Published 2026-09-25
tags: #understandingLLMs

It seems like everywhere I look there is a new piece of AI news that grabs my attention. On one podcast, there might be a discussion of a new technological advance that brings us closer to a new wave of technological success, while my phone pings with a notification from a trusted news source warning about how AI will kill us all. The constant barrage of conflicting, ever-changing information is enough to make anyone want to ignore AI altogether. Of course, if you want to prepare yourself for an AI-enabled future, you can't just ignore what everyone has to say about AI.

For those of you who want to educate yourself as much as possible about this technology, I have a framework to share with you. It's a simple framework that isn't going to give you everything you need to answer every question you have about AI, but having a way to structure the different ways I can prepare myself for an AI-enabled world has made this daunting task much more manageable. In my view, there are four pillars of AI preparedness, and each one answers a different question:

• AI Literacy: What is AI?
• AI Fluency: How do we use AI well?
• Domain-Specific Expertise: What do we do with AI output?
• AI in Context: What is the true cost of AI?

I think of these as pillars because they hold something up together: the ability to make intentional decisions about AI, whether that means using it or choosing not to. You don't finish one pillar and move on to the next. They share the load, and if one is neglected, the whole structure starts to lean. The good news is that a pillar can always be reinforced, long after the building is up.

AI Literacy

What is AI? Most people equate the Large Language Models (LLMs) that run their chatbots with AI. But it is important to recognize that there are many different kinds of AI, each with their own strengths, weaknesses, and risks. The algorithm that picks your next video, the filter that catches spam in your inbox, and the chatbot that drafts your emails are all "AI," but they work differently and fail differently. Without learning some basic terminology and ideas, "AI" becomes an umbrella term for a technology that is simultaneously powerful enough to take over the world yet dumb enough to not know which light bulb in my house to turn on. Without exposing yourself to the basic mechanics of text/token generation, you won't really understand how an LLM trained on millions of text documents could still make stuff up and get things wrong. Without understanding Retrieval-Augmented Generation (RAG), where the model looks things up in a set of trusted documents before it answers, you won't understand how tech companies have tried to reduce LLM errors.

There are different levels of AI literacy, and I'm not suggesting you need to know every last detail about "training", or "attention," or "embedding" information into hyperspace (although I personally think this is fascinating) to be AI literate. But, knowing a little bit about pre- and post-training has helped me understand what LLMs are capable of and what options we have to implement guardrails into the technology. It's also shown me some of the limits of both the technology and how much (or little) control we have of it. For example, a chatbot's polite, helpful personality is mostly added after training, which explains why companies can build in guardrails and also why people keep finding ways around them. To make sense of future versions of this technology, it is incredibly helpful to have a working understanding of how the technology works.

AI Fluency

How do we use AI well? With a strong foundation in how AI works, we can turn to how you can make AI work best for you. You might think of this as applied AI literacy or as a set of personal best practices when it comes to using AI in your life to actually get things done. Personally, this is one of my favorite areas to dive into because it transforms LLMs from giving me a product to LLMs helping me produce a product. This is the area where we can learn how to use AI to perform tasks for us that enable projects we wouldn't have even considered because we lacked a certain skillset. For example, I used AI to build a python script to analyze screenshots from my newborn daughter's sleep app so I could dive deeper into the data, something I would have never attempted when it meant learning computer vision from scratch. This is where we learn to use AI to make bespoke tools that are too small and too specific to exist as mass-produced apps, but happily live as local projects that fit exactly in our workflows and with our data.

I think it's important to recognize that this type of preparedness should be highly customized to our own personal comfort and/or experience with using AI and other digital tools. Just because someone makes a tool with AI to help with their workflow doesn't mean that's the right solution for me. It also doesn't mean that once we find a best practice for ourselves that it will remain the best as time goes on. Technologies change. People change. The nature of our work changes. And I think that makes AI fluency, well, fluid.

Domain-Specific Expertise

What do we do with AI output? Literacy and fluency are about understanding the tool. But a tool is only as useful as the person holding it. While preparing for an AI-enabled world should include exposure to and practice with AI, I think that it is equally important to continue to develop your own special expertise outside of AI. Depending on what you've learned, what you've read, and what you've practiced, you bring a unique set of thought processes and opinions to any situation. I think this is what Silicon Valley is talking about when they say that the "tastes" of their future hires are what will differentiate them. Your expertise is what allows you to know what to do with AI output. What output is too vague? What output deserves push back? What output is correct (or at least correct enough)? It is becoming increasingly important to not accept text at face value (regardless of whether the text was generated by human or machine), and it becomes easier to maintain a healthy amount of criticism when one has developed their own expertise.

My favorite description of LLMs is as "plausible text" generators. And to the untrained individual, much (if not all) of what comes out of LLMs from outside their expertise is accepted because the words seem to form a plausible statement. But experts can see past plausible text and actually evaluate whether the text is accurate and complete. If you want to stand any chance of resisting being misled (intentionally or not) by plausible-looking text, you'll need to continue developing your skills and training to become an expert in your field.

AI in Context

What is the true cost of AI? The first three pillars are about you and the technology. This last one zooms out. I don't think we can buy into AI responsibly without knowing what we are giving up. And while it might be impossible to actually know the true cost of AI (either because the numbers we need are obfuscated or because we simply cannot predict the future), taking every step we can to educate ourselves about both the positives and negatives of AI is important. How much damage are we really doing to the environment and how does that compare to the current damage we're already doing? Does all use of AI lead to a cognitive decline of its users and, if we are losing skills, how valuable are those skills in comparison to what we might gain? (This is where the pillars lean on each other: if we lose the skills that make up our expertise, we also lose the ability to judge what AI hands us.) Can AI really become powerful enough to destroy humanity and if so, are we really going to let it get to that point? Who do these systems benefit and how much disparity is there between different groups?

I don't think that we currently have everything we need to answer any of those questions (and I think those who can answer those questions confidently are either overconfident and/or trying to sell something). But even so, diving into these questions should be a priority for anyone looking to become a responsible user (or even non-user) of AI. So instead of waiting for certainty, I try to hold these questions open: I look for sources that show their numbers, notice who benefits from a given claim, and revisit my conclusions as new information comes in.

And this pillar isn't only for people who use AI. Even if you decide AI isn't for you, that decision is only as good as your understanding of what you're opting out of, and what the people around you are opting into.

Conclusion

In my own AI journey, these four pillars have been essential for allowing me to make intentional decisions about AI use. I hope they serve as a useful way to organize your thinking about AI, too. As I keep learning in real time, I plan to share resources here that fall under each of these four pillars.