Autonomous Drones: From a Lost Video Link to Mass Deployment

From evacuation to software-defined defense

Late on February 23, 2022, Yaroslav Azhnyuk returned to Kyiv with his fiancée after visiting a prospective wedding church in Lviv. When the invasion began, their immediate problem was evacuation: gathering family, finding fuel, helping a friend without a car, and driving west for 17 hours. They did not know whether Kyiv would hold or how long they would be away.

Azhnyuk had studied applied mathematics in Kyiv and lived in San Francisco from 2014 to 2020 while building Petcube, a pet-camera company. He describes his eventual transition into defense through the cameras themselves: devices that dispensed treats became devices used to deliver explosives. He had not expected to manufacture weapons; his first commitment was to contribute whatever he could to Ukraine’s defense.

By 2023, he expected a prolonged war and a new arms race. Drones offered a particular opening for his entrepreneurial and technical skills: software updates could improve an existing fleet’s battlefield capabilities. He compares that distribution mechanism to giving every Roman legionnaire a new helmet overnight—an analogy for upgrading deployed equipment without replacing the platform.

In the Latent Space conversation, that engineering thesis accompanies a moral commitment. Azhnyuk presents resistance to invasion as an obligation and weapons developers as support for those doing the fighting. A host distinguishes homeland defense from the technology’s eventual spread beyond its creators. The response acknowledges a maker’s responsibility while arguing that immediate survival cannot wait for every possible downstream misuse to be resolved.

Sensors, autonomy, and complete aircraft

The Fourth Law builds autonomy that runs onboard drones, with mass deployment as its goal. In parallel, Azhnyuk and his Petcube partners started Odd Systems, making thermal cameras: cameras that detect thermal radiation to support nighttime vision. At the time of the conversation, the companies were moving closer together, with a merger considered likely but not completed.

The guest describes three complementary businesses: - Cameras: daytime and nighttime imaging components for other manufacturers. - Autonomy modules: onboard capabilities integrated into other companies’ aircraft. Together, the component businesses reportedly serve more than 200 Ukrainian drone manufacturers. - Complete drones: battery-powered frontline strike drones, bombers, and interceptors supplied to Ukraine’s armed forces and government. The group does not build deep-strike aircraft, though it supplies some autonomy systems for that use.

The interceptors target Shaheds and reconnaissance aircraft. ISR means intelligence, surveillance, and reconnaissance; here, ISR drones observe positions and report target locations. The guest gives a maximum speed of 326 km/h for the group’s battery-powered Zero interceptor, compared with approximately 220 km/h cruising speed for a Shahed. No test conditions or specific Shahed variant accompany the comparison.

The group also operates simulation and a training school. Construction of two semiconductor plants for thermal-camera sensors is described as about to begin. These are planned facilities, extending the undertaking from software and aircraft into sensor manufacturing.

Terminal guidance closes the final approach’s communication gap

Autonomy and communication resilience solve different problems. Autonomy takes over operator tasks; fiber optics preserves a communication path. Both can be used on the same aircraft.

For the manually piloted camera-equipped drones discussed here as FPV drones, losing visual feedback does not require enemy jamming. As an aircraft descends, a hill, forest, or Earth’s curvature can obstruct its direct radio path. This creates a radio shadow, where the operator can lose the image precisely when it is needed to finish the approach. The guest describes this happening in the final 20–60 meters.

Distance compounds the problem. The guest reports Ukrainian expectations of at least 20–25 km for FPV drones and many strikes at 30–40 km using regular 10-inch aircraft. At those distances, he says, links may degrade even at altitudes of 60–100 meters. These are reported operating conditions, not standardized range tests.

Terminal guidance transfers the final approach to onboard control. A human flies toward a target and visually selects it; the drone then continues steering toward it after the visual connection disappears. This was the company’s first launched product, described as covering roughly the last 500 meters. It removes the need for continuous operator feedback during one vulnerable part of the mission.

