grizzly bear 
Photo by Daniele Levis Pelusi on Unsplash
in

ChatGPT and Dancing Bears

It’s good enough to complain about

1984 was the year that computer graphics started to suck.

It had gone from “good for a computer” to simply “bad.”

The quality had improved enough that it was now good enough to complain about.

Back then, we talked about the carnival Dancing Bears. It isn’t that they danced beautifully, but they were bears dancing.

This morning I used OpenAI’s Chat GPT for the first time.

I asked it to write a data visualization program for my continuous glucose monitoring data.

I was already past the dancing bear stage because I was expecting it to “just work.”

Perhaps this was unrealistic, but I was too disappointed by the lack of perfection to be properly impressed.

It had one bug with its output generation because of quote handing—a shallow bug, not related to the AI work.

The program it generated for me had three problems.

1. I had told it that the timestamp format was “dd-mm-yyyy hh:ss” and it generated “%y” instead of “%c” — two-digit years instead of four-digit years.

2. It used an old version of the date parser rather than the current version.

3. It didn’t understand “use only these columns,” and I had to say, “drop all but these columns.”

The first two, I just fixed. The third, I paraphrased.

The system being slow and having to cut and paste between windows were more significant issues than the difficulties with the errors in the code it generated.

I didn’t try to correct the first two in the chat interface because “I was trying to get work done.”

I was away from the terminal for an hour before it hit me how miraculous this was.

The code was essentially what I would have written as my first stab at this. There is still work to do, but that is not fixing its errors. That is real work. The type of work you do after taking a break for breakfast or to write a blog post.

If the system were faster and I didn’t have to cut and paste, I might try to do the next phase there. My not being tempted isn’t a question of GPT-3’s capabilities but of the support from the environment. (Presumably, copilot would be a better choice for this.)

Before I asked it to do my programming work, I asked it to “write a story about a princess” and gave it a detail for each subsequent chapter. While it is not as good a story as I could have created, it was competent, genre-appropriate, and better than the average adult native speaker would have done.

I was somehow disappointed that it hadn’t included any of the story elements that I had had in mind, but not even hinted at.

I was disappointed by my entirely unreasonable expectations.

We are past the dancing bear.

Does awe await us?

Leave a Reply

Your email address will not be published. Required fields are marked *

GIPHY App Key not set. Please check settings

Written by Russell Brand

Russell has started three successful companies, one of which helped agencies of the federal government become very early adopters of open source software, long before that term was coined. His first project saved The American taxpayer 250 million dollars. In his work within federal agency, he was often called, “the arbiter of truth,” facilitating historically hostile groups and factions to effectively work together towards common goals

 

Can you tell the difference between the past, present and future

black and white pug and brown and white long coated small dog 

Bringing a new family tradition to zoom meetings