Give air supremacy to AI, and the United States will have no choice

虎嗅APP

2024-05-10 19:29Posted on the official account of Beijing Tiger Sniff APP

头图｜Edwards Air Force Base

Produced by Tiger Sniff Technology Group

Author: Qi Jian

Editor|Wang Yipeng

Just over a year ago, in May 2023, OpenAI's CEO Sam Altman had mentioned the fact that AI was being weaponized at a congressional hearing and endorsed the proposal to "regulate AI like nuclear weapons."

Now, the U.S. military is in a hurry to hand over "air supremacy" to AI.

On May 2, U.S. Secretary of the Air Force Frank Kendall personally experienced the AI-piloted X-62A VISTA fighter at Edwards Air Force Base in California, USA.

Prior to this, the U.S. Defense Advanced Research Projects Agency (DARPA) had already conducted the first real-world simulation of the AI fighter X-62A VISTA against a human-piloted F-16 on April 20.

An AI-piloted F-16 fighter jet (left) flies side-by-side with a human-piloted F-16, with the two planes less than 304 meters apart from each other

The X-62A VISTA is a facelifted fighter of the F-16, with more than 4,600 F-16 fighters currently in service around the world, and is marketed in more than 20 countries among the allies of the United States.

Aren't Americans very worried about the safety of AI, why are they in such a hurry to let AI fly fighter jets?

Not enough equipment, AI to make up?

"We're taking the technical advantage and that's the key to the game." Kendall said in an interview that opponents who do not develop AI weapons will "pay the price" in real combat.

As for the planning of the AI fighter, Kendall believes that the AI fighter can be placed in some deliberately sacrificed places to "find out where the enemy is or attract firepower".

US Air Force Secretary Frank Kendall "test drives" AI fighters

Attracting fire is still relatively easy to understand, but can't you find enemies without AI?

In the matter of seeking the enemy, the main method has always been to rely on radar. In the air combat of fourth- and fifth-generation fighters, the use of radar to detect the enemy will be affected by two factors: the first is the limited detectable range of radar, and the second is the popularization of stealth fighters.

The detection range of the radar determines who detects the other side first. The basic logic of modern warfare is to discover and destroy, and even more so in air combat. In the vast sky, finding the enemy first is the first step to victory in battle.

There are two types of radar systems commonly used in air warfare: early warning radar and fighter airborne radar.

Most of the airborne radar scanning range of fourth-generation fighters is 100-200 kilometers, and the scanning range of early warning aircraft installed in various countries can generally reach 400-500 kilometers. If stealth fighters are not taken into account, the opposing side can basically be detected at a range of 400 kilometers in a modern air battle on a large scale.

However, some domestic media have mentioned that the low-altitude detection range of China's latest early warning aircraft, the Air Police-2000, has reached a maximum of 800 kilometers, significantly exceeding the E-3 and E-2 early warning aircraft of the US military. If our army's early warning aircraft can really reach such a detection range, then it makes more sense for the US military to use the AI-controlled F-16 "blindfolded mine".

AWACS

In terms of stealth, after the installation of the US military's F-22 and F-35, fifth-generation fighters have gradually become the main force in air combat, and their important feature is the "stealth" ability to evade radar detection.

In 2017, our army's stealth fighter J-20 also entered service. As an imaginary enemy of the U.S. military for a long time, the J-20 is undoubtedly a huge challenge to the U.S. military's aerial radar detection capability, and if both sides are stealth aircraft groups in future air battles, the air battlefield is likely to return to the era of "white-knuckle warfare" in the absence of a huge gap in anti-concealment capabilities.

Of course, the "white-knuckle battle" in the air needs more "cannon fodder" to attract firepower, but this does not mean that AI fighters are "blindfolded".

First of all, under the principle of discovery and destruction, aircraft radars do not actually open fire immediately after detecting the enemy. Because the range of air-to-air missiles usually does not reach the detectable range of radar, and even if they do, they are not guaranteed to hit enemy aircraft.

Fighter pilots generally choose to "get closer and then fight," allowing the enemy aircraft to enter the "non-escape range" of air-to-air missiles before firing. In actual combat, the length of the inescapable distance is not fixed. But this inescapable range is usually positively correlated with the maximum range of the missile.

The Thunderbolt-15 air-to-air missile, which is currently the main force of our army, is reported to have a maximum range of 200 kilometers and a flight speed of Mach 5. The new Thunderbolt-17, already equipped, is reported to have a range of 300-500 km.

However, the AIM-120 air-to-air missile, the main force of the US military, has a maximum range of 160 kilometers and a flight speed of Mach 4. The AIM-260, which is under development and has not yet been equipped, is widely speculated that the maximum range may not be doubled by more than 200 kilometers.

air-to-air missile

In large-scale air combat, if the missile does not have enough range to destroy the target from a long distance. It may really take the spirit of a "kamikaze" to rush into the enemy's effective range and fire "close to the face".

