I wanted to follow up on last week’s piece on what I perceive to be problems with OpenAI’s ChatGPT. In particular, whether what I was interacting with was ChatGPT or not. Some have suggested it couldn’t have been ChatGPT because there is no way to interact with ChatGPT except via OpenAI’s website.
That is true, but not the whole story. The story is somewhat peculiar, and no less worrisome.
I had tried to replicate my original experience a few times without success on OpenAI’s ChatGPT, so I went back to the original WhatsApp ‘version’ of ChatGPT who I was dealing with. I had originally been confident I was dealing with ChatGPT because the first time around it had told me:
But then, after publishing the piece and unable to replicate the experience — even closely — I later went back and asked it again:
For a moment I’d forgotten that I must have given the bot my name at some point — or else it scraped my profile name from WhatsApp. I was surprised that it was now denying any relationship with ChatGPT. So I probed further:
Needless to say, the link doesn’t work.
And I could find no companies with that name in the business the bot described, and I was, I think understandably, a little suspicious that a John Smith had popped up, along with a Doe and a Roe. So I asked for clarification about the bot’s relationship with OpenAI.
I have to admit, by this point I was worried. One of several things could be happening: I was interacting with a bot that was completely unaffiliated with OpenAI, and so my experience with it was not indicative, and my conclusions simply wrong. Another was that I was being played — that I was interacting with something, but it was probably more human than bot. And was enjoying toying with me.
Another was that I was dealing with OpenAI, but something that was not necessarily intended to be used in the way I was using it.
But I was still miffed. I sought clarity. Was the bot using the underlying engine of ChatGPT, OpenAI’s GPT-3, in any way?
Well, that was clear. But why was all the information about the company incorrect?
OK, so that’s a bit closer to the experience others have with ChatGPT — the non-threatening ‘butler response’ (my term, © Loose Wire 2023).
I don’t know why the bot suddenly backed off. But I was left with the same doubt, about myself, my research skills and what I thought I knew.
But I was still none the wiser about what I was dealing with, and whether my experience was any more or less indicative of OpenAI’s underlying technology. So I contacted the person who had created the WhatsApp interface. I won’t give his name for now, but I can vouch for his coding ability and his integrity.
He told me that the bot was not ChatGPT but was a rawer version of the technology that underpins it, namely GPT-3. At the time of writing OpenAI has not created an API for ChatGPT and so the only way for third party developers to create a way to access OpenAI’s technology, for now, has been by connecting via API to GPT-3.
In other words, I was interacting with a ‘purer’ version of OpenAI’s product than ChatGPT, which my friend told me had made some adjustments to make it a smoother experience. Those are his words, not OpenAI’s. Here is another way ChatGPT’s difference has been expressed:
It (ChatGPT) is also considered by OpenAI to be more aligned, meaning it is much more in-tune with humanity and the morality that comes with it. Its results more constrained and safe for work. Harmful and highly controversial utilization of the AI has been forbidden by the parent company and it is moderated by an automated service at all times to make sure no abuse occurs. (ChatGPT vs. GPT-3: Differences and Capabilities Explained – ByteXD
‘AI Alignment’ is taken to mean steering AI systems towards designer’s intended goals and interests (in the words of Wikipedia) , and is a subfield of AI safety. OpenAI itself says its research on alignment
focuses on training AI systems to be helpful, truthful, and safe. Our team is exploring and developing methods to learn from human feedback. Our long-term goal is to achieve scalable solutions that will align far more capable AI systems of the future — a critical part of our mission.
Helpful, truthful and safe. Noble goals. But only a small part of what OpenAI and other players in this space need to be focusing on. More of that to come.