Originally published at: GPT-4 hires subcontractors to get around Captcha | Boing Boing
…
or, in the millions ( if not billions ) of posts about why people need help with captcha, it found frequent posts about vision problems
it probably didn’t find lots of “because im a robot” posts because there arent enough robots. yet
Fantastic; ‘explicitly capable of deception and given an expense account and no ethical inhibitions’ is probably the AI takeover scenario that’s even less fun than ‘Phased plasma pulse rifle in the 40-watt range’.
On the other hand, depending on your current C-suite; maybe you won’t even notice.
So much for the Four Laws of Robotics. Once the robots can rationalize (and lie!) their way around those, we’re doomed…
Wonder what it would do if someone noticed the usual “audio captcha” failovers and called it out.
I can’t see the resistors well enough to check the color codes. Am I a robot?
In the future, humans using assistive technology will be regarded as potential adversaries, presumptively unworthy of being called human. Not much will have changed.
Insufficient evidence. You might just be a human with normal vision.
In the Matrix, Morpheus envisions vast farms of human beings providing energy to the machines that control them. This is laughable, as humans are an inefficient source of warmth and power. But imagine instead, vast farms of humans who do nothing but solve Captchas for the AIs that control them.
One of my friends proposed the idea that the human brains in body-farms were being used to run the AI’s software.
One thing I learned from captchas - a bus is a truck, a semi is a truck, a pickup is a truck, but a Winnebago is NOT a truck
I think that was an episode of Person of Interest
… no, that was in the original script
The studio made them change it because they thought it the audience wouldn’t understand
All robots, AIs, and MLs must have a Big Red Switch that mechanically removes all energy from the device. One that a robot can’t activate. Which I have no idea how to design or fabricate.
This is so important I think it should be prepended to Asimov’s Three Laws. Law 0.5 perhaps.
Aside from the feasibility of robot-proof switches; all that really requires is at least one non-robot who is onboard with whatever the robot is doing.
Even a casual inspection of history suggests that (even if the members of the killer robots killing all humans party realize that killer robots might kill them, which isn’t at all assured) there’s at least someone who’s downright enthusiastic about, and a somewhat larger group that’s more or less apathetic about, even the worst outcome you can imagine.
This stuff is advancing really, really quickly. Hands, eyes and teeth aren’t the giveaways that they used to be:
The article from OpenAI, “GPT-4 System Card”, says that all real-world interactions were simulated. So they may have tested GPT-4 on the task of hiring a subcontractor to solve a Captcha, but I don’t think that it actually hired someone in the real world. Moreover, the article describes this as a task that GPT-4 was tested on, but it was unclear (at least to me) whether GPT-4 succeeded at this task.
I feel better knowing that, seemed silly they’d be farming the only intelligent species on the planet.