Please, Use AI

Have you ever been stumped about someone’s strident objections to something they have minimal experience with? Lately, I’ve had a few confounding conversations with people who have strong objections to generative AI despite not using it much at all. A gifted writer friend dismissed AI because it gets facts wrong. A new acquaintance scoffed at my use of a chatbot to understand how to properly roast a chicken. To me, who has been using AI regularly for over a year, these reactions seem painfully under-informed.

I want everyone in my network to get smarter about AI through personal experience. That includes you! There’s a lot at stake with this technology, and we’ll all be better off if we can have informed debates about it that go beyond headlines and hearsay. Below, I’ll offer ideas on how to start using and understanding AI.

The AI Learning Curve

I can see why you might not have tried yet. It’s hard to understand what the deal is with AI.

• The hype is intense. There are overblown reactions in both positive and negative directions. Skepticism is a reasonable response.
• The technology is changing at a wildly fast rate. What you heard three months ago might not apply now. This is bothersome and confusing for the majority of folks who have better things to do than follow the play-by-play.
• Terms used in defining and describing AI are fuzzy and will likely stay that way. It’s hard to know what we’re even talking about half the time.

That said, AI is different than other recently-hyped tech.

• Unlike, say, virtual reality or the blockchain, AI is immediately accessible and useful to individuals. It’s impacting most of us whether we know it or not.
• It takes using AI to understand it. Interacting with generative AI directly challenges our mental models of what to expect from information technology. Refreshed mental models are incredibly useful in deciding how to navigate the changes afoot.

I’d like to share suggestions and resources that have been most useful to me in understanding the technology and improving my use of it. I hope this helps you actually use AI and use it enough to get past the primacy effect, where your entire perspective is disproportionately based on your first impression. Our first impressions of AI are hampered by our existing expectations of technology. Changing those expectations, even if you’re ultimately disinclined to keep using AI tools, takes first-hand experience.

You can be smarter about AI. I’m convinced that no matter what you end up thinking, you being smarter means we’re all better off.

What’s below doesn’t touch on the dramatic current events of AI (Who will sue OpenAI next? How high will Nvidia’s stock price go?). It also doesn’t touch upon the many genuine concerns about ethics and future implications (Is AI as biased as we are, or even more so? Will any of us have jobs in the future?). You’ll be well poised to follow all those debates once you’ve spent more time with the technology itself. So let’s dive in!

What To Do

I suggest you do two things concurrently. I get into how in the sections below.

1. Put in the time with AI chatbots to retrain your intuition. We’re so so used to text inputs functioning like search, but it’s not anything like search. We’re also used to computers being accurate. That’s not how this technology works either.
2. Learn the underlying technical principles so you understand why it’s bad at what it’s bad at and good at the rest.

1. Put In the Time

No technical knowledge is required to use AI chatbots. Consider signing up for an account with OpenAI (makers of ChatGPT), Anthropic (makers of Claude), or Google’s Gemini. Try out a variety of tasks rooted in language or reasoning. You can ask it for help and suggestions on how to use it and what to use it for. Approach the task collaboratively. Explore both playful and practical applications. Consider your initial use of AI as a low-stakes, curiosity-driven activity that doesn’t need to produce immediate outcomes.

I’ve used AI successfully for a variety of things: understanding peer-reviewed medical research, creating comprehension quizzes for my students, editing emails that I really wanted to go over well, sense-checking my skincare routine, fixing the tangled mess of a Ruby installation on my computer, figuring out how board game rules apply in specific scenarios, and yes, understanding how to properly roast a chicken. Not all goes well. I have twice messed up my taxes with it! (My wonderful human CPA will not be out of a job anytime soon.)

With time, you’ll break your hangover instincts from tools like search or file retrieval. It’s been popularly said that 10 hours of use will land you in a solid place. Through active engagement, you’re developing an innate sense of how to get the most out of AI tools and by extension many of the technologies that will soon have AI embedded in them.

You could also try out Perplexity for AI which is search. Or you could use Pi for encouraging, meandering conversations.

