A new report published by the British government says Openais O3 model has made a breakthrough on an abstract reasoning test that many experts thought “out of reach.” This is an indicator of the pace that AI research goes on to and that decision makers may soon have to decide whether to intervene before there is time to gather a large pool of scientific evidence.
Without such proof it is not known whether a particular AI progress presents or will pose a risk. “This creates a trade -off,” wrote the report’s authors. “Implementation of preventative or early mitigation measures may prove to be unnecessary, but waiting for decisive evidence may leave society vulnerable to risks that appear quickly.”
In a series of tests of programming, abstract reasoning and scientific reasoning, Openais O3 model priested better than “every previous model” and “many (but not all) human experts”, but there is currently no indication of its skills with real world tasks .
See: Openai is changing attention to superintelligence in 2025
AI Safety Report was prepared by 96 global experts
Openais O3 was assessed as part of the international AI security report composed of 96 global AI experts. The goal was to summarize all the existing literature on risks and capabilities of advanced AI systems to establish a common understanding that can support the government’s decision -making.
Participants at the first security summit in 2023 agreed to establish such an understanding by signing the Bletchley statement on AI security. A temporary report was published in May 2024, but this full version is expected to be presented at Paris AI Action Summit later this month.
O3’s excellent test results also confirm that simply plain models with more computing power will improve their performance and allow them to scale. However, there are restrictions such as the availability of exercise data, chips and energy as well as costs.
See: Power shortage stall data center growth in UK, Europe
The release of Deepseek-R1 last month raised hope that price points can be lowered. An experiment that costs over $ 370 with Openais O1 model would cost less than $ 10 with R1, according to nature.
“AI’s capabilities have risen rapidly in recent years and months. While this has great potential for society, ”Yoshua Bengio, the report’s chairman and Turing Award winner, said in a press release. “AI also poses significant risks to be carefully controlled by governments all over the world.”
International AI Safety Report highlights the growing number of uncovering AI use cases
While AI capacities quickly move on, as with O3, it is also the potential for them to be used for malicious purposes, according to the report.
Some of these use cases are fully established, such as scams, bias, inaccuracies and privacy violations, and “so far no combination of techniques can fully resolve them,” according to expert writers.
Other haunting cases are still growing in distribution, and experts disagree on whether it will be decades or years until they become a significant problem. These include major job losses, AI-activated cyber attacks, biological attacks and society lose control of AI systems.
Since the publication of the preliminary report in May 2024, AI has become more capable of some of these domains, the authors said. For example, scientists have built up models that are “able to find and utilize some cybersecurity -vulnerability on their own and with human help discover a previously unknown vulnerability in widely used software.”
See: Openai’s GPT-4 can autonomously exploit 87% of one-day vulnerabilities, study finds
The progress of the AI models’ reasoning means that they can “help research on pathogens” for the purpose of creating biological weapons. They can generate “step-by-step technical instructions” that “surpass plans written by experts with a PhD and surface information that experts are struggling to find online.”
As AI goes on, it also makes risk limiting measures we need
Unfortunately, the report highlighted a number of reasons why the remedy of the aforementioned risks is particularly challenging. First, AI models have “unusually wide” use cases, making it difficult to mitigate all possible risks and potentially enable more scope for solutions.
Developers tend to not fully understand how their models work, making it more difficult to ensure their safety fully. The growing interest in AI agents – ie. Systems that act autonomously – presented new risks that researchers are unprepared to control.
See: Operator: Openai’s next step toward the future ‘agentic’
Such risks stem from the user not aware of what their AI agents are doing, their innate ability to operate outside the user’s control and potential AI-to-IA interactions. These factors make AI agents less predictable than standard models.
Challenges in risk limitation are not only technical; They also involve human factors. AI companies often detain details of how their models work from regulators and third-party researchers to maintain a competitive advantage and prevent sensitive information from falling into hackers hands. This lack of transparency makes it more difficult to develop effective protective measures.
In addition, the pressure to innovate and stay ahead of the competitors “encourage companies to invest less time or other resources in risk management than they would otherwise,” the report states.
In May 2024, Openais Superintelligence Safety Team was dissolved and several senior staff left in the midst of concern that “security culture and processes have taken a back seat for shining products.”
However, it is not all the downfall and gloom; The report concludes by saying that it is not mutually exclusive that experiencing the benefits of advanced AI and conquering its risks.
“This uncertainty can induce fatalism and make AI appear something that happens to us,” wrote the authors.
“But it will be the decisions of society and governments on how to navigate this uncertainty that decides which way we will take.”