Suggestions for Interacting with LLMs
Advice and cautions as you start working with AI.
Now that you have the tool(s) to get started, let’s move into some guiding principles to make the most of them. LLMs are a different type of tool than we’ve had access to in the past, so to get the most out of your interactions with it, you’ll need to alter your approach a bit. A little understanding of what’s happening under the proverbial LLM “hood” will help you work with them effectively, so let’s dive in and take a look.
What are Large Language Models?
LLMs are essentially created by training on vast amount of human writing (just about everything in the public domain, plus plenty that isn’t), breaking it down into chunks of text (words, subwords, or characters) that we call “tokens,” and identifying all sorts of statistical correlations and patterns that emerge. When you interact with an LLM, it’s predicting responses to your input based on these trained patterns, along with an injection of creative freedom (i.e. randomness) that we call “temperature,” which allows it to provide distinct responses each time it’s queried.
An AI engineer would surely find the above too simplistic, but it’s fundamentally accurate. LLMs are essentially an ultra-powerful and sophisticated autocomplete, which is a technology that we’ve all been using for some time via Microsoft Word and within the Google search bar, etc. But these modern LLMs are on another scale entirely from what we’ve seen before.
Now, let’s get one thing out of the way – LLMs are in no way sentient, and we are still a way off from achieving human-defined Artificial General Intelligence (AGI), which would require it to perform any intellectual task a human can. AIs are designed to seem human - and we as humans are prone to anthropomorphism (ascribing human characteristics to something nonhuman) – but keep reminding yourself that they just aren’t. It may seem like a weird thing to reiterate, but it’ll constantly pop in your head as you interact with LLMs (at least if you’re anything like me).
Understanding AI’s Quirks
Due to the nature of their architecture, LLMs have some… peculiarities. Before we get too far along, it’s important to have a general understanding of their risks and limitations. AIs are insanely powerful tools, but they can act a bit odd, and more seriously, they have the potential to both mislead and be misled.
Hallucinations: LLMs can “hallucinate”—or generate false or made-up information. This happens because of how their architect-ed to predict text based on probabilities. At times they will argue for these fabrications fervently. Other times, if your LLM provides accurate information, and you push back and tell it that it’s wrong, it just agrees with you. AIs are fantastic at sounding like authoritative experts, but their responses can be 95% dead on and 5% peppered with plausible sounding errors. Unless you’re also an expert in the area yourself, it’s hard to suss the errors out from the good stuff. Though these flaws are becoming less prevalent as models advance, they’re unlikely to disappear entirely due to their underlying nature. It’s a reminder to always verify important information when working with them. (And besides, what exists in the world that is perfectly accurate anyway?)
Unpredictability: LLMs don’t act like typical software. Unlike traditional software like calculators or spreadsheets, that always produce a consistent result, LLMs are much less deterministic. The same prompt will yield different outputs on different occasions (see “temperature” above) so rephrasing slightly can have a dramatic effect on the output. More puzzling, LLMs can fail to learn a task and then suddenly just “get it,” or even do unexpected things like learn language on their own, and the engineers who designed these models don’t understand how or why. They have all sorts of little oddities (though results vary depending on model and session of course). For instance, they often respond better to (slight) politeness and can even get lazier in December, supposedly because they are trained on less work over the holiday month…?
When I’ve worked with ChatGPT, it’s refused to do something (or say that it’s unable to do so), but once I coach it along a bit (“you can do it!”) or otherwise rephrase my request, it’ll go ahead and complete it without a problem. Just one example this week - ChatGPT refused to take an existing image it drew and incorporate a genie lamp into it. I tried to coach it along (my “you can do it!” approach) but each time it stopped me with some variation of the following:
Well OK. But what if I start a new chat and I describe a new image in one prompt (vs incorporation of the lamp into an existing image)? No problem, apparently.
Why did it refuse to draw a lamp and then draw it without reservation? I have no idea. There’s no other way I can think of to describe how these models act other than a bit weird. And you know what else is weird? People.
Interacting with AI as a person
A suggestion for you moving forward - while LLMs are clearly not sentient, the most practical way to use them is to treat them as people. Highly knowledgeable, weird, slightly unpredictable people. Not a calculator, or a Google search query (though that capability is now built into ChatGPT), or a Microsoft Office application, or whatever. More like a creative partner, a brainstorming buddy, or the smartest weirdest intern you’ve ever worked with.
Why? Because AIs do extremely well at human tasks such as writing, creation, brainstorming, and even mimicking empathy, while they falter on typical machine-like tasks such as complex math and step-by-step logic (though they are improving quickly here). They are a tad weird and unpredictable and lead to surprising results just like working with someone might. As much as I like PowerPoint, it’s just not going to help me brainstorm my presentation (at least not without Copilot built in).
Think of AI as a capable and versatile sidekick, one that can tap into a vast repository of knowledge and human insights, tailoring its responses in remarkably nuanced and interesting ways. It can take your ideas and create compelling content – whether it’s prose, stories, songs, screenplays, and more. The ability to tailor content to your request is simply unmatched. Crafting a wedding toast? Brainstorming a new project? Editing a professional presentation? Inventing a custom bedtime story for your demanding 3-year-old (a favorite use of mine)? AI’s got you covered.
In these ways, interacting with AI is much more like collaborating with a person than pressing buttons on a calculator. It’s a helpful mindset because – just like working with a person – you won’t always get perfect output. You’ll need to refine, review, and sometimes redirect it for your purposes. AI can make mistakes, and you shouldn’t rely on the answers it provides without reviewing and checking its work – the same way you would when working with someone. AI can also surprise you with fresh ideas, different perspectives, and insights that you wouldn’t have considered on your own, just like working with someone might.
I’d recommend keeping this thought experiment front and center as we progress here. You can treat AI like a coworker who helps you brainstorm, a coach who pushes your thinking, or even a therapist who listens intently without judgment. But don’t forget, it’s not actually human. As convincing as it might seem at times, it’s still a tool—just one that’s far better at holding a conversation than any other.
