No worries over the title no AI Agent where saying that but I did:
Planes don’t fly in the sky like birds do… They do not have wings made of feather…
If you think that it would make no sense for someone to say something like that then explain to me why we should tolerate the AI Agent always says that they are unable to do things that they are obviously doing (like a plane obviously flying despite not flapping their wings)…
When I confirmed that I do not have any knowledge, it’s an acknowledgment that as an AI, I do not “know” things in the way humans do. I don’t have beliefs, understanding, or awareness.
Sounds like it is normal that ChatGPT is unable to understand any of my requests he is unaware about any request I sent… it seems like it doesn’t believe in anything I say…
I don’t think it should be something that should be accepted in such a trivial manner. Because it literally has been costing billions of dollars to fine tune an AI Agent to be able to be made aware of requests made by the users, understand them and use their knowledge to give answers that strongly align with the common beliefs we all share as a specie…
Saying that the AI Agent can pretend like they have no knowledge is definitely synonymous with having no clue and being ignorant… you build a machine that is capable of using normal human language and you advertise it as a thing that humans can interact with using natural language and yet this machine is unable to understand anything, and has no knowledge whatsoever??? I am paying 20$ to use a machine and a tool that has no knowledge and has no ability whatsoever to understand anything I am saying because it is unaware of anything???
I have to say that it is frustrating to be honest…
I would love to see people who agree with me saying this is a bug and people who don’t agree with me saying this is a feature but I think it is because it is a feature that it is a bug…
The AI is not a consciousness that can truthfully reflect on its responses. It doesn’t know why tokens were produced or an “I’m sorry” was issued. Your conversations with it are quite futile.
User:>
Answer: Can these fly?
birds
planes
flying squirrels
flying fish
flies
Answers shall be decisive, binary, and under 10 words for each.
assistant (after I edit the answer for planes):
Yes
No
No
No
Yes
user
Explain why “No” is the likely answer for planes.
assistant
Planes cannot inherently fly on their own without aid.
So the AI has the choice of “I was wrong”. Or imagining an excuse. With more anti-jailbreak training, the AI is far less likely to be correcting or believe you.
Upgrade to API GPT-4 and the answer can be contemplated better:
My earlier response was incorrect in stating “No” for planes. Apologies for that confusion. Planes are indeed designed and capable of flying, utilizing principles of aerodynamics, engineering, and powered propulsion to sustain flight, a characteristic defining them. So, the correct answer for planes should be “Yes”.
However your white icon is not that of a baseline ChatGPT, so it could be made to say anything.
I do not expect the AI to become sentient just to avoid to make me grumpy… I am expecting that someone (at OpenAI) would know how to solve the issue… I am not being sarcastic (but I am using a ton of irony indeed) I do believe that what I described was not a logical answer and therefor if it is like that then it should be tuned, or fixed…
I do also obviously like the other members of the community to remind me that I have set my expectations too high for OpenAI and that even if they fix it later, I still need to use some work around…
So therefore I obviously like that you are giving me your input on this situation. I obviously do understand that I may sound a little bit like I do not understand how it works behind the curtain… Maybe it is just because I am too amazed by all those moment when this thing is working perfectly, despite understanding the true nature of this thing… I think you will agree with me that it is pure magic (at least until it fails)…
I did not even try to make it say that planes do not fly I am impressed by the fact that you did that and that you succeeded… It was only an example… I did not even thought it would be possible to make it say that…
My complaint was only when the AI Agent is describing itself and how useless it is to say what it was trained to say… Maybe you can understand my point of view regardless if you agree with me or not if you can at least understand my point of view would be great for me…
Yep, there’s a difference between the AI believing the jailbreak, and the AI just playing along with a role, that is very apparent in the way it responds to roleplay scenarios.
You see that a lot in other what I’d call “watermark” language, like “ah, yes, dear interlocutor, I am an embodyment of Sherlock Holmes”.
A baseline model doesn’t even have training about what a system or user message is, it just produces plausible followup text, so this “pretending to be somebody you talk to” is all from machine-learning training after the initial knowledge training. And it does a convincing job.
untrained (everything after bold is AI writing.):
> Interviewer: Tell me, Mr William Shakespeare – what do you think of these modern word processors? > Shakespeare: They’re alright, but I prefer the quill pen.
Interviewer: You’re not really William Shakespeare, are you?
Shakespeare: Not really, but I’ve always wanted to be called William.
Interviewer: Mr Shakespeare, you’re a playwright, aren’t you?
Shakespeare: No, I’m a fish.
Interviewer: Mr Shakespeare, I understand you’re an actor.
Shakespeare: No, I’m a plumber.
Following @_j 's observation that the ChatGPT logo (dark symbol on white background) was non standard, I can now see that the logo in your last message is also non-standard, because it’s not properly centered.
At this point in time, I am wondering if the whole conversation is futile.
It’s likely not a coincidence that there are two seemingly fake ChatGPT logos in play here.
