Finding a good metaphor to explain large language models to normal people is challenging.

A couple years back, a colleague gave me the best one I’ve encountered so far when he said to me, "you know, they're just overconfident toddlers."

From where we stand today, I’d say they are more like overconfident 5 year olds.

Here's why:

They are gullible

Both large language models (LLMs) and five-year-olds lack the ability to critically evaluate information or objectively verify facts.

LLMs rely on patterns in the data that they were trained on in order to generate text. Beneath all the hype, they’re just fancy pattern recognition machines.

They are more capable than you realize

My five-year-old quickly grasps new concepts at an impressive rate. Recent proud moments include throwing and catching a baseball, counting to twenty while fluidly jumping between multiple languages, and perfecting flips on a trampoline.

LLMs are also continually impressing me with how quickly they are able to generate usable code, answer off-the-cuff questions, and solve certain problems faster than I’d be able to do on my own.

They are quick to please

Five-year-olds are evolutionarily dependent on their parents for approval. I’ve noticed my five-year-old looking over his shoulder at me all the time to see what my reaction is to whatever activity he’s doing. (Which reminds me: I should get off my phone and watch him back more often.)

LLMs are also quick to please. They’re designed to respond to your prompt with a similar tone as yours. They also tend to provide affirmative responses, aiming to be cooperative and supportive with the conversation.

They are stubborn

Five-year-olds see the world in a very specific way. If you tell them something which challenges that world view, they’ll double down on their beliefs and, if pushed hard enough, will ultimately shut down and throw a tantrum.

LLMs are similarly stubborn. When I tried using ChatGPT to write code which interacts with the raw GPT-4 APIs, it repeatedly provided the same incorrect response despite my different prompts.

LLMs also have trouble switching contexts, will dig into an initial interpretation of your prompt, and can often ignore any clarifications that might help them get onto the right path.

You gotta verify their output

When I ask my five-year-old how his day was at school, he’ll occasionally launch into diatribe about how nobody wanted to play with him and his teachers didn’t offer him any food whatsoever.

A quick check with his teacher reveals that yes, they did offer him food, he just chose not to eat it. And in fact, there are several kids in the class who enjoy being his friend.

LLMs are not designed to tell you the truth. They are machines designed to give you the most likely response to a prompt.

There’s a subtle difference between “statistically likely” and “factual.”

While the output of LLMs is often remarkably correct, you still need to double check the accuracy of any statements that they may generate.

They do better when you simplify and break down complex tasks into simpler ones

If I walk into a messy living room and tell my five-year-old to clean it up, I’ll be met with an overwhelmed reaction filled with tears and scoffing, and nothing will get cleaned up.

If I walk into a messy living room and present my five-year-old with a silly scenario which involves him parking all of his Paw Patrol characters in the pup tower, he’s infinitely more likely to do it. I can then keep asking him to grab small items and put them away, and soon enough, I have a clean living room.

LLMs similarly perform best when you break down complex tasks into smaller ones.

Let’s say I have a long article about climate change that I want an LLM to summarize. If I just say, “summarize this article about climate change and its economic impacts”, it’ll spit out a fairly decent response, but if it’s a long article, it may miss key points.

Instead, you can break it down into several prompts like this:

Create a summary of this article by following these steps:

1. Identify the main topic
2. Identify key points
3. Summarize each key point
4. Combine the summaries

By guiding the model through a set of more simple steps, you’ll end up with a richer, more complex, and likely better result.

They are great at molding multiple concepts into one

My five-year-old is a very artistic little guy. You can often find him singing songs while playing, taking lyrics from multiple different songs and mashing them up into a fun new song.

One of the best ways to illustrate how great LLMs are at blending multiple concepts into a unique response is by using a prompting technique called Tree of Thoughts.

In Tree of Thoughts, you have the model pretend to be different experts debating the best way to tackle a certain problem.

Each of the experts first explains how they would address the problem.

Then, each expert reviews what the other experts came up with and refines their own position.

After reflecting on each other’s positions, the model summarizes the refined points with a result that is often better than what it could’ve come up with on its own.

Here’s an example: my five-year-old refuses to eat pizza. I consider it one of my greatest failures as a parent. I mean, it’s pizza!