Helpful tricks

Here are three prompting techniques that seem to help across many situations:

• Reduce the possibility space – The AI starts with the full possibility space that was created by its training. That’s vast! You will get more mileage by reducing that possibility space with roles, goals, and context. Who do you want it to act like when it responds? What are you trying to achieve in this session? What additional information can you share about yourself or the situation you’re working in? (For example, tell it to act like an enthusiastic French chef who loves to help beginners perfect the basics and that you have no special ingredients at your disposal before asking it to coach you through roasting that chicken. For another example, see the ‘general tutor’ prompt linked to below.)
• Work out loud – The AI isn’t “thinking” anything in the background. It’s referencing what’s in the chat so far and its training to predict what would likely be a good response to what’s been said most recently. As such, getting the AI to work “out loud” through a chain of thought is more likely to produce good results than throwing it a big, vague ask. (For example, ask it to review some documents, summarize the main points, and then extract what might be most relevant for a particular audience, in that order, rather than ask it what’s most important in the document off the bat.)
• Provide examples – If you want it to produce a certain kind of output, giving it an example will go a long way. This technique essentially combines the last two suggestions; an example both reduces the possibility space and creates material for the AI to reference as it works.
• Bonus 4th suggestion: Remember that it’s dumb – The chatbots may dazzle you with what they can do. They will also run into walls like lemmings and not know how to back themselves out of a corner. Tell it when it’s wrong, insist that it tries harder, and feel free to scrap a chat and start over again. A common adage with AI is that hard things are easy and easy things are hard. The AI is going to have a hard time with things that may be painfully obvious to you.

The paid models tend to be significantly more powerful than the freely available ones, so consider paying for a month or two as you try it out. (See this post from OpenAI on GPT-4’s improvements over GPT-3.5; it’s the difference between miserably failing the bar exam and passing it with ease.) The $20ish investment is small compared to what you’ll gain by giving the best models an earnest go.

2. Learn About The Tech

Here is a tight selection of my favorite resources. Each of these created a huge unlock in my understanding of AI. They’ll do the same for you!

Generative AI exists because of the transformer, an animated story from the Financial Times

• The creation of a technology called transformers is foundational to the current capabilities of Generative AI. This animated piece is the most elegant, non-technical explanation of them that I’ve found yet.

One Useful Thing, a newsletter by Ethan Mollick

• A Wharton professor who studies and teaches entrepreneurship, Ethan is bullish on AI. Untechnical and action-oriented, his upbeat newsletter is rich with accessible, motivating articles.

Intro to Large Language Models, an hour-long video by Andrej Karpathy

• Karpathy has held influential technical roles at Tesla and OpenAI. He’s a well-respected communicator on these topics. This talk can get a bit technical, but it pays off. It’s worth revisiting as you get progressively more familiar with the tech.

Artificial Intelligence: A Guide for Thinking Humans, a book by Melanie Mitchell

• This book offers an excellent background on key ideas in AI and its evolution. Having worked in AI for decades, Mitchell has her heart in it. She can also be pragmatically jaded. The result is a balanced take that lets the reader think for themselves.
• Reading this book, you’ll learn: why AI has been around so long but only recently got good, why comparing AI to human-level abilities is deeply misleading, and why AI is bad at common sense (and what even common sense is and why we take it for granted).
• The book was released in 2019 before OpenAI’s release of GPT-3 triggered the AI frenzy. While some of the technical limitations described in the book are no longer true, it offers a much more sober look at the field than most things published after GPT-3’s release. I find this refreshing.
• Some passages are a tad technical, but skimming over those parts in no way hindered my ability to understand later parts. You need not be daunted!

Concepts to learn first

In exploring AI, you’ll run into a lot of new concepts. Below are a few that have been helpful for me to understand as I use chatbots. I’m hoping to streamline your self-education by pointing out the handful that have been most helpful as a user.

• context window
• tokens
• hallucination
• training versus fine-tuning
• weights
• retrieval augmented generation (RAG)

I’ll refrain from defining them here as I’m not necessarily qualified to do that. Besides, you could just ask any decent AI to help you with these concepts!

How would you use a chatbot to help you get the gist of these concepts? I recommend using the “general tutor” prompts from Lilach and Ethan Mollick to get the best results. Just start by adding the prompt text to a new chat, and engage in the dialogue that ensues.

Thank you for using AI

Really, thank you. I can’t wait to hear your concerns and criticisms about AI once you’ve spent some time really giving it a go!