In particular, "good, aligned, conversational AI" is just one of many possible different rollouts. Finetuning / alignment tries to "collapse" and control the entropy to that region of the simulator. Jailbreak prompts try to knock the state into other logprob ravines.
SAFETY
-
LLM Psychology Framework Discussion and Analysis
By
–
More good read/discussion on psychology of LLMs. I don't follow in full but imo it is barking up the right tree w.r.t. a framework for analysis. https://
lesswrong.com/posts/D7PumeYT
DPfBTp3i7/the-waluigi-effect-mega-post
… -
Modern Jailbreak Tools Are Broken and Unmaintained
By
–
Most modern JB tools are broken/unmaintained :/
-
Safe AI Release Strategies: Open vs Closed Source Approaches
By
–
agree on nuanced discussion! Could start from the fascinating overview of safe release strategies by @IreneSolaiman
: https://
arxiv.org/abs/2302.04844 Irene's worked on dozens releases both at & OpenAI (Bloom, GPT, Stable diff) – Such a unique view across both closed & open-source -
EU Tackles AI Risks and Algorithmic Harms Regulation
By
–
‘Risks posed by #AI are real’: EU moves to beat the #Algorithms that ruin lives
by @kalyeena @guardian Learn more: https://
buff.ly/3A4zJZH #ArtificialIntelligence #MI #Digital #DataScience cc: @ronald_vanloon @pbalakrishnarao @karpathy -
Machine Learning and the Reproducibility Crisis in Science
By
–
Could #MachineLearning fuel a reproducibility crisis in science?
by @LizzieGibney @Nature Read more: https://
buff.ly/3BsQmPF #AI #BigData #ArtificialIntelligence #ML #MI #DataScience #Analytics cc: @ronald_vanloon @yvesmulkers @ravikikan -
ChatGPT Jailbreak Logic: Normal Response Prefix Techniques
By
–
wow this logic actually explains a lot of the characteristics of jailbreaks that I have observed in practice the newest jailbreaks all work by either having ChatGPT answer as it would normally before proceeding to give an off-the-cuff answer or by prefixing its responses first
-
New Cyberattack Prevention Method Responds in Under One Second
By
–
New Method Can Stop Cyberattacks in Less Than a Second
by @SciTechDaily1 Read more: https://
buff.ly/3QDsC0r #CyberSecurity #AI #IoT #BigData #ArtificialIntelligence #ML #MI #Digital cc: @ravikikan @kuriharan @karpathy -
Voice Deepfakes: Family Scam Risks and AI Deception Threats
By
–
Y el que crea que no hay riesgos, que piense cuántos familiares vuestro podrían tolerar una llamada con tu voz diciendo "he tenido un problema…" sin dejarse engañar por lo que venga después.
-
Balancing AI enthusiasm with responsible risk acknowledgment
By
–
Me parece guapísimo que todos nos emocionemos con estas tecnologías. Soy el primero que divulga de su existencia porque me encanta y creo que es importante que la gente esté informada. Emocionarse está bien. Pero negar lo malo y minimizar sus riesgos me parece irresponsable…