An Interface To Necessary Complexity

It strikes me that Large Language Models (LLMs) are a great interface between people and necessary complexity.

What do I mean? Here’s an example.

Many government programs that are meant to provide help to people in need don’t end up helping very many people. Matthew Desmond describes this reality in heartbreaking detail in his book Poverty, By America. A lot of the waste is a reflection of the complexity of these aid programs and their implementation, which makes it difficult for people to figure out how to get aid they’re entitled to receive.

I suspect that these programs are complex because people are complex, and in order to implement them we need to abstract our judgment of who should get what kind of help, when, and in what form — at scale. Right now that judgment at scale is based on static rules! This is where modern AI (a.k.a. LLMs) could be useful.

Imagine a program like the Supplemental Nutrition Assistance Program (colloquially known as food stamps) where instead of having a generic list of approved vendors and food items that the recipient is allowed to purchase, every transaction was run through an LLM and given the full context on that person’s individual case — are they sick? Diabetic? Celiac? Is this a brand new local store or farmer’s stand? Is it their birthday? — and the transaction was either approved or denied based on individualized judgment at scale as provided by the LLM. Maybe a birthday cake on your birthday is OK.

Imagine further that there was a chat interface you could talk to in order to find out if you were eligible for help in the first place, and could make applying for it straightforward. What if that chat bot were infinitely patient and absurdly well informed, like Mr. Incredible in The Incredibles when he’s working his day job at an insurance company.

This is the kind of messy, necessary complexity that LLMs are well suited to digest and interpret in straightforward conversational ways. It will always be complex to design and administer government aid programs, but it doesn’t always have to be complex for us as individuals to interact with these programs.

What other places do we have necessary complexity, where LLMs might be a helpful interface for us?

• Taxes and insurance claims — this one has already occurred to many people, and you might even argue that the complexity is unnecessary, but the tax code is currently complex and an LLM will make a great parser and guide, much like the idea above about navigating government aid programs.

• Connected home devices, especially in the kitchen. Your blender, your pressure cooker, your oven, your air fryer are likely all from different brands. They all need to be operated differently depending on how many people you’re cooking for and what the recipe is. The complexity here calls for a natural simplifying interface. (Disclosure: We at Alsop Louie have an investment in a company called Fresco which is using AI to make smart appliance cooking easy and reliable in the home kitchen). 

• International regulatory landscapes. Countries tend to have their own rules for what’s allowed in their products based on their own values, and those differences are unlikely to change. Each populace should and will continue to have different ideas about everything from privacy (hello, EU!) to food safety (hello, Argentina!), so this will remain a complex landscape. Having an LLM as the scalable interface to that complexity would be a great win.

• General contractor work. The coordination of different construction specialties, schedules, and materials required for home renovations or commercial construction projects. Checking those projects for compliance with laws and regulations adds an extra layer of necessary complexity, and an LLM could serve as an excellent interface for complicated construction projects.