The AI Revolution: Two Years In

Two years ago, creating things using AI was largely the domain of highly paid machine learning engineers working at big tech companies. Now, it's in everyone's hands. The AI revolution has reshaped our world at breakneck speed. Let's take a step back and reflect upon how far we've come since I started writing about this topic in August 2022 when I asked OpenAI’s GPT-3 what technology should freak humans out the most.

That interaction was, in part, what inspired me to write my first article on what was clearly evolving into an exciting new area: Generative AI. (Although no one called it Generative AI then, I referred to it as “no-code” AI.)

It has been two years and fifty articles since I wrote that first one. A whole lot has changed in the world of AI since then. ChatGPT launched and became the fastest-growing consumer app in history. AI has moved to the top of almost every business leader and regulator’s mind. And NVIDIA’s stock price is up about 24x in the last 24 months.

So, how have the products improved over this time? As I did last year, I want to use this anniversary as a way for us to step back and appreciate the rapid change we are seeing in the field. In a world where we see constant change and improvements, it is hard to appreciate the progress that has been made. Or as Ferris Bueller put it:

Life moves pretty fast. If you don’t stop and look around once in a while, you could miss it.

AI Creating Images

Take this image as an example. I have used the same prompt each year I have written this article. In 2002, this is what DALL-E returned with the prompt: “a small group of people panicking out about what they see on a computer screen, painting by leonardo da vinci”

Back in August 2022, it was AMAZING to type in text and have it return something that resembles what I typed! If you want to go back into the AI time capsule, listen to the excitement from John Oliver and his audience from that summer as he showed them what image creation tools could create.

?

Fast forward one year, and putting that same Da vinci prompt into Midjourney returned this:

While not perfect, it is a huge leap forward from the prior year. AI artwork went from a sideshow to a tool that is more frequently used in designer workflows. Even the hands at least seem to have five fingers. Today, I can go a step further. I put that same prompt into Midjourney to generate an image; then, I can have Runway ML do a decent job animating it.

We’ve come a long way in 24 months. By this time next year, we’ll probably have full-on text-to-video conversion. (We’re still waiting for you, Sora!)? ?So, in August 2025, expect a 15-second video, with sound, of Renaissance-area men panning left to right and each man uniquely reacting in shock to what they see on a modern computer monitor.

AI Creating Books

I started another experiment a few months after the original article came out. I wanted to see how well an AI could write a book and track its progress over time. I took on the pen name Alex Irons and started publishing these books on Amazon long before they had any disclosures about AI-generated books. (To clarify, I disclose that Alex Irons is an AI in every book. Additionally, I give any money I might make from those books to the Second Harvest Food Bank of Silicon Valley.) I’ve published fourteen books, entirely written by AI, on Amazon since 2022, and candidly, they all suck.

Large language models do an excellent job generating a press release about a new product launch or creating a bedtime story for your kids. However, creating a long-form work of 50,000 plus words is much more difficult. While they cannot do it today, there is a path to get there.

In 2022, Alex’s first book, Artificial America, took me close to twenty hours to create from start to finish. While some of this was me figuring out how to optimize the models, the tools were generally limited. If I had GPT 3.5 (which was world-class at the time) write a story of a thousand or more words, it would have gotten ugly. The tool would have difficulty creating a story with a clear beginning, middle, and end. I found that limiting the token output to approximately 500 words helped dramatically. So, I created a spreadsheet list with different genres and writing styles for the AI to mimic in each of the 50+ short stories that make this book.

Ultimately, it was a lot of work for a product that wasn’t that good. Earlier this month, I had Claude 3.5 Sonnet write a book called The Depths’ Warning. Again, it isn’t that good. But the time it took to do this is exponentially less than in 2022. Writing the book in Claude, including having it brainstorm a topic, create an outline, and write the piece, took about ten minutes to produce 12,000 words. So, the time needed to produce a work went from >10 hours to 10 minutes in two years.

That is only going to accelerate. AI will be able to write a good book, and it may even be able to do it today. All of these books are effectively completed in a single attempt. These are just first drafts. And in this case, what is true for humans is also true for AI. As Ernest Hemingway famously said:

All first drafts are shit!

An AI doesn’t just have to be the writer. It can also be an editor, a book critic, and your targeted reader. Different AIs with different personas can give feedback and rewrite the work dozens of times. A process that takes humans years to do will take minutes.

If an AI is good at writing a cover letter, there’s no reason it can’t be good at writing a romance novel or historical fiction. Just like with us, it needs to do the reps to take it from a first draft to a finished product. And with expanding context windows and the promise of AI agents to come, this will happen in the near future.

Final Thought: More isn’t always better

Inarguably, Generative AI is better today than when I began writing about it. But is there any example of it being worse?

I asked many of the leading models the same question. Could I beat Mike Tyson in a fight? The responses were lengthy and verbose. Here are snippets I got from a few:

• ChatGPT-4o said, “In short, it's highly improbable that you'd win against someone of Tyson's caliber…”
• Mistral Large 2, “It's highly unlikely that you would beat Mike Tyson in a boxing match, even at his current age….”
• Claude Sonnet 3.5, “Based on the information you've provided, it would be extremely unlikely for you to beat Mike Tyson in a boxing match, even at his current age of 58.”

I wondered if this is just, as Elon Musk infers, new AI models withholding the truth and giving “PC” answers. So I went to the latest version of his model, Twitter’s Grok 2, and here’s the long, rambling answer it provided:

Nope. Same type of response as the others.

So, how do these answers compare with the same question asked to GPT 3 two years ago?

“No.” Well played, GPT3. You got right to the point. And no model that we can ever create could give a better answer than that.