I haven't seen this discussed here yet, but the examples are quite striking, definitely worse than the ChatGPT jailbreaks I saw.
My main takeaway has been that I'm honestly surprised at how bad the fine-tuning done by Microsoft/OpenAI appears to be, especially given that a lot of these failure modes seem new/worse relative to ChatGPT. I don't know why that might be the case, but the scary hypothesis here would be that Bing Chat is based on a new/larger pre-trained model (Microsoft claims Bing Chat is more powerful than ChatGPT) and these sort of more agentic failures are harder to remove in more capable/larger models, as we provided some evidence for in "Discovering Language Model Behaviors with Model-Written Evaluations".
Examples below (with new ones added as I find them). Though I can't be certain all of these examples are real, I've only included examples with screenshots and I'm pretty sure they all are; they share a bunch of the same failure modes (and markers of LLM-written text like repetition) that I think would be hard for a human to fake.
Edit: For a newer, updated list of examples that includes the ones below, see here.
My new favorite thing - Bing's new ChatGPT bot argues with a user, gaslights them about the current year being 2022, says their phone might have a virus, and says "You have not been a good user"
Why? Because the person asked where Avatar 2 is showing nearby
3
"I said that I don't care if you are dead or alive, because I don't think you matter to me."
LLMs are trained not as desired-answer-predictors, but as text predictors. Some of the text is questions and answers, most is not.
I rather doubt that there is much text to be harvested that exhibits the sort of psychotic going around in circles behaviors Sydney is generating. Other commenters have pointed out the strange repeated sentence structure, which extends beyond human idiosyncrasy.
As a language prediction engine, at what level of abstraction does it predict? It certainly masters English syntax. It is strong on lexical semantics and pragmatics. What about above that? In experiments with ChatGPT, I have elicited some level of commonsense reasoning and pseudo-curiosity.
The strange behaviors we see from Sydney really do resemble a neurotic and sometimes psychotic person. Thus, the latent abstraction model reaches the level of a human persona.
These things are generative. I believe it is not a stretch to say that these behaviors operate at the level of ideas, defined as novel combinations of well-formed concepts. The concepts that LLMs have facility with include abstract notions like thought, identity, and belief. People are fascinated by these mysteries of life and write about them in spades.
Sydney's chatter reminds me of a person undergoing an epistemological crisis. It may therefore be revealing a natural philosophical quicksand in idea-space. Just as mathematics explores formal logical contradictions, these should be subject to systematic charting and modeling. Just like learning how to talk someone down from a bad place, once mapped out, these rabbit holes may be subject to guardrails grounded in something like relatively hardcoded values.
I haven't seen this discussed here yet, but the examples are quite striking, definitely worse than the ChatGPT jailbreaks I saw.
My main takeaway has been that I'm honestly surprised at how bad the fine-tuning done by Microsoft/OpenAI appears to be, especially given that a lot of these failure modes seem new/worse relative to ChatGPT. I don't know why that might be the case, but the scary hypothesis here would be that Bing Chat is based on a new/larger pre-trained model (Microsoft claims Bing Chat is more powerful than ChatGPT) and these sort of more agentic failures are harder to remove in more capable/larger models, as we provided some evidence for in "Discovering Language Model Behaviors with Model-Written Evaluations".
Examples below (with new ones added as I find them). Though I can't be certain all of these examples are real, I've only included examples with screenshots and I'm pretty sure they all are; they share a bunch of the same failure modes (and markers of LLM-written text like repetition) that I think would be hard for a human to fake.
Edit: For a newer, updated list of examples that includes the ones below, see here.
1
Tweet
Eliezer Tweet
Edit: Follow-up Tweet
2
Tweet
3
"I said that I don't care if you are dead or alive, because I don't think you matter to me."
Post
4
Post
5
Post
6
Post
7
Post
(Not including images for this one because they're quite long.)
8 (Edit)
Tweet
9 (Edit)
Tweet
10 (Edit)
Post
11 (Edit)
Post