Way, way back in the dim and distant past — circa 2021 — I was watching the emergence of machine learning models capable of some forms of text-to-image generation. I was intrigued. Learning to draw has long been on my Good Intentions list. I have no skill with drawing at all, and I’m frustrated sometimes because I want an image to illustrate something I’m writing, and sometimes I’m unable to find any suitable stock images. I’m only just now learning how to approach actual artists for commissions. I’ve tried creating my own, but rarely succeed. Did I mention “no skill?” So when DALL-E came out, I thought that maybe it was the answer.
I tried with DALL-E (various versions) and then ChatGPT to generate an avatar for twoprops.net. At the end of the process, I wasn’t terribly happy with the result after I had invested a lot of time in the project. I had this aching wish that I had spent that time and energy trying to learn to draw. Oh, I have no illusions that I could have achieved what I wanted in that time, but I would have learned something. Instead, I spent a lot of time and significant money and ended up with an avatar that met my requirements, but I felt as though it had no soul.
When I told people of my disappointment, at various points in the years-long endeavor, I always seemed to get the same advice:
- “You need to improve your prompting skills. Maybe watch some YouTube videos or take a class.” Great, so I just didn’t waste enough time to get the result I wanted.
- “You were using a model that was too old. They’re much better now.” On repeat with each new model release.
- “Maybe your expectations are too high.” Yes, maybe.
The worst, though, was after I finally settled on an avatar. I did yet another image search and I could quickly and easily identify much of the original art from which the LLMs had borrowed to generate the images I was shown. Really, there was nothing original or creative in any of them. They were just collages with better blending.
At that point, I was beginning to learn more and more about how LLMs worked and the massive environmental impact. I saw one after another major corporation abandon all their environmental initiatives “because… AI!” I lost interest in using LLMs.
I’ve observed a conundrum in times past. Someone will ask me to read a book or watch a video. My time for such things is pretty limited — my list of books I want to read seems infinite. When someone suggests a work by someone whom I know to have lied or deceived in the past, who uses circular or fabricated references, I see no reason to invest more time in analyzing their work. But inevitably someone will say “how can you say it’s bad if you haven’t read/watched it?” Good question, but I certainly don’t want to chase every crackpot theory that someone sets down in a YouTube video. But as the rhetoric around LLMs started heating up, I thought I should at least give them another look.
I created an image of Infrastructure Man for twoprops.net. It isn’t bad, but it also isn’t good. And as time has gone on, I’ve become increasingly embarrassed by it. I have commissioned a real, live, breathing artist to re-do it, and I have a strong feeling it will be much, much better and less the product of other folks’ work.
I also tried using several “frontier models” to try to solve a coding problem. Three of them said it couldn’t be done. One gave a completely bogus answer. In fact, every coding problem I have ever submitted to an LLM has resulted in code that was wrong. Not just “not up to my standards.” Wrong, as in, it will not possibly work. That’s not to say the answers were always useless; sometimes they gave me a new perspective and helped me solve the problem. But still: Wrong.
I figured out how to do the job anyway. I coded it, wrote it up, and it works phenomenally well. I wrote about it in these very pages. Some months later, I thought it would be interesting to try again. This time an LLM gave the right answer. Mine. Without attribution.
I remain largely unimpressed. I find the more I know about a topic, the less satisfactory the results from LLMs are. When I factor in that the actual cost of running an LLM is currently massively subsidized by venture capital, the whole LLM business starts to look like a real house of cards. When I do an incompetent job of writing software, I don’t declare that my program “went rogue” and did bad things. I say “ooops, my bad, that was a bug” and I fix it. LLMs have no more agency than helloworld.c.
Does that mean they’re useless? Not at all, but the current “our LLMs are going to grow so powerful they’ll destory humanity… BTW, we want to sell our LLMs to the Department of War so they can control strategic weapons” hype is just… overblown.
—2p