At my promotion ceremony, I reflected on how, despite tech's evolution in our military journey, it's our enduring human bonds and discernment that truly define us.
One thing I wanted to push back on a little though was the part where you wrote "I was discussing the use of large language models with another group of leaders a few days ago, and we kept coming back to one point: These systems don’t know how to handle uncertainty. They accept information at face value, without discerning its quality or context. They then present information as if 100% true, despite being wrong a significant fraction of the time."
While I think this is right in certain contexts, I think recent-gen LLMs might be able to provide a bit more nuance than you've experienced previously. I also think a lot depends on how they were trained (e.g. maybe the ones you've been using are rewarded for accepting your inputs at face value). Would you mind providing an example of a situation where you feel they've fallen short?
Thanks, appreciate the comment. I agree with you that the current generation has improved on this to an extent (and I think we're also setting our standards higher than they need to be...but that's another post).
In terms of personal experiences, I can't share specifics from a work context, but in a personal context I've seen a lot of, say, X-Y problem issues. For example, a few months ago I was experimenting with Flask and I asked ChatGPT how I could run Flask on GitHub Pages. It gave me several pages of tutorial instead of saying "those two things don't go together, which one is more important?" or something similar. That's been fixed now, but similar types of things are what I'm talking about.
Glad you're writing regularly again Chase.
One thing I wanted to push back on a little though was the part where you wrote "I was discussing the use of large language models with another group of leaders a few days ago, and we kept coming back to one point: These systems don’t know how to handle uncertainty. They accept information at face value, without discerning its quality or context. They then present information as if 100% true, despite being wrong a significant fraction of the time."
While I think this is right in certain contexts, I think recent-gen LLMs might be able to provide a bit more nuance than you've experienced previously. I also think a lot depends on how they were trained (e.g. maybe the ones you've been using are rewarded for accepting your inputs at face value). Would you mind providing an example of a situation where you feel they've fallen short?
Jonathan,
Thanks, appreciate the comment. I agree with you that the current generation has improved on this to an extent (and I think we're also setting our standards higher than they need to be...but that's another post).
In terms of personal experiences, I can't share specifics from a work context, but in a personal context I've seen a lot of, say, X-Y problem issues. For example, a few months ago I was experimenting with Flask and I asked ChatGPT how I could run Flask on GitHub Pages. It gave me several pages of tutorial instead of saying "those two things don't go together, which one is more important?" or something similar. That's been fixed now, but similar types of things are what I'm talking about.