Humble Beginnings

It seems like just yesterday when we were all agog over dogs and pagodas (if you know, you know). The images themselves were generally underwhelming, but they pointed to an exciting future, to the idea that we were on the precipice of fundamental change. 10 years later, and at least in the creative space, that realization has been accomplished. Finally, the hardware and software capability has caught up to the vision. Data sets used for training have become absolutely massive, and the modern results seem astonishing.

Great Marketing

Although pattern replication algorithms have been around for a number of years, it is only recently that they exploded in the mainstream. And with it has come a strong push to brand them as AI. And why not? At first glance they appear to be quite aware and sentient. Their response to inputs is fascinating and frequently surprising in its depth.

But if you look past those initial results, it starts to become apparent how stunted these algorithms actually are.

Expansion

And today, its become so much more. From generative art, we have made the leap to other creative endeavors in audio and text with algorithms, trained on broad datasets but focused so specifically on their subject, that they appear sentient to the everyday person, or anyone seeing the effect in passing.

Emergent Problems

I’ve had this same discussion where I state that ChatGPT is not an AI in a number of discord channels, and it has been widely rejected. Even the most techy people want it to be AI and will look past the shortcomings.

Seeing past the curtain

• Repetition
• Subject matter expertise

One of the quickest ways to see the limitations of ChatGPT is to simply ask it similar prompts, or the same prompt multiple times. What happens, is a number of similarities show up and you can see that whole sections of text are simply repeated, with maybe a few different descriptors.

Another way is to ask it very specific questions about an area of expertise you have. Coding is a great example. Non-coders see a wall of codey looking code and a youtuber telling them it works, and they are amazed. But, professionals look at the code and see that it’s full of gibberish and non-essential elements. And ERRORS. So, so many errors. And that’s the problem

It doesn’t work the way you think it does

I think the number one issue that keeps people from easily seeing that these are algorithms and not AI is because they don’t understand how it works. Many people assume that the algorithm is working on the photo they see, or the words they read, or the music it makes. But in reality, it’s using underlying patterns and structures invisible to humans to replicate things based on the equivalent of usage heat maps. The algorithm has absolutely no understanding of the inputs it is replicating, nor of the output it produces.

Conclusion

• Memory
• Self awareness
• Continuing education

ALtho the modern algos are flashy and seem quite capable, they are still a long ways off from true artificial intelligence. How long until we bridge that gap? I really don’t know, it could be tomorrow just as easily as it could be another 50 years.