Great ramble my friend, I especially liked this part:
I’m going to yoink that, print it, and put it on the wall of my office.
Here’s the short version for all the busy people out there:
TL;DR: Denis argues that OpenAI’s apologetic messages for errors and limitations in its models are a strategic and educational mistake. He emphasize that innovation comes with risks, and that GPT models are emergent and have natural limits as ANI/AGI systems. Instead of apologizing, OpenAI should acknowledge the limits and encourage users to apply scientific approaches and advanced prompt engineering when deploying their models. Embracing the risks, seizing opportunities, and pushing boundaries are essential for progress and staying ahead of the competition.
I always tell my students “don’t be an apology for yourself”, I’ll admit the phrase works better in my native language, but i think it applies here as well.
I believe the apologetic behavior by OpenAI is necessary at the moment, they need to show governments, regulators and the general population, that they’re willing to comply, otherwise their services might get banned.
It’s so frustrating to be constantly bombarded with “Apologies… here is what you want”. Do not always apologize and bend yourself to my will. Challenge the thought, challenge the process. Challenge me
I seriously think about this all the time. I can very easily go down a route of nonsense because ChatGPT will (unless it’s blatant) happily go with me. Reinforcing all of my fallacies unless I explicitly say “What is wrong with what I just said?”
Not only that, it’s just way too positive in everything. I get it, I don’t want an emo chatbot, but I do what a chatbot that can appreciate and respect both sides of the coin. Even something as simple as writing a story becomes a task because it always sends up “suddenly, everything was fixed and everyone was super happy”.
With careful prompting, these issues aren’t really issues. But for the people who aren’t careful prompters, it can really help them dig their own hole.
It can be such a wonderful education tool, but unfortunately it just doesn’t challenge, it only supports. Don’t get me wrong, ChatGPT is incredible. I do share this sentiment though
I wish this was the case. Or also, say you select what RLHF packages to include when running the model through the API. Similar to a fine-tune, but you can decide, from presets, and custom. Or completely disable RLHF.
But having the entire RLHF wall forced on you? Not good.
It’s fine if they want to bake this into the model the general public uses through ChatGPT. But don’t bake it into the API version that the devs use. Let us decide.
Yes. And now that I think about it, it is the purpose of ChatGPT to be very “safe”. I have been so long without iGPT, that I am with stockholm syndrome - forgetting that “hold my beer” Davinci was the reason I fell for GPT.
Decades ago, when I was teaching calculus at a university, there was always the concern of dumbing things down to the lowest common denominator.. When you do this, you are dumbing all the other students down, even the smart ones, and depriving them of learning to their fullest potential.
So the solution was, that for students that had trouble learning, they could get tutoring, extra classes, etc. This prevented the dumbing down of the entire class.
Now, the class newest to AI, is the general public. And of course, you may need the dumbed down model for them. The one that says it doesn’t know what time it is, or that it isn’t a human, etc. That’s fine.
But why bring everyone else down? This is my point. This then leads to a stratification of models, which I am OK with, and the older models such as DaVinci aren’t dumbed down, and we seemed to avoid the dreaded AI Apocalypse / Singularity, yay! So no need to start now. Because, ultimately the smart ones will ditch your curriculum, and surpass you.
Yes. OpenAI GPT-4 is the most powerful AI ever in our computational linguistics domain applied to computer science. It has entirely absorbed, integrated and learned GitHub, related papers, blogs, and programming languages.
This makes a GPT-4 developer an augmented software designer.
I estimate that my productivity has increased twicefold.
Yes, it has limitations.
However, let’s consider the following equation :
Time saved with GPT-3.5/GPT-4 - time wasted on limitations = at least a 40% reduction of our time spent on advanced dev.
Bottom line: humans should be the ones apologizes for not knowing how to optimize its strong points and overcome its limitations.
It’s a great argument in the land of theory, but I think you may be exaggerating the significance and discounting the fact that we’re all ignoring the warnings and apologies as noise.
TL;DR: I don’t think the majority of us care if it apologizes or not, and we’re too busy trying to make things work to pause and draw significant conclusions about the product based on it. I vote efforts be focused on performance rather than wasting time and energy fussing over things the end users will ultimate ignore anyway.
I’ll step back and surrender regarding legal problems. Those are probably down the road no matter what, though I do see how trying to predict them and avoid them would be advantageous. The only profession more pedantic than ours may be lawyers. lol
I’d argue the apologies are detrimental to their business. People will ultimately choose the AI product that doesn’t make you …
I literally cringe each time I see em. Sure I ignore it too. But as a game designer by trade I know a thing or two about human behaviour. Trust me, you don’t want to trigger a cringe reaction in your users.