In order to help me figure out a way to get him to try pizza, I gave ChatGPT this prompt:

There are three experts considering the best way to get a 5 year old child to eat a meal that they've never tried before:

1. A child psychologist
2. A kindergarten teacher
3. A grandparent of a child who went on to become a Michelin-star chef

For each expert, give their opinion on how to best get the child to try a new food such as pizza. They will then reflect on each other's viewpoints and reconsider their own perspective. At the end, combine the strengths of each perspective to create a final suggestion.

Walls of text like that are sometimes tough for me to follow, so here’s a flow chart of what’s happening here:

And here’s the result:

That final combined suggestion feels like a great approach. I’ll report back in a few weeks on how the implementation goes. 🍕

One of the best pieces of advice I can give you to get the most out of an LLM like ChatGPT is to talk to it like you would talk to an extremely confident five-year-old.

Explain exactly what you want it to generate. Speak to it using clear and complete instructions. Try asking it to do something in several different ways until it understands what you’re asking for. When it finally gives you a reasonable response, verify its accuracy.

And if it’s acting stubborn and not spitting out what you expect, maybe take a deep breath, gently shut your laptop lid, and take a walk around the block. That usually works for me.

Links Worth Reading

One of the goals of this newsletter is to empower smart people like you to try new ways to get computers to do useful things for you.

This talk from Maggie Appleton gave me a definition for those smart people: barefoot developers.

The term “barefoot developer” is inspired by the “barefoot doctor” concept deployed during the 1960s in China.

Most doctors at that time resided in urban environments, leaving the 80% of the population who lived in rural areas with poor or no healthcare.

In order to get some of their knowledge into the hands of rural Chinese citizens, villagers would go get educated on tasks like administering vaccines and diagnosing simple ailments.

They’d then return to their homes and act as “barefoot doctors”, providing preventative care to their neighbors. They were still barefoot villagers like everyone else, but they were empowered with skills to help those around them.

As a result, life expectancy during that time rose from 35 years to 63 years.

In 2024, we have access to tools which can empower someone comfortable enough to write a basic function in Excel to write code that’ll solve real problems.

You might not be a professional computer user, but you can learn enough about these tools to help you and your neighbor get more out of these machines.

David Smith is one of my favorite developers and a real inspiration to me as someone who has built a platform by which he shares everything he’s learned as a prolific independent developer.

This post reflects on the closing quote from Steve Jobs’ excellent commencement speech: “Stay hungry. Stay foolish.”

Those words hit differently depending on what stage of life you’re in.

When you graduate college, it’s a call to arms. You have nothing to lose, so get out there and try stuff. Be foolish and risky. Build cool stuff with no hesitation.

As you get older and see any sort of success, your charge turns into defending what you’ve built. You become aware of how taking risks can compromise the success you’ve already seen, so you filter your ideas through a lens which protects your fortress.

The key takeaway here is to keep some part of your mindset hungry and foolish. Keep yourself open to new opportunities, and don’t cling too much to the successes you’ve experienced.

In other words: “stay hungry, stay foolish”, and “stay hungry, stay foolish.”

I finished reading this long article during the process of writing this newsletter, and I cannot recommend it more highly.

It’s a gripping story of teenage hubris, the underground hacking scene, how shockingly simple it is to take down major parts of our internet infrastructure, and how the authors of the Mirai exploit redeemed themselves.

I’m trying to get back into reading once again, but I can’t seem to find pieces that keep me wanting to come back consistently. (It doesn’t help that Dropout releases incredible new TV shows daily.)

It took me about four nights to read this whole piece, and for the first time since probably high school, I found myself excited to launch Instapaper and pick the story back up.

And there we have it, issue 3 in the books!

I really appreciated all of you who reached out and told me what you thought of the last issue. I got a lot of great feedback about formatting the text better so it’s easier to skim, and also a lot of “this is good, keep going!” which is very encouraging to me.

For this issue’s call to action, I’d love to hear from you what you’d like to hear me ramble more about. I’ve focused mostly on AI so far, but there are so many aspects of technology I could venture into. What sorts of tech topics are you curious about? Is there any other tech topic you’d like me to cover, or should I keep sharing more of my thoughts about large language models and artificial intelligence?

Reply to this issue and let me know!

Thanks again for reading, and we’ll see you in a couple weeks with another issue of 🐵🔧!