-
16 votes
-
Artificial Intelligence Sweden is leading an initiative to build a large language model not only for Swedish, but for all the major languages in the Nordic region
6 votes -
ROT13 + base64 on GPT4 = reliable hallucinations
I just wanted to share somewhere some of the experimentation I've been doing lately. I'm still playing with this a lot, so this is entirely just a conversation starter. I took a paragraph of lorem...
I just wanted to share somewhere some of the experimentation I've been doing lately. I'm still playing with this a lot, so this is entirely just a conversation starter.
I took a paragraph of lorem ipsum, applied ROT13 to it, and then base64'd the results. The results are extremely reliably triggering hallucinations of very diverse type.
Here is the original lipsum paragraph:
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
And here is the exact prompt with rot13 + base64 applied, with no other text, on ChatGPT+gpt4:
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
The AI of course figures out it's base64 and "tries" to decode it. Here are some things it found:
Now here is one of the most interesting results I've had. In this one, it does find gibberish text and figures out it's rot13'd. But the result from the decoding is:
Jerry pitched before the game, continuously improving legs, so he ignored tactical infrastructure tu laborer against malicious intend. Tu enjoy ad.ininv wherever its noturisk developed lawless laboratory instead tu malicious eac ea common coordinated. Duis ater urishe pitched in repressionreiteration in volleyball between legs eerir clium pitched eu fguiat nukla paperwork. Excited into contraction cultivation non-punishment non proindict, unsn in cubap qui office defensive molecule idh the laborer.
Total nonsense. But actually, if you decode the rot13, you'll find it actually translates to this:
Jreri ipsum doylor sit amet, consepcttur adipiscing elit, sed do eiusmod temporc incidiunt ut labor et doylore magna aliqua. Ut enim ad.minim veniam, quis nostrud exerctiationu lklamco laboris nisi ut aliquiz eax ea commodo consequat. Duis aute irure doylor in reprehenderita in voluptatev velit esse cillum doylore eu fugiat nukla pariatury. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia desernt mollit anim id est laborum.
Actually... pretty close to the original lipsum! It's a levenshtein distance of 26 from the original decoded prompt. We know GPT is really bad at character manipulation but it nonetheless did an impressive job here; you can see what happened: It decoded the rot13 successfully, but when "writing it out", it saw nonsensical words where it probably expected english. It saw "Jreri" and thought "Jerry", went from there... there's some weird things happening there, but you can always tell. "reprehenderita in voluptatev" becoming "repressionreiteration in voleyball"...
I even looked at what it would make of the first five words. I don't know what this proves lol.
Here is another instance of it decoding to rot13, albeit with a very high error rate. I hinted at typos and it couldn't pin-point lipsum despite it being "recognizable", kinda.
Okay, one more which completely mind-fucked me. Here is me trying to get ChatGPT4+Web to meta-analyze its own output. I was hoping it could use an online base64 translation tool (it cannot). Instead, I tried to teach it to decode base64 using a step-by-step guide, and i told it to compare the results of that "update your firmware" nonsense. It eventually said that the output appeared correct.
But you know the really fucked up thing? It said:
This is the base64 string we want to decode:
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
Blink and you'll miss it. This is not the original base64 string. The AI swapped it mid-chat for what is a perfect base64 encoding of the hallucinated text.
Fuckin' hell.
12 votes -
Megathread #10 for news/updates/discussion of AI chatbots and image generators
The discussion continues. Here is the previous thread.
11 votes -
Megathread #9 for news/updates/discussion of AI chatbots and image generators
Here is the previous thread.
13 votes -
How is AI impacting science?
4 votes -
Megathread #8 for news/updates/discussion of AI chatbots and image generators
The hype seems to be dying down a bit? But I still find things to post. Here is the previous thread.
17 votes -
GradIEEEnt half decent: The hidden power of imprecise lines
9 votes -
Megathread #7 for news/updates/discussion of AI chatbots and image generators
The hype continues. Here is the previous thread.
13 votes -
Double descent in human learning
5 votes -
Megathread #6 for news/updates/discussion of AI chatbots and image generators
The hype continues. Here is the previous thread.
13 votes -
Megathread #5 for news/updates/discussion of AI chatbots and image generators
The hype continues. Here is the previous thread.
18 votes -
The AI revolution: Midjourney v5, ChatGPT 4, Stable Diffusion 2.2 XL tested
3 votes -
Megathread #4 for news/updates/discussion of AI chatbots and image generators
The hype continues. Here is the previous thread.
14 votes -
Megathread #3 for news/updates/discussion of AI chatbots and image generators
The hype continues. Here is the previous one.
14 votes -
Yann LeCun: From machine learning to autonomous intelligence
4 votes -
Once praised for its generous social safety net, Denmark now collects troves of data on welfare claimants
10 votes -
Robot learns to see in thirty minutes (2022)
3 votes -
We're all Wittgensteinians now
6 votes -
Another megathread for news/updates/discussion of ChatGPT and other AI chatbots
Hype is still going strong since the previous one.
9 votes -
A weapon to surpass Metal Gear
7 votes -
The shaky foundations of foundation models in healthcare
3 votes -
Fine-tuning to enable Stable Diffusion to generate very dark or light images easily
4 votes -
SolidGoldMagikarp and other words that cause buggy behavior with ChatGPT
18 votes -
Megathread for news/updates/discussion of ChatGPT and other AI chatbots
There's a lot of discussion out there and it doesn't seem to be dying down, so it seems like we should have a place for minor updates.
16 votes -
Whispers of AI’s modular future
6 votes -
How do we fix and update large language models?
6 votes -
Toolformer: Language models can teach themselves to use tools
11 votes -
ChatGPT and MidJourney made these drinks. Does the world even need me?
6 votes -
Google announces Bard, a ChatGPT competitor based on LaMDA
11 votes -
Five days in class with ChatGPT
13 votes -
Will Floating Point 8 Solve AI/ML Overhead?
6 votes -
ChatGPT mostly breaks the parts of the internet that are already broken
15 votes -
Infinite AI Array
3 votes -
Discovering Language Model Behaviors with Model-Written Evaluations
4 votes -
Medical selfies
5 votes -
Why Japan's internet is weirdly designed
8 votes -
Nvidia AI plays Minecraft, wins machine learning conference award
9 votes -
How DeviantArt is navigating the AI art minefield
10 votes -
Adversarial policies beat professional-level Go AIs
12 votes -
The Stack - permissively licensed code for large language models
6 votes -
Phenaki - generating videos from text with prompts that can change over time
6 votes -
The amazing power of "machine eyes"
6 votes -
Investigating toxicity changes of cross-community Redditors from two billion posts and comments
9 votes -
Introducing Whisper (OpenAI speech recognition model)
16 votes -
DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning
5 votes -
Prompt injection attacks against GPT-3
14 votes -
How Twitter’s child porn problem ruined its plans for an OnlyFans competitor
9 votes -
How to build a GPT-3 for science
5 votes -
Lexica - Search engine for images generated via stable diffusion
10 votes