The F-35 is no longer useful

The X-62A VISTA is the first attempt of AI on the main fighter of the United States, and other types of AI fighters similar to it have begun testing in the militaries of various countries as early as 10 years ago, including the American X-47B, XQ-58A, and China's "no detection" unmanned reconnaissance aircraft.

The X-47B and XQ-58A unmanned fighters of the United States have tried AI-controlled take-off, landing and flight, and the X-47B has also completed AI-controlled in-flight refueling.

However, these drones, which do not have a cockpit at all, basically cannot fly at supersonic speeds, and the onboard weapons are limited. The main function of this UAV is reconnaissance and relatively simple strike missions, and its air-to-air combat capabilities are relatively limited, which is completely incomparable with the F-16, F-22 and F-35, which are the main forces of the US military.

Of course, the price can't be compared.

The Valkyrie-type UAVs of the XQ-58A cost $7.5 million each. The cost of each F-35C is $90 million, and the annual operation and maintenance cost is as high as $6.6 million.

XQ-58A

Even with the purchase and maintenance, the cost of one F-35C in a year can buy 14 lighter AI UAVs XQ-58A.

Currently, the US Navy's Ford-class aircraft carriers can carry about 70 or so fighters. With the establishment of the US Navy, an F-35C squadron has about 24 aircraft, and the total cost is 2.16 billion US dollars. The cost of a Ford-class aircraft carrier is about 12.7 billion US dollars, and the cost of our army's Fujian ship is about 50 billion yuan.

In addition to the manufacturing cost, the R&D cost of the AI system is not high.

In 2020, U.S. contractor EpiSys Science, Inc. (EpiSci) was awarded a $7.4 million research and development contract from DARPA to develop credible tactical artificial intelligence (AI) algorithms for combat autonomy, a contract that is likely related to the X-62A VISTA seen today.

$7.4 million in R&D is quite cheap for a fighter jet.

The total cost of research and development of the F-35 is about $400 billion, and since 2019, the U.S. government has also set up an eight-year modernization plan for the F-35, with an estimated annual R&D expenditure of $1.9 billion.

To fight a war is to spend more money than whomever it is, and to spend it more efficiently.

AI fighters that can save money have been considered by the U.S. Navy to be the future on the flight deck of aircraft carriers. The U.S. Navy has proposed that it wants to replace the F-35 with a part of the AI UAV in the form of a wingman to form a new battle group of "F-35 + UAV". The U.S. Navy predicts that up to 60% of carrier-based aircraft wings will be made up of UAVs in the future.

However, cheap AI drones can fly only half as fast as fourth- and fifth-generation fighters, and it remains to be seen whether they can really win in many-to-one or cooperative battles in actual combat.

Manpower is too expensive, and it is impossible to recruit people

In addition to reducing the cost of advanced fighters, AI pilots can also solve some HR problems for the US military.

The first must be to reduce casualties. In recent decades, the total number of Air Force combat casualties in five U.S. overseas military operations (code: OIF, OND, OEF, OIR, OFS) is 1,064 (as of May 7, 2024).

Statistics on the number of U.S. military casualties in foreign military operations

Although the number of casualties in the five military operations was only 1/4 of that during the period of resisting US aggression and aiding Korea. However, these military operations have hardly engaged in large-scale air battles with an enemy of equal strength.

This is followed by a shortage of pilots. According to the U.S. Air Force Personnel Center, as of March 31, 2024, there were 11,921 pilots below the rank of lieutenant colonel on active duty in the U.S. Air Force. The U.S. Air Force has long wanted to have 13,000 pilots on active duty and about 8,000 more pilots in the Air National Guard and Air Force Reserve.

The U.S. military believes that a long-term lack of pilots is becoming a "potential crisis", and Kendall told the U.S. Congress in May 2022 that the U.S. Air Force is actively addressing the shortage of pilots.

The advent of AI fighters makes people wonder if this is one of Kendall's solutions to the shortage of pilots.

In fact, the salary level of US military pilots is very high, and the basic salary of a mid-ranking officer (such as a captain) plus other benefits is more than $100,000 a year. Pilots with higher ranks or longer service will also have more allowances. By comparison, the average annual salary in the United States was about $55,000 in 2022.

According to the 2024 U.S. military salary data released by the U.S. Department of Defense, pilots at the O-1 level (second-class lieutenant) receive a basic monthly salary of $3,826.2 when they serve less than two years. As the number of years of service increases, the basic pay increases. For example, an O-3 (captain) pilot receives a base salary of $5,783.7 per month after two years of service.

U.S. military salaries

U.S. military pilots can also receive an additional Monthly Flight Allowance (AvIP) based on the seniority they fly, which ranges from $150 to $1,000. To encourage staying, the U.S. Air Force also gives pilots a high AvB (AvB), which can reach $35,000 per year.

