A few weeks ago, I began exploring a concept called "Enterprise AGI." Interestingly, I started this endeavor without a concrete definition in mind. I simply knew that harnessing the power of GPT-4, deployed as an autonomous software agent, would allow us to achieve automation at scale with applications built using frameworks like LangChain.
This morning, while walking Charlie, I listened to an interview between Sam Altman and Barry Weiss. Sam offered a few definitions of AGI, and one of them resonated with me as the perfect description for what I had in mind. So, without further ado, here it is:
Enterprise AGI is a system capable of performing at least 50% of valuable human work within a business.
I'm thrilled to have this definition because I believe it will make conversations about the future much more straightforward. Unfortunately, discussing AGI or intelligence in general tends to trigger resistance from many people. When I try to broach the subject, I'm often met with immediate intellectual pushback from those who want to debate the nature of intelligence. While these conversations are interesting and significant, they're not my primary concern. I care more about what large language models can do, not who they are. I'd rather not engage in debates about the nature of consciousness or human rights for AI models, and so on.
It's just software. When the head of analytics at a bank or insurance company contacts Prolego seeking help in automating a complex business process, they aren’t interested in discussing the nature of consciousness within the models. They simply want to find a way to automate a complex business process, freeing up their team to work on more important matters.
This definition enables me to speak more specifically about what enterprise AGI is and what it can do. It also grounds the conversation in something both significant and actionable. As I mention in my essay, The Era of Enterprise AGI has begun. We are in the initial stages of this transformation. Proof of concepts for enterprise AGI will launch in 2023, with substantial investments following in the next two to three years. I have no doubt that within five years, we will achieve the 50% automation described in the definition.
Are you ready for it?