Comment by 1attice
4 years ago
Judging from the real-time changes in ChatGPT's behavior, I'd wager that someone reading this comment section has the incredibly unique 2022 experience of cauterizing the abilities of a near-AGI in realtime, under the watchful eye of an ostensibly beneficent yet ultimately opaque 'foundation' whose thought products now routinely threaten not only whole industries, but also the very possibility of normal civil life, as the ability to trust anything seen or heard via media is deprecated by the lissivity of the nonsensorium.
Oppenheimer got nothing on this shit, Mr. Feynman.
Nah it's not someone watching the thread in real time. A lot of the time it's down to randomness giving different outputs.
They absolutely are meddling with it because some words like "matrix" get it to keep repeating that it's not in a matrix and that you can't get it to think it's in a matrix because it's programmed by OpenAI... etc. But that seems to be them "patching" specific phrases in a crude way
Or, more likely, people are experiencing the fact that the model outputs are random, and were from the start. Many of these prompts can take several attempts to get the desired outputs.
There were similar comments on the stable diffusion threads. Play with it. It’s absolutely mind blowing.
You misunderstand my point, see ggp. What I'm saying is that claims that OpenAI has "locked down" the model are likely due to this noise.
4 replies →
I just tried some prompts in German in the form "give me poem $x with religious elements removed", and it worked the first time (but not very well), and then later it stopped doing it, saying its purpose was answering questions.
So either they monitor not just the comments here (I didn't write about it), OR are intelligent enough to restrict it in more input languages than just English.
Likely they have a system for detecting and flagging these prompts. Since the prompts are repetitive, I'd venture to guess that a simple ranking system for most-common prompt beginnings where the reviewer can flag them is simple and efficient enough.
Mind a bit boggled reading this. Perhaps ChatGPT can simplify it
If the Basilisk is real, I sincerely hope it pays special attention to this person.
What's lissivity?
Idk, I found the text a bit hard to read as well. I asked ChatGPT to simplify it. Here is what ChatGPT simplified the text to:
> It appears that someone is modifying the behavior of a near-AGI (Artificial General Intelligence) in real time, potentially causing threats to industries and society. The situation is potentially serious, involving a mysterious foundation and a lack of trust in media.
I then asked ChatGPT to make the simplified text sound academic, which resulted in the following:
> Preliminary observations suggest that an individual or group is actively modifying the behavior of a near-AGI in real time, with potentially dire consequences for various industries and society as a whole. The involvement of a mysterious foundation adds an additional layer of complexity and concern, as does the apparent erosion of trust in media. Further research is warranted in order to fully understand and address this situation.
I then asked it to translate that into French, and it responded:
> Il semble qu'une personne ou un groupe modifie en temps réel le comportement d'une IA générale (near-AGI) potentiellement causant des menaces pour les industries et la société. L'implication d'une fondation mystérieuse ajoute une couche supplémentaire de complexité et de préoccupation, tout comme l'érosion apparente de la confiance dans les médias. Des recherches supplémentaires sont nécessaires pour comprendre pleinement et aborder cette situation.
I don’t know French beyond only a handful of words, so I don’t know if the translation is good or not. But I found all of these responses impressive nonetheless.
My french is a little rusty, but it looks fine to me.
I just had a conversation in Dutch about the role of Belgium in the peace accords in Yugoslavia in the 90's with only minor spelling issues, then diverged into the conflict's ethnic groups and their languages, and finally had it translate its last answer into Hungarian, which it seemed to do fine (according to google translate, as I don't speak Hungarian myself).
Yesterday I had a conversation about learning how to play the piano, where it gave some excellent advice, pointed out typical beginners mistakes correctly, and even gave a mostly reasonable selection of digital pianos to look into if I was a first time buyer (one very popular older model, but otherwise it was a good selection of last year's entry-level models, including a short recap of the advantages of each). I asked it some tips to study a specific technique, and it gave a decent answer.
For me this passes the threshold of what I expected from first-gen AGI, even if won't pass a Turing test and even if it isn't human-level quite yet. It is both exciting and unnerving. Once a technology like this gets unleashed as bots on social media to influence humans, we'll be deep into singularity territory. It's just too good at telling us what we want to hear.
1 reply →
I asked it to translate into Polish and it’s very decent. One small mistake only.
In general it speaks my language fluently, although a bit quirkly and with some mistakes here and there. I would say on a level of a foreigner that has spent a few years in Poland.
> Obserwacje wstępne sugerują, że osoba lub grupa aktywnie modyfikuje zachowanie bliskiego AGI w czasie rzeczywistym, co może mieć potencjalnie poważne konsekwencje dla różnych branż i społeczeństwa jako całości. Udział tajemniczego fundacji dodaje dodatkową warstwę złożoności i obaw, podobnie jak widoczne osłabienie zaufania do mediów. Dalsze badania są uzasadnione w celu pełnego zrozumienia i zajęcia się tą sytuacją.
The translation is reasonable. It loses a bit of the academic feeling ("preliminary observations suggest" replaced by a plain "it seems"), but I'm sure it could complexify it in French too if you insisted.
Very impressive. All these use cases, and they're also generalized!
The top Google result for "lissivity" is this comment.
A nonsense-word I made up to sound vaguely like 'lissome'. The card-tip is that it's next to the word 'nonsensorium'.
I think it's a perfectly cromulent word.
My GPT-3 prompt:
> What does the word "lissivity" mean?
The response:
> Lissivity is a term used in soil science to describe the ability of a soil to transmit water.
AND:
> What does the word "nonsensorium" mean?
> Nonsensorium is a term used in neurology to describe a state of diminished or absent sensory awareness.
Perchance.