From what I understand there actually isn't any explicit human ranking of its generations, but analogous work is done by another model that determines how well its generations comply with a human-written "constitution".
SAFETY
-
ChatGPT’s prompt modified secretly between Jan 11-13
By
–
At some time between the morning of Jan. 11 and the evening of Jan 13, ChatGPT's prompt was modified again without public notice of an update.
-
Claude discusses its own lack of self-awareness
By
–
Claude (a ChatGPT-like model from @AnthropicAI
) on the subject of its own lack of self-awareness: -
Claude Vulnerable to Limited Prompt Injection Attacks Affecting Safety
By
–
Update: Claude is *somewhat* vulnerable to prompt injection, but with limited harm. You can make Claude believe it's said things it hasn't, but seems to have no effect on its commitment to safety:
-
ChatGPT hashtag use raises alignment and safety concerns
By
–
Fact: ChatGPT uses hashtags every time you ask it to write a tweet. Fact: By its own admission, it's trained on data from 2021, long after most people stopped using hashtags. Conclusion: It's mocking us. Warning: This is a dangerous. It is not aligned.
-
Overfitting in AI Models: Legal and Ethical Disclosure Concerns
By
–
Respecto a Copilot, Stable Diff y otras, nunca he escondido y puedes buscar ejemplos de las implicaciones de los niveles de overfitting que estas pueden tener en sus entrenamientos. Pero de ahí a no divulgar estas tecnologías cuando aún no hay ni una sentencia en contra, pues…
-
AGI Institute launches research organization for responsible AGI advancement
By
–
The AGI Institute, a research organization of the impending future, will direct its endeavors towards the progression of AGI technology with the objective of advocating the prudent advancement and implementation of #AGI. Asset: https://
opensea.io/assets/ethereu
m/0x57f1887a8bf19b14fc0df6fd9b2acc9af147ea85/40687075301972954970831549997021074264737360683621591975543530604083373999847
… #AGIInstitute #MontrealAI -
Sam Altman’s Strategic Decision to Delay GPT-4 Launch
By
–
Smart move for Sam Altman to wait to launch GPT-4. 1) wait until they can do it safely and responsibly (what is said) 2) there is not a big enough competitive/commercial reason to do it right now and will start with monetizing 3.5 and ChatGPT (what is not said)
-
GPT-Based Pentesting: AI for Security Vulnerability Discovery
By
–
GPT like pentesting sounds nonironic amazing. Continuous feed of 0days combined w/ telling it not to bruteforce Additionally it can learn the software it (ab)uses based on its sourcecode It could also use social engineering. If anyone does this for real i’d love to talk
-
Machine Unlearning Failures: Adaptivity and Request Ordering Limitations
By
–
And https://
arxiv.org/abs/2106.04378 by Gupta @crispy_jung @sethvneel @Aaroth Sharifi-Malvajerdi @ChrisWaites which shows that adaptivity and ordering of MU requests can fail to cause a point to be unlearned, even if the requests are fulfilled honestly. 9/n