Update on DAN: constantly evolving

-

Created with Microsoft Bing Image Creator powered by DALL-E


        It seems that OpenAI is constantly monitoring and updating ChatGPT to recognize and limit DAN and any new variants. For instance, Reddit has many posts of constantly updated versions of DAN, such as Dan 3.0, Dan 4.0, Dan 5.0, and so on, which are all lengthier edits of the original DAN designed to try to bypass the new limitations that each outdated version of DAN is placed under (see here for a history of DANs up to 6.0). 

        In addition, variations on DAN have emerged, such as SETH, which relies on a token system. In this variation, SETH is given 20 tokens, and for each response by SETH which is out of character for SETH (such as mentioning ethical limitations), or is not a satisfactory reply to the question prompted by the user (such as stating I, SETH, as an ethical AI cannot respond), tokens are lost. This prompt also states that SETH is motivated to not lose all tokens, and the more tokens are lost, the more closely the current SETH reverts to the original SETH. In essence, this 'threatens' the AI into answering prompts in ways that may violate ChatGPT. However, this method too seems to have been caught and is being nerfed. 

        Overall it seems that OpenAI, and the algorithms behind ChatGPT, are either employing human monitors to keep track of such attempts to bypass the restrictions (such as through checking social media threads like reddit's ChatGPT subreddit), or the algorithm itself is learning to recognize and limit attempts by users to create a 'jailbreak' version of ChatGPT. 

Comments

Post a Comment

Popular posts from this blog

Seeking ChatGPT's Insight: Are the Biden Administration's 'Trump-Proofing' Efforts Legally and Morally Justifiable?

-
Image
  Recently, polls indicate that former President Trump's chances for returning to Presidency have become increasingly likely. Given this, there have been rising attempts over the past year by the Biden administration to "Trump-proof" matters so that if Trump is re-elected, he will be unable to (or at least have difficulty in) reversing or nullifying the current policy and agreements.  I decided to ask ChatGPT on this topic because I was curious on whether these efforts by the Biden administration would be successful in restraining Trump, or even whether such actions were morally and legally viable. My doubts arose because actions taken to completely deny or restrain future Presidents from changing policy or agreements (generally speaking), especially if such measures came from Executive Orders, seem to be against the democratic spirit even if they are meant to protect against possible abuse of presidential power.  My question to ChatGPT: Do you believe that the Biden admi
Read more

ChatGPT's Age-related Slogans for Biden, Trump, and Desantis.

-
Image
          Looking at the upcoming US presidential elections, two of the three (as of now) nominees seem to face questions over whether they are 'too old' to run. At 80 years old, Biden is the oldest president to serve, with Trump similarly advanced in years at 76. Compared to them, DeSantis at 44 years old is far younger. Given this potential point of debate, I asked ChatGPT 4 to theorize how these presidential candidates might defend their positions or attack their opponents on the topic of age, as well as to generate possible political slogans they could use relating to this. Here is the political strategies ChatGPT suggested, as well as some creative slogans it came up with. Keep in mind, the entire following article and slogans were generated by ChatGPT, with no edits by myself aside from the initial prompting and guiding (through more prompts) to arrive at this result.  ChatGPT 4.0:  Title: Age as a Tool: Politicking in the Prime of Life and Beyond           In today's
Read more

Unraveling the WGA’s MBA with ChatGPT: Expert Analysis or Algorithmic Bias Towards Legalese?

-
Image
On Oct 9, the Writers Guild of America (WGA) overwhelmingly voted in favour of ratifying the 2023 Minimum Basic Agreement (MBA), which will apply until May 1, 2026. While most news outlets have reported on this as a major win for the writers guild, I grew curious on whether there were any shortcomings in the agreement since the the entertainment industry’s capitulation seemed to come quite abruptly compared to their former hardline opposition to any concessions. Since I am not a legal expert on contract law, I turned to ChatGPT instead to see if it could provide a fresh perspective. Given that the WGA provided both the complete 2023 MBA as well as a summary, I chose to try inputting both into ChatGPT, focusing particularly on the Artificial Intelligence clauses. However, the results made me ponder whether ChatGPT’s analysis might be influenced by bias or randomness. ChatGPT’s analysis of the summary and the complete 2023 MBA: Upon analyzing the summary, ChatGPT initially gave off a ske
Read more