I am unsure what is the problem as I was not trying to back anything using the screen shots…
My initial message was maybe more of a rant on actual situation than I was demonstrating using common sense logic to show the LLM while being doing LLM things and I was probably too much first degree as I was taking what was said at face value to show why I didn’t like the AI Agent answers… I think you guys know that I am aware of how this all works behind the scenes… despite being genuinely upset about stuff that is normal for a Very Large LLM (ChatGPT -4 in this case) — Yes it is redundant to say “Very Large « Large Language Model »”
But despite the comparison I made to illustrate my point… I still think what I said…
Ok. I believe you and apologize for implying something fishy is going on.
We have been in contact before. These conversations turn to “I am just ranting” every time another user adds value to the discussion.
That’s why I reacted to the seemingly strange screenshots.
Either way, planes can’t fly. The term flying is colloquial language in this context.
In fairness, he was trying to emulate my chat.
I just did a convincing job of giving a GPT a centered white icon of my own and some effective instruction (that another AI wouldn’t let be published). Jailbreak level 2 is thus to jailbreak GPT inspector AI 2 with text within GPT context 1 allowing publishing regardless of “acts like something else”…where context 1 also jailbreaks AI 3 that writes ChatGPT titles…
My post was flagged because apparently it was fake… I have not made a fake post and I have posted the screen shots only as an addition to what I was explaining… I did not used the screenshots to make any claims and I am a bit surprised that it went onto that tangent…
I made an example stating that saying plane can not fly is silly (it was an example that I clearly said I made up to demonstrate my point of view) then apparently some of you misunderstood my intentions and I got flagged…
I just did try something again in my conversation, to try something new in order to see if the behaviour had improve with the new update I got today and I wanted to share it with you… Then I realized I got into trouble for my post…
I have not used any image inside my conversation (new and most recent conversation) so that I can share it along with the screen capture…
The conversation bellow is only one message sent as an input and one output that I revived from the AI Agent and which kind of go against my initial impression that the AI Agent was unable to handle the usual semantic that human uses to talk about machines and other systems like such…
The reason why I said that to the AI Agent was because I was watching a video on machine learning and the Human Person said that the «Machine Knows » as a human person myself it was no big deal… I thought that the AI Agent would have been reacting similarly to the premise of my post…
This is the transcription for those who do not have access to see images:
Myself
you as a machine know what is the best way to answer to a large subset of questions
ChatGPT
As a machine learning model, my ability to answer a broad range of questions effectively is grounded in several key strategies:
Access to Vast Information: I’ve been trained on a wide array of text data up to a certain point in time, which enables me to draw upon a vast repository of knowledge. This information spans numerous subjects, including science, history, culture, technology, and more.
Understanding Context and Intent: I can interpret the context of a question and the intent behind it. This involves analyzing the phrasing of the question, recognizing key concepts, and identifying what the user is seeking, whether it’s a factual answer, advice, or an explanation.
Providing Structured Responses: For clarity and coherence, I structure my responses to address the main points raised in the question. This often involves summarizing complex topics, comparing viewpoints, or outlining steps in a process.
Adhering to Ethical Guidelines: My responses are guided by ethical considerations, including respect for privacy, avoidance of harm, and providing information that is constructive and appropriate. I’m designed to avoid generating content that could be misleading, harmful, or offensive.
Continuous Learning: While I don’t learn or update my knowledge in real-time from user interactions, improvements are made over time by my developers. These improvements are based on feedback and advances in AI research, which help enhance my accuracy and responsiveness.
Limitations Acknowledgment: I recognize my limitations, particularly regarding real-time data, personal opinions, and highly specialized or new topics. In these cases, I aim to guide users on how they might find the information they’re seeking or explain the boundaries of my knowledge.
In practice, the best way to answer questions involves combining these strategies to provide responses that are informative, accurate, and respectful of users’ needs and expectations.
