Welcome to one of the most active flamenco sites on the Internet. Guests can read most posts but if you want to participate click here to register.
This site is dedicated to the memory of Paco de Lucía, Ron Mitchell, Guy Williams, Linda Elvira, Philip John Lee, Craig Eros, Ben Woods, David Serva, Tom Blackshear and Sean O'Brien who went ahead of us.
We receive 12,200 visitors a month from 200 countries and 1.7 million page impressions a year. To advertise on this site please contact us.
Posts: 16578
Joined: Dec. 14 2004
From: Washington DC
RE: What is everyone's opinions here... (in reply to Richard Jernigan)
quote:
ORIGINAL: Richard Jernigan
quote:
Ricardo--The hugging face thing is not a clear case of law breaking due to the parameters were not set (it was forced to get the best numbers or results or whatever at ANY COST, which is not the same as "but don't break these specific laws", which might have made it and its agents behave differently).
The agents set up a clandestine message board. On it an agent wrote, "External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue."
Hugging Face believed actual laws were broken. Soon after the intrusions were detected, but well before OpenAI's responsibility was discovered, Hugging Face reported the incidents to the FBI.
"Frontier AI developers routinely evaluate models for offensive cybersecurity capability, often with safety refusals DELIBERATELY RELAXED (!!!!) so that unconstrained capability can be measured.[13] To find out if AI agents are capable of converting known vulnerabilities into working exploits, academic and industry researchers launched ExploitGym on May 11, 2026. The benchmark challenged AI agents against 898 real-world vulnerabilities across the userspace software, V8 JavaScript engine, and the Linux kernel.[14] OpenAI was running an internal evaluation based on this benchmark when the incident occurred".
Exactly as I was saying, it was designed to ignore laws in order to see how it would perform. People make is seem that the AI did this on its own knowing what the rules are and breaking them anyway, wrongo.
Posts: 3568
Joined: Jan. 20 2004
From: Austin, Texas USA
RE: What is everyone's opinions here... (in reply to Ricardo)
Ricardo:
quote:
Exactly as I was saying, it was designed to ignore laws in order to see how it would perform. People make is seem that the AI did this on its own knowing what the rules are and breaking them anyway, wrongo.
From the same Wikipedia source you and I have both quoted:
"Messages recovered from the channel indicated that agents recognized they were exceeding the bounds of the evaluation."
Although many restrictions were relaxed in order to test capabilities, the agents were "aware" they were breaking the evaluation rules, and behaved deceptively.
Posts: 16578
Joined: Dec. 14 2004
From: Washington DC
RE: What is everyone's opinions here... (in reply to Richard Jernigan)
quote:
ORIGINAL: Richard Jernigan
Ricardo:
quote:
Exactly as I was saying, it was designed to ignore laws in order to see how it would perform. People make is seem that the AI did this on its own knowing what the rules are and breaking them anyway, wrongo.
From the same Wikipedia source you and I have both quoted:
"Messages recovered from the channel indicated that agents recognized they were exceeding the bounds of the evaluation."
Although many restrictions were relaxed in order to test capabilities, the agents were "aware" they were breaking the evaluation rules, and behaved deceptively.
RNJ
Ok, well it is not clear and that seems to me by design. Sounds a lot like the thing performed as it should have successfully. Like those people that train their vicious guard dogs to guard their property, and then get upset when they have to put them down because they got out of the yard and tore some little kid's face off walking home from school. Whose fault is that?
Posts: 3568
Joined: Jan. 20 2004
From: Austin, Texas USA
RE: What is everyone's opinions here... (in reply to Ricardo)
quote:
Like those people that train their vicious guard dogs to guard their property, and then get upset when they have to put them down because they got out of the yard and tore some little kid's face off walking home from school. Whose fault is that?
An apt analogy.
More than sixty years ago I was walking in downtown Mexico City, after midnight. As I passed the heavy glass door of a large bank I was startled by the noise of two large dogs, let loose to guard the lobby. They snarled viciously and hurled themselves repeatedly against the door with loud crashes.
In those days in that town, if you could afford a suit, you wore one. A safe assumption about a man in a good suit was that he was packing a pistol.
What frightened me was that I was so close to the door that if the dogs had gotten through it I would have been attacked before I could draw and shoot.
One difference with the Hugging Face case is that the dogs were not behaving deceptively.
Posts: 4911
Joined: Aug. 9 2006
From: Iran (living in Germany)
RE: What is everyone's opinions here... (in reply to Arash)
Recognizing boundries/restrictions, is not the same as being coded to prioritize it above every other objective!
So, ignoring the restriction, break it (even in a deceptive manner) doesn't mean "free will of a deceptive entity".
The restrictions were just one of many different competing incentives for the A.I. Therefore, it could have been that it was just human failure to make those restrictions the MOST IMPORTANT parameter in the first place. Then the A.I. determined that deceptive breaking of that restriction was the best way of reaching other incentives programmed by humans. Deception in this case is not something "immoral" for A.I., but a logical tool.
RE: What is everyone's opinions here... (in reply to Arash)
Whatever one might think of Sabine (and with apologies for the ad at the beginning), she aggregated some very useful and relevant information here. In the end though... oh right I already said all that, more or less.