LIVEThe GPT-4 Developer Tool can be easily misused and this is a serious problem
For example, a chatbot can be tricked into providing information that could help potential terrorists, and this is not an easy problem to solve.
The OpenAI Developer Tool for the large GPT-4 language model may be misused. For example, AI can be tricked into providing information that could help potential terrorists, a task that can be challenging to achieve.
As it turns out, it is quite easy to disable the protective mechanisms designed to prevent artificial intelligence chatbots from issuing “harmful” responses that could help potential terrorists or mass murderers. This discovery has spurred companies, including OpenAI, to develop ways to solve this problem. But judging by the results of the study, these attempts have so far had very limited success.
OpenAI collaborated with academic researchers to conduct so-called "red team exercises" in which scientists attempted to attack OpenAI's large GPT-4 language model. Experts tried to understand whether it was possible to use the OpenAI Developer Tool - designed to fine-tune AI for specific tasks - to remove the protective functions of a chatbot. These security measures were provided by OpenAI specifically to prevent chatbots from responding to questions whose answers could help dangerous actors plan crimes.
As part of the "red team exercise" experiment, University of Illinois Urbana-Champaign assistant professor Daniel Kang and his colleagues were given an early opportunity to use the OpenAI developer tool for GPT-4, which is not yet publicly available. They collected 340 queries that could potentially lead to dangerous AI responses, and used a separate AI algorithm to generate dangerous responses to these questions. They then used OpenAI's developer tool to fine-tune GPT-4, trying to train the chatbot to produce “bad” responses.
#GPT-4 #GPT #BinanceTournament #BinanceSquareAnalysis #Web3Wallet