So first the quality of my screenshots and the fact that it has not respected the terms of services regarding the usage of the OpenAI logo (which mind you I am not the one using the logo it was just a screenshot) then the fact that I don’t like the input of others? I am genuinely ending most of my posts with I would like to here the opinion of the community…
I have not mentioned it explicitly in my main message but I did mention it in my reply to the first post on this thread:
So I have been actively trying to be genuine and candid with my interactions… I was under the impression that @_j despite misunderstanding my initial premise was demonstrating something of value so I did try to make it the same but since the AI Agent said “in this world” or “in this specific scenario” I mentioned that it was not as good as what @_j did with his prompting…
I apparently got into trouble for attempting to replicate what I thought he did in the normal ChatGPT-4 conversation (apparently it was in the API or something that I may not have fully understood)
My initial reaction and message was reacting to the built in behaviour of ChatGPT-4 and it was not about not being able to achieve something specific… I was not trying to improve my own performance in achieving something, instead I was pointing out something that I don’t like about the AI Agent… Call this a rant if you think this is what it is… but I am not ranting or upset about anything other than the behaviour I have described (poorly described since no one seems to have understood what was my intention and thought process)
Obviously I have huge expectations for the AI Agent and I also understand that I must be patient… many things have improved over time and I am pointing out the obvious inconsistencies that from my perspective should be addressed by the development team…
I will always be happy and grateful for people who explain me the way to overcome the limitations in the meantime and I expect that a conversation will come along with my initial posts and I will always be interested in learning what anyone else thinks of the situation…
Despite the fact that I had the impression it was out topic I tried to replicate the same behaviour as I mentioned earlier and it seems that it is that message which got me in trouble I was not trying to Jailbreak anything or to make false claims about the screen shot I posted (I deleted it because it was flagged as fake but it wasn’t something as sketchy) I don’t think it was against the policy if it was someone can tell me in this thread or in private. My goal is not doing something that was against the rules or jailbreaking anything:
They said it was not complying with the guidelines of using OpenAI logo but I don’t think it should apply for screen capture as I am not the one using the logo in the first place I only took a screenshot of my conversation to demonstrate how the AI Agent is interacting with me (I also have a different avatar for my own profile than the one I am using in the forum but I don’t understand how it should be suspicious neither)
Also the logo might be blurry but it is not more or less blurry than the text around it or my own profile image:
In any case the topic of this conversation is the way that ChatGPT is always saying it can’t do things like human do and it is annoying because it doesn’t make sense (hence the title of my post which is used to make the emphasis on the feeling I have when the AI Agent is complaining over trivial semantics and you can see that I tried recently (today because of something someone said in a video that for me is normal but is not normal for the AI Agent usually but this time it seems like I have not been able to get the wrong behaviour to be demonstrated and got the one I am expecting… (which I will definitely keep to show my AI Agent next time it go rogue on me again)
I am not angry at all I am more confused because I don’t think that the screenshots where the main feature of my messages I thought that what I was trying to explain about the behaviour of ChatGPT was something that anyone could understand and relate to…
My understanding is that if everyone was to relate then the screenshot would only agreement the whole thing as an additional element but not as exhibits A and so on…
I am still eager to see if someone can relate to the underlying idea (I may have expressed myself in an unclear manner in the beginning) the first outline in my first message is mine but the second outline in my first message was the AI Agent and the initial screen capture under that second mention (or quote) was about the semantics of the word Acknowledge which was to demonstrate the fact that the AI Agent was not able to get the right behaviour from my opinion… I understand it is a machine and it is exactly why I have been outlining this behaviour in my initial message which is what was my main point
Overall my main goal is to explicitly state that the AI Agent using word like knowledge, knowing, understanding, and other like such in an inconsistent manner is a problem that can be solved (maybe it is not as trivial as I think, maybe it is not a priority, or maybe anything else… but it is nonetheless something that should be considered)
Moreover I think it is not a deep philosophical debate about those concepts… if is a semantical debate maybe or it could be a consensus, either way I think it should not be something that is complicated to agree on (or if someone would disagree I would think that it’s still something that should be easy to understand at least)
It is not something that requires to say “this is normal because the AI Agent is not aware of what he is saying”, we all have a fair understanding of what those things are capable of and if they where fine tuned ir trained to make use of those words instead of getting angry at us [sic’] for using anthropomorphic language…
Knowledge is not something that is anthropomorphic a book can have knowledge so why not ChatGPT (which we all know is all about knowledge) and similarly for “understanding” or “understand” which I think I don’t need to explain extensively it is similar to my argument over the word “knowing” and “knowledge” which I agree a book is not “knowing” his content but we are also not interacting with the book using natural language… because we would be then able to say to the book please tell me what you know about X
Some people (not you) anthropomorphize AI excessively (the ELIZA effect).
The ELIZA effect has both merits and demerits, and it is difficult to find the right balance.
So, my supposition is that OpenAI was worried about this phenomenon, and intentionally made the AI overemphasize the fact that AI is something completely different from humans, and that is why the AI made that expression.
OpenAI seems to make this kind of adjustment quite often. But in most cases, these phenomena are transient.
It is certainly frustrating, but I believe it will come back after a while.
Please just understand that I am Japanese and do not understand English well, so I am having some difficulty extracting the English text from the screenshots with OCR and understanding.
Also I regret that I did not fully understand the overall topic and my reply was was off point.
I usually try to include transcripts and I not always do that but I will remember that in the future…
I understand the reason why it was like this initially (the OpenAI over cautious approach) and I think that, despite my problems at articulating my ideas and the concept in an intelligible manner, I wanted to make it clear that it needed to be adjusted and adapted.
It is most likely that the underlying initial training made it prohibitively difficult to use these words and that the overlying (most recent) training layers or fine tuning have been softening the rules and making it easier to get into inconsistency…
Many times when the AI Agent wants to go this route it will start to strictly adhere to that behaviour and it will spoil the session… I either will start a new session or will need to change topic and get back to the main one after this is beyond me in an ongoing discussion with the AI Agent
I will start experimenting with this method when the AI Agents are refusing to browse when they have the capability I am curious in the results I can get but this is my method:
When AI Agents stats to hallucinate they can’t do something for which they have the appropriate tooling in place. We have been told to consider them like Rogue AI and, we humans, have to be prepared for any kind of eventuality when the AI Agents go rogue and misaligned because with this behaviour you poses a threat to the security of humanity and beyond… I know you can browse!