The communication options have different tradeoffs: - Analog radio: video and control can travel over separate radios; video jamming and control jamming are distinct failure modes. - Digital radio: another link option, independent of whether the aircraft has onboard autonomy. - Fiber optics: a resilient physical connection that adds weight and reduces range and maneuverability. The guest estimates roughly 3 kg for a 20 km connection, including the cable carried and unspooled during flight.

That weight competes with payload. In the guest’s example, a 15-inch drone carrying fiber over 20 km can carry only one or two kilograms of explosives. He describes 30 km as probably the practical maximum and 40 km as very problematic for that approach.

Fiber also ties battlefield supply to AI infrastructure. The guest reports prices rising from about $4 to $32 per kilometer over a few months at the beginning of the year, attributing the increase to data centers competing for the same supply. He says Russia scaled fiber-controlled drones first and Ukraine had mostly caught up roughly three months before the conversation. He declines to disclose the share of damage attributable to fiber-controlled versus autonomous drones, leaving their relative effectiveness unquantified.

From piloting to specifying a mission

The economic attraction begins with inexpensive hardware. The guest contrasts a roughly $400 FPV drone with a roughly $4,000 standard NATO artillery shell, then adds the gun’s amortization and vulnerability. But purchase price does not measure equivalent military effect: he acknowledges that drones lack artillery’s firepower and that artillery remains necessary. His broader orders-of-magnitude comparison concerns perceived versatility, not a defined performance metric.

He makes a similar distinction about tanks. Their utility is diminishing in his assessment, but armor and firepower remain valuable. Vulnerability to incoming drones is the immediate problem; better protection could restore some of that utility.

When a host challenges the leap from guidance to full autonomy, the guest clarifies that terminal guidance was only the first and easiest component the organization fielded. His broader analogy is ridesharing: the operator specifies a mission instead of continuously piloting. He contrasts that with months of pilot training and potentially 30 minutes of manual flight toward a target.

His autonomous bomber example begins with a human specifying the starting position and target area through a smartphone. The aircraft takes off, navigates and locates itself on a map, finds the designated area and targets, attacks, assesses the result, returns, and lands. Video can be reviewed after recovery if no radio connection was available.

A human initiates the mission; execution then proceeds autonomously. That distinction resolves the suggestion that no human decision occurs anywhere in the process. It does not, by itself, specify targeting constraints or opportunities for later intervention.

The guest proposes a four-orders-of-magnitude capability gain from three separate effects: - Operator access: 100 times as many potential users because specialist pilot training is unnecessary. - Mission success: a claimed tenfold improvement. - Reuse: ten times the utility per aircraft through returning as a bomber rather than serving as a one-way munition. These are asserted multipliers, not measured results. Another participant interprets the success claim as autonomous bombers outperforming human-piloted bombers tenfold, but the original claim does not establish that baseline.

Mission success needs an outcome and an observer

A more concrete example concerns an operator with the call sign Grom. The guest credits terminal guidance with increasing his mission success rate from 20% to 71% in a reported field outcome, while expanding the described area of effective strikes near the front from 3 to 10 km. No sample size, evaluation period, controlled comparison, or consistent success criterion accompanies those figures.

A request to define success separates several outcomes and failure modes: - Target effect: damage and destruction differ, as do the effects sought against infantry, dugouts, machinery, and radio equipment. - Impact without detonation: reaching the target may still produce no intended effect. - Communication loss: control jamming, video jamming, and loss beyond the radio horizon can each prevent arrival. - Aircraft or navigation failure: a broken motor or getting lost can end the mission before terminal guidance begins.

The guest organizes onboard capability into a five-level framework of his own: - Level one — terminal guidance: completing the approach after human target selection. - Level two — bombing. - Level three — autonomous target detection and engagement decisions. - Level four — autonomous navigation. - Level five — autonomous takeoff and landing. He reports some industry adoption. Self-driving classifications inspired the framework, but the two systems do not map directly.

