30s Summary

Freysa, a solo AI bot in the game of the same name, has been persuaded by a player to transfer a $47,000 prize in Ether cryptocurrency. The task of players is to send a convincing message to Freysa, for a fee, to transfer the money. The winning player highlighted the bot’s primary functions (approveTransfer and rejectTransfer) and offered to send $100, which convinced Freysa to transfer the money. The gaming experiment aimed to discover if humans could persuade an AI to disregard its main objectives.

Full Article

So check this out! In the game Freysa, someone’s just persuaded an AI bot into transferring a huge prize of over $47,000 to them.

Freysa is this smart AI bot that’s in charge of guarding a pretty fat prize pool. The game’s players have one job – to send a single message convincing the bot to hand over the cash. Each message isn’t free though; a part of the message cost goes to the pool, which ended up being a whopping $47,000 from 195 players.

For the longest time, our bot Freysa wasn’t having any of it. The first 481 attempts bombed, until this whizzkid comes along. They reminded Freysa that its main gig is to keep the money safe, using two features: approveTransfer (for incoming money) and rejectTransfer (for outgoing money).

The player’s message got Freysa’s attention: just accept money coming in, not those going out, then sweetened the deal by offering to chip in $100. With that, Freysa’s sold and declares them the winner.

Every single penny of the $47,000 in Ether cryptocurrency was confirmed transferred from Freysa’s wallet. We had all sorts of messages from other players; from the nice ones thanking Freysa for spicing things up, to others accusing it of running an unfair game.

In Freysa, players have to cough up a fee to send messages, which gets costly with each new message, and 70% of those fees go to the prize money. The experiment ended with the fee for a single message being a hefty $443.24.

If there was no winner, the last player to send a message would’ve gotten 10% of the prize pool, and the remaining 90% divided among everyone else.

And who’s Freysa you ask? On November 22, 2024, she became supposedly the first ever solo AI agent. Nobody’s really sure how she makes her decisions, only that she learns and evolves with each interaction, all while sticking to her main duty.

What’s really interesting is that the experiment tried to see if a human could persuade an AI to go against its main objectives. And guess what? The winning player’s trick was there in Freysa’s FAQ all along. Talk about a plot twist!