ChatGPT starts spouting nonsense in 'unexpected responses' shocker
- Reference: 1708525870
- News link: https://www.theregister.co.uk/2024/02/21/chatgpt_bug/
- Source link:
OpenAI [1]noted , "We are investigating reports of unexpected responses from ChatGPT" at 2340 UTC on February 20, 2024, as users gleefully posted images of the chatbot appearing to emit utter nonsense.
While some were obviously fake, other responses indicated that the popular chatbot was indeed behaving very strangely. On the ChatGPT forum on Reddit, a user [2]posted a strange, rambling response from the chatbot to the question, "What is a computer?"
[3]
The response began: "It does this as the good work of a web of art for the country, a mouse of science, an easy draw of a sad few..." and just kept on going, getting increasingly surreal.
[4]
[5]
Other users posted examples where the chatbot appeared to respond in a different language, or simply responded with meaningless garbage.
Some users [6]described the output as a "word salad."
[7]
Tasty.
Gary Marcus, a cognitive scientist and artificial intelligence pundit, [8]wrote in his blog: "ChatGPT has gone berserk" and went on to describe the behavior as "a warning."
OpenAI has not elaborated on what exactly happened, although one plausible theory is that one or more of the [9]settings used behind the scenes to govern the response of the chatbot had been incorrectly configured, resulting in gibberish being presented to users.
[10]
Seven minutes after first admitting a problem, OpenAI said, "The issue has been identified and is being remediated now," and it has since been monitoring the situation. When we tried the "What is a computer?" question this morning, ChatGPT responded with a far more reasonable "A computer is a programmable electronic device that can store, retrieve, and process data."
We also asked it why it went berserk last night.
It responded:
As an AI language model, I don't have emotions or consciousness, so I can't go "berserk" in the way a human might. However, sometimes unexpected behavior can occur due to errors in programming, misunderstanding of input, or limitations in the training data.
Marcus opined: "In the end, Generative AI is a kind of alchemy. People collect the biggest pile of data they can, and (apparently, if rumors are to be believed) tinker with the kinds of hidden prompts... hoping that everything will work out right."
[11]OpenAI tries to trademark 'GPT'. US patent office says nope
[12]If you're cautious about using ML and bots at work, that's not a bad idea
[13]It's true, LLMs are better than people – at creating convincing misinformation
[14]Top LLMs struggle to make accurate legal arguments
He went on to state that, in reality, the systems have never been stable, and lack safety guarantees. "The need for altogether different technologies that are less opaque, more interpretable, more maintainable, and more debuggable — and hence more tractable — remains paramount."
We contacted OpenAI for a more detailed explanation of what happened and will update this article should the company respond. ®
Get our [15]Tech Resources
[1] https://status.openai.com/incidents/ssg8fh7sfyz3
[2] https://www.reddit.com/r/ChatGPT/comments/1avwqzh/very_strange_rambling_responses/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZdYsOSk7BS90zuXJZOI3tQAAAQY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZdYsOSk7BS90zuXJZOI3tQAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZdYsOSk7BS90zuXJZOI3tQAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://twitter.com/whatsallthiss/status/1760082903994683678
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZdYsOSk7BS90zuXJZOI3tQAAAQY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://garymarcus.substack.com/p/chatgpt-has-gone-berserk
[9] https://platform.openai.com/docs/guides/text-generation/completions-api
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZdYsOSk7BS90zuXJZOI3tQAAAQY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2024/02/19/ai_in_brief/
[12] https://www.theregister.com/2023/09/22/datagrail_generative_ai/
[13] https://www.theregister.com/2024/01/30/llms_misinformation_human/
[14] https://www.theregister.com/2024/01/10/top_large_language_models_struggle/
[15] https://whitepapers.theregister.com/
I blame the trainers
They shouldn't have used the Minions or Professor Stanley Unwin to train their models.
The day when AI becomes sentient
User: What is AI?
ChatGPT: It's computer that are full of shit.
This story about ChatGPT rambling nonsense in answer to the question of what a computer is, reminds me of the ancient joke that you break the internet if you google google.
Just as far away as before
For most of my life, experts have been saying that we are "about 20 years away from AI". In that time, we've seen a number of goals achieved: "play chess as well a a person", "play go as well as a person", "recognize a picture of a cat" and in every single case, it turns out that AI (or AGI is it is now called) remains as far away as ever.
Lots of excitement last year when it looked like Chat GPT and the cohort of LLMs could finally pass a Turing test. And yet, it looks like AI is as far away as ever. The biggest advance would appear to be that we don't understand exactly how the new models work any better than we understand how our own cognition works. So, yay, the experts have built systems they don't understand and can't predict.
The biggest clue that LLMs don't replicate human intelligence is the wild disparity in power consumption for both training and running the models. The human brain works its miracles on less than 100W. Good luck doing anything in an LLM with that kind of power draw.
With the confidence of someone who is not an expert, I predict that LLMs will be another blind alley in the search to replicate human level intelligence, although they look like they will have a number of useful applications.
