In the hilarious Bill Murray* comedy Stripes, the misfit platoon is late for basic training graduation, and the general asks them where they’ve been. Murray’s response is, “Training, sir.” To which the general asks, “What kind of training?” And in classic Murray fashion, the answer is, “Arrrmy training, sir.”
(https://www.youtube.com/watch?v=rHcMxOJ5BN4, 2:15 minutes)
[Image Credit: https://www.deviantart.com/topher147/art/Bill-Murray-Army-Training-591625354]
One might ask this same question, “What kind of training?”, about ChatGPT and other AI based tools we may access regularly to have fun, do work, explore, create, etc. When we engage with these “entities” (as they often seem to interact with us in almost conversational ways), how much do we know regarding the data used to train AIs or GPT models?
How does the data used to pre-train the GPT affect the results we receive, and can we detect biases and limitations based on the results? How can we rely on algorithms which most of us don’t understand?
This might be a challenge unless we have some way of validating or corroborating the results.
For example, I recently did some work on a project for my college reunion, and we tried out ChatGPT to collect details about the world during and since we were in school. The results generally tracked with my memory of the past decades. But I don’t know what was excluded, what emphasis was placed on specific historical events. However, even the OpenAI help page cautions that ChatGPT can provide incorrect information and that users should check other sources. In my case, my possibly faulty memory. (https://help.openai.com/en/articles/6783457-what-is-chatgpt)
The challenge of understanding or being alert to how the training data impacts the responses confronts all of us. Do we just the technicians trained the model properly, using data we feel is appropriate? Do we trust the developers that created the reward system that tells the GPT that a response is correct? Can we rely on the tech companies to install the proper guardrails, so AI tools are helpful without being harmful? Do we apply the same judgment we use in evaluating human truths? Are any of us even equipped to understand how AI training will manifest?
Ted Chiang addressed these questions brilliantly in his novella, The Life Cycle of Software Objects. (https://en.wikipedia.org/wiki/The_Lifecycle_of_Software_Objects) In the work, the main characters educate the software entities over a period of years, much like raising children or pets. They rewarded behavior, exposed the entities to new information, and experience, and so on. The owners/parents trained the software entities to be like adult humans, but it took many, many years. It’s only been a couple of years since ChatGPT was released.
Using this paradigm, should we evaluate AI’s performance, searching for nuance and effects from how it was trained, the same way we look at how upbringing affects us and the people we meet?
On the darker side, in a climactic scene in the movie M3gan, the question of understanding how an AI will respond based on training comes up in an unfortunate way. The robot is busy attacking the main character in a dark and dismal basement, and it confronts Gemma:
Gemma: [shakes her head] Look, this is all my fault. I didn’t give you the proper protocols …
M3gan: You didn’t give me anything! You installed a learning model you could barely comprehend, hoping that I would figure it out all on my own. … [Ref: https://www.imdb.com/title/tt8760708/characters/nm6761813]
But not to be too dark and dreary about AI, Star Trek, The Next Generation, as it does, gives us hope. In Season 7, Episode 23, Emergence, the Enterprise creates a new conscious entity. In the ending scene, Captain Picard provides insight into how what we used to train AIs may determine what result we get:
“The intelligence that was formed on the Enterprise didn’t just come out of the ship’s systems. It came from us. From our mission records, personal logs, holodeck programs, our fantasies. Now, if our experiences with the Enterprise have been honorable, can’t we trust that the sum of those experiences will be the same?“
[Ref: https://memory-alpha.fandom.com/wiki/Emergence_(episode)]
And in the end, does the process matter or is the outcome what’s important? Did the general in Stripes care that Murray’s unit learned the drill in an unconventional way, or was he just satisfied that the drill was completed successfully?
It’s going to be an interesting time, to understate the obvious, as we learn how this new technology will alter our world. Let’s hope it’s more Star Trek and less M3gan.
Leave a comment and let me know what you think.
Thanks for stopping by.
*My reference to Murray’s work in Stripes does not condone or apologize for his alleged inappropriate behavior.
[Disclaimer: Please accept my apologies for any ads that pop up before the linked videos. They do not reflect my position, nor do I endorse any of the products—it’s just a YouTube thing I can’t get around.]