that is finely tuned to their operational needs.

The Balance Between Universality and Specificity

The future of AI in the enterprise space is a delicate balance between the universality of off-the-shelf products and the specificity of custom solutions. As AI continues to evolve, it’s likely that we’ll see more sophisticated off-the-shelf tools that offer greater customization options. However, until these tools can truly adapt to the intricate and varied requirements of different organizations, tailor-made AI agents will remain the more effective choice.