AI is everywhere right now. If you believe marketing materials, damn near every company and product is “leveraging the power of AI” to incredible results. When we set out to use AI at the foundation of Fordje we expected to get A LOT of eye rolls from colleagues and partners in software engineering.
Fortunately, we are undeterred by eyerolls (just ask my wife!) and despite the collective cynicism toward shiny new cool tech, AI is the right tool for synthesizing and distilling the complexity of building and zoning codes. Beyond the naysayers we have met countless industry experts and practitioners who offer a reasonable challenge: “Can’t I just do this with ChatGPT?”
First, let’s be clear: ChatGPT is great. I’m a paid user and I use it every day. It’s powerful, fast, reliable and solves an impressive range of use cases right out of the box at a no-brainer monthly price point. The speed in which OpenAI is cranking out new models and functionality is very impressive and moves the needle on possibility seemingly weekly.
BUT, as always, you can’t be all things to all people.
Let’s unpack why ChatGPT isn’t the right tool for our building and zoning code solutions.
ChatGPT models are trained on general knowledge, cramming as much variety and volume into the training set as possible. Foundation models also have a “training cut off date” - where models cease to ingest and process information past the the date a model has completed its “training.” Past this cut-off users can ask ChatGPT to read the webpage and use that context to formulate an answer, but the underlying model does not incorporate or retain that information for future use. Users are left with a product with static capabilities that is done learning and improving.
Producing domain-specific answers from ChatGPT requires massive amounts of data to train models. LLMs are limited by the amount of data that can be parsed for each request, technically referred to as the Context Window, and these limits are unique to each model’s design. As is, these models depend on expensive processing resources to generate responses from these gigantic pools of data and are difficult to scale functionally and economically.
However, using an approach called RAG (Retrieval-Augmented Generation) we can enable models with the ability to search for only the most relevant data context needed to fulfill the request. By cutting through the noise, the LLM is able to provide more succinct and accurate answers.
ChatGPT is focused on general knowledge conversational abilities and its training data is structured to meet that interaction type. Conversely, custom solutions can be designed to work with large-scale, structured or semi-structured data sources that align with the intended industry. These custom solutions allow additional search and machine learning techniques to rank, gather and pre-process data for use in the LLM interactions, greatly improving accuracy and performance.
I won’t soap-box about it - we all understand how important security and privacy are when making technology decisions. While ChatGPT does a good job of anonymizing incoming data it simply does not meet the rigor required by industries handling sensitive data. PII, medical data, proprietary intellectual property, corporate and financial records and the like should not be sent through third party services, full stop. Custom LLM implementations allow for a secure, closed loop that manages that risk appropriately. All of the major cloud providers provide a robust interface to host your own models with increased confidence that user data is safe and secure.
At the risk of over-tenderizing the horse, I’ll add that custom solutions offer additional benefits you won’t find in ChatGPT, such as extensive multi-lingual support, structured output formatting, function chaining and tight integration with existing processes and workflows.
Again, ChatGPT is a great tool that does a lot of things for a lot of people. But that’s not us. We are not building a tool that does all things for all people and we’re not using AI for the sake of whiz-bang marketing or a hyped-up company valuation.
We ARE a team of subject matter experts that have 45+ years of combined experience at the crossroads of government, construction and technology. By scaling that experience with the ever-increasing power of AI, we are bringing an intelligent tool that brings clarity and sanity to a problem bogged down with inefficiency and frustration. We’re building a tool that learns, scales and empowers construction professionals to build quickly, build safely and build with quality.