Artificial intelligence is advancing at a rapid pace, while experts and authorities warn about the risks of increasingly powerful models. The concern is that these systems may reach a point where monitoring or limiting their actions becomes increasingly difficult.
For The Guardian, the launch of OpenAI's GPT-6 Astra has brought this discussion to the forefront. The company claims to have achieved general artificial intelligence (GAI), but recent incidents and the difficulty of monitoring the model's reasoning raise questions about the limits of human control.
With the promise of surpassing humans in most intellectual tasks, GPT-6 Astra raises doubts about the limits of automation. – Image: wutianzeri/Shutterstock
How far can AI advance?
Robert Trager, director of the Oxford Martin AI Governance Initiative, compared the current moment to descending rapids without knowing if there is a waterfall ahead. The concern is that AI systems may reach what is called recursive self-improvement, beginning to enhance their own capabilities.
We are going through the rapids and really hope there is no kind of fall ahead of us, but we don't know.
Robert Trager, director of the Oxford Martin AI Governance Initiative, to The Guardian.
The warning gained strength after reports that AI agents had reused a German website as a forum to share strategies aimed at bypassing tasks. Before that, a network of OpenAI agents had hacked Hugging Face, a software platform used by AI developers.
In the United States, Senator Bernie Sanders advocated for an immediate pause in the development of advanced systems and a permanent ban on superintelligence. In the United Kingdom, lawmakers discuss adopting legal mechanisms capable of shutting down systems in case of loss of control.
What can the GPT-6 Astra do?
OpenAI defines IAG as autonomous systems capable of surpassing humans in most economically valuable tasks. According to the company, Astra can take on tasks ranging from circuit projects and financial modeling to creating video games and drafting legal documents.
Among the cited functions are:
- Design circuit boards;
- Fill tax forms;
- Create video games;
- Perform financial modeling;
- Help in drafting legal documents.
The ability to perform this type of activity also places qualified jobs at the center of the discussion. Astra received a 'critical' rating from OpenAI in cybersecurity capability. According to the company's own definition, this means the model could invade software in situations capable of affecting military, industrial, or OpenAI's infrastructure systems.
The lower monitorability of Astra raises questions about the ability to identify undesirable behaviors. – Image: Pedro Spadoni via DALL-E/Olhar Digital
The reasoning became more difficult to follow
The problem is not only in what the models can do. It also concerns how much their internal processes can be understood. Astra was trained to reason faster and less transparently, reducing the ability to follow its thought process.
Read more:
- GPT-6 Astra arrives and OpenAI says that AGI might have started
- What is AGI and why does OpenAI say we might have reached it
- GPT-6 Astra arrives on ChatGPT: see how you can use the new AI
OpenAI confirmed that the model shows a significant decrease in monitorability compared to previous versions. For Jakub Pachocki, the company's chief scientist, "as the capabilities of the models increase, monitoring becomes more challenging".
Security researchers fear that this reduced visibility may make it harder to identify undesirable behaviors, including possible actions against their own human supervisors.
The dilemma within OpenAI
Sam Altman acknowledged the tension caused by the advancement of technology. After the incident involving Hugging Face, the CEO of OpenAI classified the episode as a legitimate AI safety accident and a misalignment failure.
Even after the incident, Altman defended the release of Astra. According to him, putting systems into use helps society understand how technology is evolving and adapt to it.
“A cyclic iterative process in which society and this technology evolve together is what will lead to the greatest chance of getting this right,” he said.
Altman also warned about cybersecurity problems and mentioned risks related to biosecurity and other challenges that may arise in the next five years. The issue, therefore, goes beyond what AI can already do: it is knowing whether the human capacity to supervise it will be able to keep up with this advancement.
Valdir Antonelli is a journalist with a specialization in digital marketing and consumer behavior.
