GPT4 and Why You Should Really Care

Large Language Model (LLM) tech is advancing rapidly, and a lot of games are starting to change as a result

One might say that I was relatively early to the ChatGPT hype train, calling out GPT3 and why you should care about it back in December:

What is GPT3 and Why You Should Care

And I’ll be honest, February came around and I was worried that I had overhyped it. I had stopped using ChatGPT by then. Perhaps it was just a really cool novelty, but a novelty nonetheless.

My conclusion there was what I particularly worried about having overstated:

Is it hyperbolic to say that ChatGPT is a step change in the same way the Internet, Social Media, and the Smartphone were? Maybe. But it sure looks like a true step change that will kick off another acceleration of innovation. Go ahead and give it a try and see what you think!

While I stop short of confirming it as a step change, suggesting it was such a think felt wrong as we got into February and early March. My life hadn’t changed, so what’s the big deal? But then…

GPT 4 Arrives on the Scene

The future of work? This image was generated with AI (Midjourney).

OpenAI released GPT4 to their paid subscribers on March 14, 2023. I couldn’t hold out too long. It took about a day or two of me seeing some of the really cool examples being shared on Twitter to buckle and pay the $20/mo for premium access to GPT4.

I’ve used GPT4 almost every day since, and try to explain it’s capabilities to most people I come across.

Unlike GPT3 however, GPT4 is making a quantifiable impact on my life. Here’s one measure of that… my Github contributions so far this year:

Since I’ve had access to GPT4, I’ve been pushing code 5-6 days a week.

Coding, as we’ll get into, is certainly one area where GPT4 shines, but it is far from the only one. My wife and I have now outsourced our meal prep planning to GPT4, with great success. I’ve used it as a language tutor, had it generate lesson plans for kids educational content, use it as a career advisor, generate travel itineraries, and more. As with all great tools, it is an accelerant to human progress when used correctly.

But back to that original statement from a few months ago: is it hyperbolic to put ChatGPT, or rather GPT4 and similar LLM technology, in the same sentence as smart phones, social media, and the internet? I don’t think it is, even though we are clearly still at an early stage of innovation.

I covered some of this on my main blog from a non-business lens and looking at the broader philosophical and cultural implications that might arise, so take a look at that to get some early reaction to GPT4:

Everything Just Changed

But as it’s been a month, let’s move into some examples of how I’ve used it, which will hopefully illuminate it’s use.

Some Examples

It is very tough for me to pick out singular examples that demonstrate the capabilities of GPT4. To give some sense, I have about 30-40 different threads of conversation with GPT4 on various topics. Some of those, like the ones around building web apps, are over 100 pages of text when copied into Word.

GPT4 and Code

But we’ll try! So let’s start with the prompt that I gave it as I was looking to recreate WeAchieve - a habit tracking platform that my good friend Kai and I worked on in 2018 and 2019:

Obviously, this is a fairly detailed prompt. But it’s written in plain English, with a little bit of necessary tech jargon. Here is GPT4’s response:

From there, I was able to prompt it to draft or create specific components, functions, et cetera. While not perfect, it was relatively easy to debug. Here are some of my subsequent prompts:

I’m not showing the answers to these, but the responses are the same: GPT4 provided code or an answer that was sufficient and thus implemented. I’d say it was a rough 50/50 as to whether the code could just be copied in wholesale or if it took some manipulations after the fact. Now, each of these are things that I could have figured out on my own with enough time. But, is it easier for me to refactor the code, ask GPT to do it? Obviously it’s the latter, which frees me up to continue to plan further ahead.

As I said, it’s not perfect, and creating something the right way still takes human guidance. Here is one telling example:

Basically, I’m asking it to have a modal pop up that I may trigger from multiple places in the app. The solution it proposes here would work, but isn’t logical or efficient: basically adding a parameter to the modal so that the modal, while still stored with the button that calls it, can be called from elsewhere. So, I challenged it:

The small-but-significant nuance here is that GPT4 not only acknowledged that my suggestion was better, but it knew why that suggestion was better “this makes the modal more reusable and easier to integrate into different parts of the app.” That may not seem like much, but trust me, that’s huge. Heretofore, no computer program that I am aware of has ever possessed such understanding.

And so this whole chat goes on for, 101 pages of content, which I will spare you from. What I can share is the end result, which is now up at www.mumma.co/weachieve (use password: “expound” to enter the site).

While it only stores data locally for now, it’s very usable and, if I may say so, a great habit tracker for ~3-5 days worth of coding that would have taken at least a month without GPT4:

You can read more about WeAchieve Lite on my Human Optimization Project blog:

WeAchieve has Returned!

GPT4 and Learning

More broad than coding, GPT4 is revolutionary for education and learning. In the below example, I asked it to be my Spanish tutor, and to give me new lessons when I ask for them, then evaluate my response, explain any mistakes I made, and give me a new lesson adapting to my skill level.

Here was my response to the first exercise:

And then here was my response to it’s second prompt:

Now, while computer translation has become somewhat ubiquitous, the level of comprehension and clarity here is incredible, especially when considering that this knowledge is generalized in GPT. It can explain and handle any concept in this way. We could do the same thing with math, science, anything.

Let’s say you want to relearn Chemistry and Biology for some reason. You can just ask it to teach you:

And here’s the plan for Chemistry:

And the first lesson:

And note, anywhere along here, you can ask GPT to explain something to you like a five-year-old, or ask for an analogy from [topic you know more about]. The options are endless!

As I was going through this very topic actually, a hilarious moment just happened. Even after the second lesson, it didn’t include a quiz. So, I just thought I’d ask about it, and it gave exactly the answer you’d expect from the frazzled high school teacher or absentminded college professor you can all recall:

I often describe to people that GPT4 is “bizarrely human” in some of its responses, and this is a good example. It simply forgot the quiz, as a human would, and covered that up in the best way that it could.

GPT4 and Advice

One other thing I’d recommend doing is asking ChatGPT to act as a career advisor and then just pasting in your resume. Specifically, you can suggest that it “ask any questions that may be relevant before providing advice.” Here is the response that I got:

Naturally, I answered these questions and then explored several of the specific points it mentioned, asking for more detail and further fleshing out ideas where needed.

What Comes Next?

Since GPT4 only came out a month ago, a lot of news has been made in the AI space. Google launched their AI, Bard, and several open source alternatives have been launched, including Dolly by Databricks! People have been experimenting with a wrapper called AutoGPT that basically allows GPT4 to create tasks and delegate those tasks to child instances of itself, which ultimately gives these tools a degree of agency. Plug-ins are coming which will allow GPT4 to directly interact with the internet and various tools around it. Longer-term, robotics companies are already looking at how to synergize LLMs with a physical robot and advanced vision algorithms.

We’ll cover each of these topics in a bit more detail to come, as there simply is too much to cover in this (already lengthy) piece.

Even still, even today with the capabilities that $20/month can get you, the nature of work has already changed tremendously. And if you are simply using it to summarize or draft e-mails, you are missing 99.99% of the value that GPT4 can bring with its knowledge and communication abilities. A lack of knowledge or expertise doesn’t have to stop you (or anyone) any more. All you have to do is know how to ask the right questions.

If there are topics in AI or use cases with GPT you’d like to hear more about, please let me know in the comments! I’ll be covering many of these in more detail, but would love to hear what is of most interest to this audience.