While the world’s major intelligence agencies are racing to test the capabilities of obstetric artificial intelligence models, sharp questions are waving on the horizon: Who will settle the race? America, which has the finest technology? Or China, which shows a faster embroidery?
On the day that Donald Trump was inaugurated as President of the United States, a Chinese company called “Deepseek” launched a major linguistic model (LLM) that won global recognition. The event was not crossed in the corridors of the American intelligence, as Trump described it as a “alarm bell”, while Senator Mark Warner, Vice -Chairman of the Senate Intelligence Committee, approved that the American intelligence community, which includes 18 agencies, “was completely surprised”, according to the “Economist” magazine.
With fears of Chinese superiority in this regard, the administration of former President Joe Biden launched a plan to enhance the experiences of state institutions with the most advanced models.
Direct directives were issued to intelligence agencies, Pentagon and the Ministry of Energy, to urge them to expand their cooperation with the leading artificial intelligence laboratories, led by “Openai”, “DeepMind”, “Google” and “Anthropic”.
In a remarkable step, the Pentagon granted in mid -July decades of $ 200 million to four companies, including Xai, affiliated with Elon Musk, to test what is known as “smart agent models”; They are models that are able to analyze complex tasks, divide them into stages, and implement them automatically, to control devices such as computers, or vehicles.
While the field is experiencing an increasing momentum, clear signs of artificial intelligence appear to enter a new stage in intelligence work environments. Last January, Microsoft announced that 26 of its cloud services became accredited to American security use. In June, the company “Anthropic” revealed a model for government use, named “Claude Gov”, and it has already begun to be used in higher agencies within the American national security system.
But these models are not used as they are, as companies subjected to accurate adjustments that enable them to overcome public safety restrictions, and train them on languages and dialects for governments, as well as operate them within closed and reserved environments. In some cases, agencies to build their own models began relying on internal security infrastructure.
Europe enters the line
The movement is not limited to Washington and Beijing. In Europe, efforts are accelerated to reduce the gap. A senior British source said that the UK intelligence services have access to very advanced models, designed to deal with very secret information.
The French company «Mistral», the official partner of the French Military Intelligence Agency, is the most prominent players of the continent in this field. Its model is characterized by its superior ability to understand the Arabic language, and regional languages such as Tamiliya, as a result of his training in data from the Middle East and South Asia.
As for Israel, it has revealed that the Israeli army’s use of the “Openai” model “GPT-4” is increased by 20 times after the outbreak of the Gaza war.
Slowen or rational caution?
Despite the accelerated moves, some voices indicate the slow adoption of artificial intelligence in national security institutions. Catherina Malligan, a former defense and intelligence official and is currently the head of security partnerships at Openai, says that use is still without aspirations. While the National Security Agency (NSA) is an exception thanks to its accumulated experience in techniques such as sound recognition.
Tarun Chapra, director of national security policies at “Anthropic”, says that the real shift lies in reformulating the methodology of carrying out tasks, not just using smart chat interfaces.
The risks behind the smart agent
On the other hand, researchers are warned against excessive optimism. Dr. Richard Carter, from the British Alan Torring Institute, confirms that the intelligence services need tools that achieve consistency, reliability and transparency. Therefore, UK agencies are focused on RAG enhanced generation techniques (RAG) to reduce errors.
It is believed that the company «Mistral» offered to some parties an experience of intelligent agents specializing in analyzing certain types of data, from satellite images to audio communications, which speeds up response decisions. But Carter warns that these models may produce serious cumulative errors due to their repeated self -mechanisms, pointing to an evaluation conducted by “Openai” that showed that the “Chatgpt Agent” model suffers from a hallucinations of an average of 8 percent.
China in the background … monitor and advance?
While American companies are progressing in developing models, China is increasingly concerned about its adoption speed. “We don’t know much about the extent of China’s use of models such as (Deepseek) in the military fields,” says Philip Riner, of the Silicon Valley Institute of Security and Technology. “They may be able to achieve faster results, because they are not subject to the same regulatory controls.”
In an attempt to remedy the matter, the Trump administration issued directives to the Pentagon and intelligence agencies to conduct an urgent evaluation of the speedy adoption of these technologies compared to China and a plan for continuous adaptation.
Experts agree that delay in adoption may be more dangerous than delay in development. “We were not effective enough to monitor China’s progress,” says Senator Warner. As for Moligan, she adds, “The real danger is not a failure to develop artificial intelligence, but to lose the race to use it.”
0 Comments