The AI space is moving at breakneck speeds. I see first hand in my work the real divide between folks that may have tried ChatGPT a year ago and wrote it off as unreliable (or who just have Copilot through their O365 subscription) and folks connecting agent harnesses to different systems to create real value in their organizations.
I think it is important every organization gives space to an ambitious builder that wants to see these tools blossom into the value drivers we are all being promised they will become. Too often, organizations are plagued with folks in the opposite camp, and for good reason.
Why the skepticism is there
These are some core reasons why people may have a bad taste, ranging from personal experience to more of a macro view:
- They tried it a few times for some basic tasks and it was wrong. Many of us don't give things another chance.
- There is a real fear that it will impact their jobs in a negative way. If the machine can do it better than me, then why am I needed?
- The more hype a topic gets, the more folks seem to reject it. Music plays this out well. A song becomes popular and people start to hate it. It's played out.
- The more macro view: this huge build out is not good for the environment and the world as a whole. Energy consumption is massive and supply constraints are everywhere. This is all leading to super inflationary pressure on a wide range of goods and services.
This is by no means an exhaustive list, but these are the ones I am confronted with most often. And I don't think all of them deserve to be waved away. The fourth in particular is a real tradeoff. Part of building responsibly is being honest that the cost is real, and then being deliberate about where the spend is actually worth it. Which, as it turns out, is most of the argument I want to make.
What's actually working
Because on the other side of the coin, folks that are given the space and freedom combined with some creativity are seeing their ability to add value multiply by orders of magnitude. We are seeing agentic workflows connected to various resources — databases, knowledge graphs, software stacks — giving people the capability to provide rich context to these models in order to guide them in ways only the largest, most well run teams could do in the past (if that).
I think there are some patterns emerging on this front:
- The solution isn't just one product you pay for. It's a combination of products, local infrastructure and controls, and a flexible but well documented change management procedure.
- Build on a code base that has sufficient data access for your goals, and a framework that provides solid structure for reliable agent development. There are a ton of examples here, and they could be combined.
- With the two above defined, determine when deterministic and non-deterministic processes are needed. A good example of deterministic is something like order entry. We need an order to be entered — and all the business rules around it — in a consistent fashion. We don't want to incur token costs for this process. Non-deterministic may be better suited when we are exploring the most effective ideas in demand planning before settling on a plan to write deterministic code for the process. In many cases we are finding non-deterministic processes (aka using the models) precede a deterministic (code) result.
These are just some notes on my current observations. They are changing fast. One aspect I find interesting is knowledge graphs and how we can deploy them in small to medium teams in controlled but effective ways. I like the work Garry Tan is doing with gbrain on this front, and we are seeing a lot more smart builders following suit.
I love exploring these ideas and applying them inside the environments I am working in. If you have any ideas or feedback I'd love to chat. Hit me up at nick@segosolutions.com.