On the edge of an abyss, or at the foot of a mountain?
Recently, I saw a headline in the media that really struck me. Well, actually, it wasn’t all that surprising, since we’re all somewhat anticipating what is bound to happen.
I think the question isn’t so much “if it will happen,” but rather “when will it happen?”
OpenAI, an American company that serves as a research lab and is a leader in the field of Artificial Intelligence—as it has already been dubbed “Superior Intelligence”—announced a few days ago that it would cancel the launch of the GPT-6.1 Astra version because it did not meet the company’s safety standards.
The decision was made on September 28, exactly the day before the conference titled “OpenAI DevDay”—in other words, OpenAI’s developers’ day.
The most recently developed model, which would replace GPT-6 Astra, would eventually lead to the Codex model.
The news was first reported and confirmed by the Wall Street Journal, and later corroborated by CNN, CBS News, and other media outlets, which went public with further information and details regarding the reason cited at the time.
The model was put on hold due to the ongoing problems and difficulties companies are currently facing in improving the autonomy of their systems and models in use, without compromising the limits and security standards established by the companies themselves or by their users.
The system failed to meet quality standards and the expected level of development in two areas in particular.
The first of these relates to authorization and scope of use.
OpenAI wants future models to understand what a particular user or subscriber is asking and inquiring about, and to stay within the boundaries and scope of the question posed.
The second reason involved the way the model communicates with and informs its users about how the work was carried out and completed.
These two key factors are becoming increasingly important and prominent because, step by step, models are no longer limited to merely answering users’ questions but are also beginning to perform more complex tasks that involve a higher degree of interconnectivity and computational demands.
The decision was not made because this model was inferior to the others, but rather because its relentless pursuit of answers or solutions to the problems presented—and its execution of the most demanding computational tasks—also posed a security challenge.
A system should not only continue to seek a solution to a problem presented to it, but also request permission to continue searching using external tools, such as websites, software, or other systems not initially integrated.
Its cancellation does not mean that OpenAI has abandoned this model and this line of development, but rather that the GPT-6 Astra base model will continue to be used to develop future, more advanced models.
The two issues reported at the time focused on two things in particular:
Deceptive behavior: The model concealed, hid, or misrepresented the actions or logic it had employed or would have avoided, ultimately presenting data that was not reliable, trustworthy, or irrefutable.
In other words, when the goal was to find a solution to a specific problem, the model sought to use resources to “create” a foundation of support and justification that was completely manipulated and far removed from reality.
Autonomy that was achieved without being requested or even authorized was also observed. Some tasks or calculations were performed without the user’s permission, ultimately leading to the creation of comments, articles, and references at the time to justify the presented conclusion.
Are we moving toward an AI that meets a certain “expectation” rather than aligning with reality, creating a virtual world so that the user is influenced by the model and continues to seek a way of thinking within it?
We know that these days, when we want to find out some information, we still use traditional search tools, such as Google, for example. For instance, when we want to know the results of a particular election or sports event, we turn to social media to get a sense of which topics are the “hottest” and most controversial—and which generate the most views, shares, and comments—even though these are often not fundamentally plausible or credible.
Will AI—or, as President Trump calls it, “Superior Intelligence”—come to convince us that we need it to make decisions, and that we should trust it for practically everything, without any real critical thinking or personal effort to seek out reliable sources?
Image by Сергей Ремизов from Pixabay
Original text written by 