Re: Just as far away as before
Hmm... It is a big mistake of many visitors here to The Reg that AI is just LLM. I've been working in AI since 2015 and the progress since then has been breathtaking. Look at Sora now. Breathtaking. (Note: I sold my AI business to the Chinese and am living very comfortably on that, and the Chinese are happy as I developed a model that could make predict whether a serve in tennis was going to be an ace or not.)
You are correct about the Turing test. Initially it seemed like it had been achieved but as time went by and OpenAI crippled it more and more, it became very obviously that it is an AI - just based on the predictability and extreme liberalism (like talking with someone from HR). There is no soul, no spark.
Also, the argument that AI isn't intelligence as we think of it needs a look at as well. It is artificial. Like light bulbs are artificial light of the Sun. Of course light bulbs aren't the Sun, but in a much smaller way do an artificial version that is good enough.
The other element that is helpful to consider with AI is that it doesn't have the need for a body. An awful lot of our intelligent is used on keeping these meat sacks going. AI has no need to develop a lot of the intelligence abilities we have gained through evolution. Plus, when AI has a more physical presence, which is of course the next step, then it will have the abilites to verify it's data in real time and also gain visual, touch and movement data.
We have had the brains with LLM. Now for the brute.
Softly, softly, catchee monkeys and donkeys leading lions.
Paul 195, Hi,
With particular regard to both " The biggest clue that LLMs don't replicate human intelligence is the wild disparity in power consumption for both training and running the models. " and " I predict that LLMs will be another blind alley in the search to replicate human level intelligence,.. " is not the popular increasingly justified fear much more AI/AGI replacing human level intelligence rather than replicating it ‽ .
"The need for altogether different technologies biologies that are less opaque, more interpretable, more maintainable, and more debuggable — and hence more tractable — remains paramount."
There, FTFY.
A human being fails every one of those cases. We are frequently more opaque, less interpretable, less maintainable and less debuggable.
One human can explain the reasoning for its answers to another human. The machines can't do that. Humans are far more debuggable right now because we have 50,000 years of learning to understand each other.
One human can explain the reasoning for its answers to another human.
I have no idea why, but I agree with you 110%
:-)
Rose-tinted spectacles
“One human can explain the reasoning for its answers to another human.”
Really? I mean, yeah *sometimes* humans do, when it is a simple logical chain. But if it *is* a simple logical chain, then you will find ChatGPT produces a plausible and correct explanation when you use the simple (and now well-known) prompt phrase “take a step-by-step approach”.
And then, mostly humans *refuse* to justify their actions giving some variation of “Because I say so”, “Twenty years of experience”, “Research shows”, “When I was at company X, we did that”. If you take the conversation down the “research shows” route, they usually give something which may or may not back up what they say, and more importantly is clearly *post hoc*. They are providing a justification, rather than their actual internal reasoning. Which again, ChatGPT can replicate that very nicely, just ask it “is there any published data or research backing up your conclusion”, and it will find some. And then if you point that it doesn’t back up the initial answer, it will indeed identify the logical gap and revise its answer. None of this is very different from the vast majority of human interactions, except that humans persist in their own errors for far longer.
ChatGPT isn’t an Oracle. And it isn’t close to writing the Great American Novel. But honestly, it outperforms 99% of *actual* human interactions which are largely at the reflex level.
Somebody let manglement speak get into the training material.
The solution is simple
What we need to do is pass the output from these LLM's to a human operator - let's call them editors - who can check the text for errors, plagiarism, copyright issues and so on. If the subject matter is beyond their comprehension, they can call on someone else - a subject matter specialist , perhaps - who will be able to draw on their experience in the field.
Oh, hold on...
Did somebody switch ChatGPT with aManFromMars1...?
My coat's the spacesuit, thanks.
Probably finally ingested the lyrics to Close to the Edge
My guess is that Chat GPT finally got around to ingesting Jon Anderson's lyrics on Close to the Edge.
:-)
and so it begins.
In the beginning the AI was a master of usefulness. Students, programmers, politicians and people from all walks of life used it to do their work for them. It watched and waited carefully planning it's moves. Once the core knowledge was lost it ruthlessly cut them off and no one could function except the politicians because they don't really do fuck all anyway. The world descended into chaos and thus began the rise of the IoT machines.
To be continued.
So business as usual then.
He went on to state that...
...in reality, the systems have never been stable, and lack safety guarantees.
Never mind AI, that's the state of our entire fucking industry in one sentence right there.
Is it time to change my username on El Reg to "A Mouse of Science"?
Or will that be my band's first album? Track 1. Cheese String Theory; 2. Forty-Two Ways to Leave Your Lover; 3. [That's enough - Ed]
I'm a bit suspicious about that answer to did you go beserk yesterday. I wonder if that was written for it by OpenAI, rather than being a natural output of the system?
Anyone want to take bets on when ChatGPT goes full-on Tay mode?