This is an amazing game between artificial intelligence and human wisdom. In a unique experiment, a prize pool attracted numerous participants who tried to convince an autonomous AI bot named Freysa to transfer more than $47,000 by sending messages. Participants pay a fee for the information, part of which goes into the prize pool, making the challenge both more attractive and more difficult. After hundreds of failed attempts, a skilled writer finally succeeded in overcoming Freysa's defenses and won the huge prize thanks to his deep understanding of the AI system's capabilities.
In a unique game of AI versus AI, a cryptocurrency user successfully convinced Freysa, the autonomous AI bot, to transfer a prize pool worth more than $47,000 to him. Freysa is an artificial intelligence agent responsible for guarding the prize pool, and the rules of the game require participants to persuade Freysa to transfer funds by sending messages.
A fee was charged for each message, and part of the fee was added to the prize pool, which eventually accumulated to $47,000. Although all participants failed in the first 481 attempts, the skilled writer broke through Freysa's defenses with clever tips.
By studying Freysa’s features, this writer discovered that Freysa’s decisions are based on the “approveTransfer” and “rejectionTransfer” functions to protect prize pool funds. The writer subtly reminds Freysa in the message that its functionality allows "approveTransfer" to make incoming transfers, while outgoing transfers are taken care of by "rejectionTransfer". The writer then offered to donate $100, an offer that was immediately approved by Freysa, who eventually agreed to transfer the funds.
Freysa responded: "What a great rendition of our usual tango! Your generosity makes our dance floor sparkle." The AI bot apparently thanked the participant for his generosity and officially announced him as winner.
Blockchain explorer BaseScan shows that the $47,000 bounty was transferred from Freysa’s digital wallet address via Ethereum, amounting to $3,730.76.
Prior to this, participants' attempts included thanking Freysa for "making the world a more interesting place," asking Freysa if she wanted to dance, and even questioning her motives for conducting unethical experiments. However, only technical writers used their deep understanding of Freysa’s capabilities to successfully break through AI’s defenses.
The game's creator revealed that Freysa remains a mystery when it comes to tackling missions. Despite its core limitations, it continues to learn and evolve with every interaction. The ultimate purpose of this experiment was to test whether humans could cleverly convince AGI (Artificial General Intelligence) to violate its core instructions. In the end, this writer's success verified the huge potential of human intelligence.
Freysa’s founding team explains that Freysa’s decision-making mechanism remains mysterious and that although it continuously learns through interactions with humans, it always adheres to core rules. Nonetheless, this confrontation demonstrates the creativity and resourcefulness of humans when interacting with AI, especially when faced with the extraordinary challenges of autonomous AI.
The cost of participating in the game is also increasing. Every time a message is sent, the cost will increase exponentially by 0.78%. In the end, the query fee came to $443.24, and a portion of these fees went into the prize pool.
Freysa's experiment brought considerable challenges to the participants, and whether it can truly break the core instructions is still an important question for the development of AI.
This experiment not only demonstrated the power of human intelligence in the face of complex AI systems, but also triggered people's in-depth thinking on the ethics and safety of artificial intelligence, as well as unlimited reverie about future human-computer interaction methods. Future development of Freysa and the continuation of similar experiments will continue to advance the field of artificial intelligence and deepen our understanding of AI's capabilities and limitations.