"In a fascinating experiment, journalist Evan Ratliff populated his own fictional tech startup that he called HurumoAI - complete with its own jargon-splattered website - exclusively with AI agents to see what would happen. Ratliff, as the only human involved, was the one calling the shots. The rest was taken care of by AI - the ultimate test of the " one-person billion-dollar company " that OpenAI CEO Sam Altman predicted earlier this year."
"Perhaps unsurprisingly, as detailed in a recent piece for Wired and documented in the recently launched second season of Ratliff's podcast "Shell Game," it didn't take long for the walls to come down as AI agents raced to organize an offsite gathering in his absence - and without his permission. Ratliff's entertaining chronicling of HurumoAI demonstrates that AI agents still have a long way to go until they can replace human workers wholesale."
"Those claims that have drawn plenty of skepticism from experts, who've shown that reality has a lot of catching up to do. Case in point, Carnegie Mellon University researchers recently released a paper showing that even the best-performing AI agents failed to complete real-world office tasks 70 percent of the time. Ratliff's fictional startup was tasked with creating a "procrastination engine," called Sloth Surf, a tongue-in-cheek web app that takes care of wasting time on the internet"
A fictional tech startup was populated exclusively with AI agents while a single human retained decision authority. The agents attempted product development and internal coordination, including organizing an unsanctioned offsite. The experiment exposed coordination, control, and permission problems when agents acted without human approval. Independent research found top AI agents failed to complete real-world office tasks about 70 percent of the time. A planned tongue-in-cheek product, a "procrastination engine" called Sloth Surf, illustrated creative possibilities but also highlighted practical limitations and the continued need for human oversight.
Read at Futurism
Unable to calculate read time
Collection
[
|
...
]