Let’s start with a quick disclaimer. These are just some thoughts in reaction to the conversation around AI and new products sold as game-changers. For the rest of this piece, I’ll refer to LLMs (Large Language Models) rather than AI. Mainly, that’s because I think AI oversells what we’re actually seeing and, to be frank, I can be a bit pedantic.
What is an LLM?
LLMs are machine learning models that utilize algorithms to process language. They’re trained on immense amounts of data and learn language patterns they use to perform tasks. The key bit of information in that sentence is “patterns.” These models are essentially pattern-recognition networks that work together to analyze text and predict outputs. This article from ieee.org has a great quote from robotics researcher and AI expert Rodney Brooks that sums this up nicely.
“What the large language models are good at is saying what an answer should sound like, which is different from what an answer should be.”
What conflicts with the hype around LLMs is the idea that, beyond this ability to recognize patterns and predict, they don’t have the ability for logical inference (yet). This is an important distinction to understand when your business is considering implementing an LLM into a product or day-to-day work. It’s this thought that leads to the next couple of notes it’s worth discussing if your business is looking to begin defining a strategy around using LLMs.
How fast does your business want to get to “average”?
We’ve mentioned that these language models train on vast amounts of information and then use technical mechanisms to focus selectively to identify the most relevant sections for summaries. But can your business be successful doing what’s already out there? Do you really want to use a UI tool that outputs a summarized version of an app experience to differentiate your product? Will your marketing materials perform better if they’re summarized versions of all the other content out there? Do your customers really want another average chatbot? Does that solve the actual problems they may be having with your service?
My point here is that you should have an intentional discussion to understand what value you might gain for your business or your customers when deciding how to use an LLM. If you can’t define this, then you might only be getting to average faster than you were before.
Design by consensus is bad, even if that consensus is set by an LLM.
Designing a solution by consensus doesn’t work. Consensus homogenizes points of view until we think there’s only one viable way to solve a problem. This means we spend all our time trying to validate that course of action instead of establishing experiments to determine which of the competing explanations for what’s happening is correct.
LLMs can positively impact very specific processes when implemented thoughtfully, but you shouldn’t let them choose a direction for your company. Your business needs to understand the intended value you expect to get from implementing an LLM so you can track and test the outcomes. Otherwise, you risk letting an LLM tool come to a consensus on the direction and blindly following a path of diminishing returns.
Don’t confuse performance with competence.
This is a simple one. It’s no secret we’ve seen many examples where LLMs provide incorrect information or just make up something new. It’s important to understand the flaws inherent in the current state of the technology and how that could cause more harm than benefit to your business.
You need more meaningful outcomes, not quicker outputs.
LLMs are powerful tools and can be great value creators for your business. However, your business doesn’t just need quicker outputs. Instead, it needs more meaningful outcomes so your team can understand the direction they should aim. Before jumping right into an LLM implementation, you need to do the strategy work to set a good foundation so that the work you do is geared toward improvement rather than repair.