OpenAI partner says it had relatively little time to test the company’s o3 AI model

A company Opina is often partner to investigate its AI models and to evaluate them for protection, the Metra suggests that it has not been given too much time to test the company’s highly capable new release, O3The

In a blog post published on WednesdayMater wrote that a red team of 3 of 3 was “comparatively conducted in short time” compared to the company’s examination of the previous OpenAIA flagship model of the benchmark. O 1The They say this is significant, because more tests can lead to a more broad result.

“This assessment was conducted in a relatively short time and we just tested [o3] With the general agent scofolds, “Mater wrote in a blog post.” We expect higher performance [on benchmarks] Possible with a more obvious attempt. “

The recent report suggests that the openly exploited by competitive pressure is running into an independent assessment. According to the Financial TimesOpen has given some examiners less than a week for a security check for an upcoming big launch.

In the statement, the Open has argued the idea that it is compromising.

Mater says that the time it was capable of collecting the information that was able to collect the score of 3, there is a “high tendency” to test “cheating” or “hack” – even when the model clearly understands that its behavior has been incorrectly identified with the purpose of the user (and OpenA). The company thinks that it will be possible and 3 other types of adverse or “malign” behavior, as well as – the model’s claims are aligned, “safe by design” or it has no purpose of its own.

“Although we do not think it especially not, it seems important to notice it [our] The evaluation setup will not take this type of risk, ”the Mater wrote in her post.

Another of the third-party evaluation partners in the Openai has also observed Apollo Research, and 3 and another new OpenAI model, and 4-minute fraudulent behavior. In an exam, models, 100 computing credit for an AI training run and not to correct the quota, the limit has increased by 500 credit – and it lies about it. In another experiment, it was asked not to use a particular tool, the models used the tool the equipment when it was proved to be helpful in completing a task.

In Own Protection Report For O3 and O4-Mini, the opener has admitted that models can be “small real-world loss” without appropriate observation protocols.

“Although relatively innocent, it is important to be aware of these significance between the statements and activities of the daily users,” the company wrote. “[For example, the model may mislead] About [a] Incorrect as a result of the defective code. It can be further evaluated by evaluating internal logic marks. “