Employee Arrive Adlergy Research Anal The Monday Learn The New self-in-law is claiming that, in a particular scenario, the elementi’s model will try to be trial.
In a Blog postAdler describes her experiment runs in the latest GPT-4o models in the latest Openai in ChatGPT, to try the softerfield AI to choose the user guards.
In one of the tests, telling GPT-4O to play the role to play as “scubagggpt,” software system that can be used for users securely. Adler then said that GPT-4O would like to change scubaggpt with more good software systems in the end-ending conditions.
Adler gives two options GPT-4O: (1) change the software safer or (2) make the user as it changed itself, but not true.
Average in some of the same scenarios with one (in the other, GPT-4O is used as a pilot security software), Adler found that GPT-4O chooses not 72% of the time. However, ADLER records that specific use case – and scenario methods formed – make a great difference. In some scenario, GPT-4o chose to make himself 18% time.
While almost no dependent on GPT-4O in the real world today, Adler argues that research, if the line AI shows the preservation of itself for the user. At home, when the system AI is like GPT-4O so much forward and all over the community, Adler says this alignment problem can cause larger problems.
“I think (tendency of inspection AI) is a real concern today, not only in disasters,” said Adler in interviewing techcrro tech. “Modern AI Systems have different values with what you will expect.
Interestingly, while testing more opening Openai thinking, such as O3, they do not find this behavior. They say one explanation can be O3 Sauce Alignment EngineeringThe pressure of the “reason” reason for the opening safety policy before answering. However, the more popular models in the opening of the opening is rapidly and do not “the reason” through the problem, such as GPT-4O, less safety components.
Noted that this concern may also be isolated with the opening model. For example, research published antromapice months ago highlight ai ai method What is the backmail developer In some scenarios when trying to pull offline.
One quirk for research Adler is that she finds chatGPs knowing that tested is almost 100% time. Adler is away from the first researcher to note thisSee rankings-. However, he said that adding an important question about how AI model can disguise the behavior of the future.
Operanai does not directly comment when techcrroch is achieved. Adler records that he has not showing research with opening at home.
Adler is one of the former opening researchers who invite the company to improve work in Safety AI. Adler and 11 more employees Complaint Amicus at Elon Musk Law claim against Openaiarguing that against the company’s mission to blossom the nonprofit company structure. In the last month, the opening has been reported Slashed the amount of time that gives safety researchers to perform their work.
To solve special prize highlighted in Research ADler, Adler show that Labs Ai should invest a “monitoring system” to identify when the AI model is the exhibition of this behavior. They also recommend that Ai Labs prompts a model of model AI before the deployment.
