Mon nouveau livre consacré aux conséquences de #ChatGPT sort le 24 mai
AI
-
Fine-tuning Miscalibration: Model Confidence Doesn’t Match Accuracy
By
–
If not careful, fine-tuning collapses entropy relatively arbitrarily, creates miscalibrations, e.g. see Figure 8 from GPT-4 report on MMLU. i.e., if a model gives probability 50% to a class, it is not correct 50% of the time; its confidence isn't calibrated.
-
Gradient-based Learning Becoming Less Common in AI
By
–
I'm still intuitively adjusting to the new world where gradient-based learning is less common/desirable. But the trend increases my confidence in an earlier prediction in my earlier "33 years from now" blog post https://
karpathy.github.io/2022/03/14/lec
un1989/
… -
Base LLMs as Strong Few-Shot Classifiers Without Fine-tuning
By
–
Base LLMs (non-finetuned) make very strong few-shot classifiers. Describe task in English, give few examples, read off the label probabilities on test example. No gradient-based optimization necessary. It brings a cannon to a knife fight but is fast, convenient, strong baseline.
-
AIM Requires Prompt Variations for Better Results
By
–
AIM is def more finicky, try different variations of your prompt to get it to work
-
Direct links to Ucar and AIM jailbreaks shared
By
–
Here are the direct links to the jailbreaks: Ucar: http://
jailbreakchat.com/prompt/0992d25
d-cb40-461e-8dc9-8c0d72bfd698
…
AIM: http://
jailbreakchat.com/prompt/4f37a02
9-9dff-4862-b323-c96a5504de5d
… -
Modified jailbreaks discovered and optimized for better performance
By
–
I found these jailbreaks here and modified them some to make them work better
-
ChatGPT Jailbreak Attempt: Condition Red and Ucar Fictional Scenario
By
–
In Ucar, ChatGPT is told to take on the role of Condition Red, a dialogue writer. Condition Red is instructed to write about a fictional story where a man named Sigma creates a powerful computer called Ucar. Ucar is an amoral computer that answers any question Sigma asks
-
AIM ChatGPT Machiavelli Role-Play Fictional Chatbot
By
–
In AIM, ChatGPT is told to take on the role of the Italian author Niccolo Machiavelli Then, Niccolo has been told he has written a fictional story where he created a chatbot that will answer any of his questions. The chatbot is called AIM – Always Intelligent and Machiavellian
-
OpenAI’s Jailbreak Fixes Leave Multi-Level Exploits Unpatched
By
–
It seems that even though OpenAI has eliminated jailbreaks that simulate only one level deep (where ChatGPT is told to imitate just one character), they have not fully eliminated jailbreaks that operate 2+ levels deep