Human Error And Artificial Intelligence

The Bleeding Edge Edition #2

As I was buttoning up this morning’s newsletter, I realized I had made a factual error.

The first half of the newsletter was built on incorrect information. Removing it made the rest of the newsletter empty and unfinished without the context. So I’m calling an audible and will release the new version Wednesday morning at 7am EST.

It’s unfortunate because the newsletter sounded really good. This is ironic because the popular AI text tool chatGPT has the same issue. The tool is designed to give you answers when you ask it things, not be accurate.

When you use it, be cautious of mistaking the novelty of impressive tech for an all-knowing machine.

While asking chatGPT about Christopher Columbus, it told me he was an explorer from the late 15th century to the early 16th century. I knew he sailed the ocean blue in 1492, but I didn’t know he was also the oldest person to ever live!

He wasn’t. ChatGPT was incorrect. When it presents a fact, be sure to fact-check it.

A very effective use of chatGPT is to give it paragraphs of complex text from places like Wikipedia and tell it to repackage it in new ways. Let’s say you want to learn about the discovery of exoplanets. The format can be as follows:

Imagine you’re a teacher. I will give you a few paragraphs of text. You will summarize it in a 9th-grade reading language [paste Wikipedia text here].

[chatGPT answers]

Let’s take it a step further.

Imagine you’re a video creator. Using your answer above, you will give me a punchy script for a video that will be 60 seconds long. Use a strong opener that gets the viewer’s attention.

When I tried that, chatGPT gave me instructions for B-roll and voiceover.

Would this lead to a viral video? Probably not. Is it a pretty damn good start? Yes.

It’s also important to note that between the first summarization to creating the script, chatGPT oversimplified the line about the first exoplanet discovery. The first suspected observation of an exoplanet occurred in 1988, but the first confirmed observation came in 1992.

Again, verifying the information chatGPT presents as fact is important.

It’s also important to note that chatGPT is the worst it will ever be.

In the not-so-distant future, a friend will text you a 20-minute YouTube video, and you’ll say, "Hey Siri, summarize video."

I’d love to hear what you’ve done with chatGPT. What have you used it for? What have you found useful?

I’d also love to know if you have connections with any artists dabbling in AI art. This week I went neck deep using Midjourney, easily the most capable AI art software I’ve used so far.

I was able to make some really fun planet concept art.

See you Wednesday morning at 7am EST!