In addition, the cost of training pilots in the US military is considerable. According to a 2019 study by the RAND Corporation, the cost of training a basically qualified pilot in the U.S. Air Force ranges from $1.1 million to $10.9 million.

The cost of training for F-16 pilots is about $5.6 million, and for F-22s it is $10.9 million; The cost of training bomber pilots is also high, 7.3 million for the B-1 and 9.7 million for the B-52; The cost of training transport and maneuvering pilots is relatively low, from 1.1 million for the C-17 and 2.5 million for the C-130J; The cost of pilot training for RC-135 and other aircraft responsible for command, control, intelligence, surveillance, and reconnaissance operations is about 5.5 million.

RAND Corporation's estimated costs include the cost of flight hours, student accommodation and compensation, but this is only a part of the cost of pilot training, which also requires many other expenses, such as simulator usage, teaching materials and instructor fees.

AI fighters are not sci-fi, but far away

When handing over fighter jets to AI, the first thing to worry about is not the sci-fi "AI apostasy".

Since the outbreak of large language models, all sectors of society have been worried about the imminent arrival of AGI and strong artificial intelligence. As a result, many people also have the fear that AI is about to destroy the world.

But handing over a fighter jet to an AI isn't that scary. Because the AI that controls the fighter is mainly discriminative AI or decision-making AI, these systems do not have the strong generalization ability of ChatGPT, which has little to do with today's large language models and generative AI, and is also far from AGI.

Because military equipment is too unfriendly to AI large models.

The first is computing power, Palmer Luckey, CEO of the American attack drone manufacturer Anduril Industries, once said: Tesla's artificial intelligence is better than that of American military vehicles, and the computing power of the iPhone is much more powerful than the systems commonly used by the Department of Defense.

Military chips, especially those used in high-end military equipment such as strategic nuclear submarines, F-22 supersonic fighters, and intercontinental missiles, have high requirements for reliability and the ability to withstand harsh environments, so they usually use larger process sizes than civilian chips.

Mainstream AI chips such as GPUs and TPUs generally use smaller process sizes such as 7nm and 5nm, while many military chips use process sizes of up to 250nm.

Larger transistor sizes have greater current carrying capacity and a thicker oxide layer, which can more effectively resist environmental aggression and have high resistance to harsh conditions such as high temperatures, radiation, and electromagnetic interference. Small-sized chips are also more susceptible to radiation-induced charge accumulation (e.g., single-event flipping caused by cosmic rays) due to their tiny feature size.

Large-size transistors also have better heat capacity and heat dissipation than smaller-sized transistors, which allows them to work better in high-temperature environments without being prone to performance degradation.

On the other hand, AI models used in the military industry need to have high stability and interpretability, and the performance of large models represented by Transformer is still imperfect in this regard.

To apply Transformer to military equipment, additional explanatory mechanisms or tools need to be integrated to increase the transparency of the model. In terms of stability, the stability of the model can be ensured through a refined testing and verification process, combined with more stable training techniques, such as regularization methods and robustness optimization.

"Although the military industrial unit is conservative, it can also feel the impact of the AI model." An expert engaged in related scientific research told Tiger Sniff that some people are already trying to "bring" large models such as Transformer on the plane, and some military scientific research units in China are studying AI-related technologies, including using AI to help fighter jets optimize take-off and hovering.

However, the expert also said that although the X-62A VISTA has completed a "dogfight" (dogfight) with a human fighter, it is still a long way for AI to fly a fighter completely autonomously in the real world, and even participate in combat.

A spokesperson for DARPA, describing the problems with AI-piloted fighter jets, said that AI could push the aircraft's flight control surfaces beyond its limits. When performing high-intensity maneuvers, it is physically restrained and damages the aircraft. To this end, DARPA has set up a "safe stroke" in the X-62A VISTA to prevent the AI program from overloading the airframe.

In fact, to achieve full AI autonomy of the F-16 fighter, the AI must be able to handle extremely complex battlefield situations, including rapidly changing tactical environments, identification of friend or foe, threat evasion, and performing a variety of missions such as reconnaissance, air combat, and ground attack. These decisions need to be made in the seconds of a second, requiring AI systems to be highly intelligent and responsive.

In addition, the battlefield environment is volatile, and AI also needs to have the ability to learn on its own to adapt to new tactics and technological developments. This requires AI systems to continuously learn from experience and optimize their own behaviors and strategies.

"AI and fighter pilots are more collaborative, and pilots can hand over control to the AI for a minute or two." Judging from Kendall's description of AI, the role of "AI driver" should be similar to that of assisted driving and the popular concept of "Copilot" in large language models in the short term.

View original image 47K

• 

• 

• 

• 

• 

• 

• 

• 

•