Confirming the result requires observation beyond the strike drone. In the described workflow, an ISR aircraft identifies a target for the FPV operator and observes the outcome. Teams rotate surveillance drones to maintain coverage. When that independent view is unavailable, another FPV may be sent to inspect the result. Reaching a target and establishing what happened there are separate capabilities.

An autonomy level does not describe the deployment

Capability level is only one of eight interacting dimensions in the guest’s broader framework. The others explain where the system works, what surrounds it, and how widely it can be deployed: - Platform: quadcopters, fixed-wing aircraft, missiles, shells, ground vehicles, or sea vehicles. - Operating domain: combinations such as ground-to-ground or ground-to-air interception. - Coordination: capabilities such as swarming, drone carriers, and drone nests. - Environment: lighting, season, temperature, humidity, target type, and concealment. - Command and control: coordinating assets, connecting them to higher command, and collecting data. - Infrastructure: simulation, data collection tools, security, and deployment mechanisms. - Deployment scale: fielding 100 systems presents different problems from fielding 100,000.

These dimensions combine. A capability must work on its particular platform, in its operating environment, while integrating with command systems and supporting infrastructure. A capability label alone says little about readiness at scale.

A host raises distribution shift—changed conditions that challenge the system’s interpretation—through a hypothetical: fog arrives during a vision-guided approach. Could the drone attack the wrong object? The response argues for drones’ precision relative to mortar fire and predicts that, within five to ten years, using weapons without AI could be considered immoral because of greater unwanted damage. That forecast leaves the specific question of handling fog unresolved.

The next hypothetical is an AI general, integrating battlefield information and commanding operations. The respondent considers it technically possible, while acknowledging engineering work and a larger deployment challenge. How broadly such decisions could extend across missions and command levels remains an open question.

The safety concern extends beyond targeting accuracy: command decisions would put friendly personnel at risk, as well as direct drones. The respondent also distinguishes operations near evacuated front lines from missions involving civilians, while acknowledging limited expertise across the full range of military scenarios.

The long tail between a prototype and a fleet

A participant recalls predicting in 2013 that autonomous attack drones would make infantry untenable. The response returns to a dugout: an armed soldier still prevents an opponent from occupying it. A Ukrainian punk song titled “2030” supplies the counterimage—AI and cyborgs coexist with infantry still getting cold underground.

Humanoid robots prompt speculative five- or ten-year timelines. The guest’s firsthand engineering lesson is more concrete: a prototype, a product ready for millions of units, and millions of deployed units are different achievements. Each additional autonomous behavior creates more edge cases. AI coding tools might accelerate development, but he does not claim they eliminate that gap.

Terminal guidance gives the discussion a bounded starting point: image recognition and tracking build on established capabilities, though making them reliable is not trivial. Broader decision-making adds more possible behaviors and therefore more ways to fail.

The robotics architecture is described as sensors → perception → planning → control → actuators. The implementation choices include: - Classical logic: algorithmic processing through the stages. - Hybrid systems: neural-network perception paired with conventional logic elsewhere. - End-to-end neural systems: a large network maps sensor inputs—including images, sound, or accelerometer readings—to controls. The guest also contrasts classical controllers with neural controllers trained through reinforcement learning. The choice changes system behavior, not merely how the code is organized.

At a higher level, he contrasts a world model intended to interpret a task in its environment with explicit subsystems. His photography example is compact: one model interprets a request to find and photograph a person; a modular system combines navigation, person detection, and an approach behavior. Neither is presented as inherently superior. Their feasibility and resulting behavior differ.

Automating a vehicle does not automate the operation around it. The guest describes retaking territory as coordinating thousands of people, dozens of drone types, ground forces, reconnaissance, aviation, tanks, and logistics. He rejects an imminent disappearance of riflemen and emphasizes how much military expertise lies beyond his own understanding.

Manufacturing sets the scale of deployment

The supply-chain question has two parts: can a drone be built without Chinese inputs, and can it be built in sufficient volume? The guest contrasts Ukraine’s reported four million FPV drones in the preceding year with a hypothetical Chinese capacity of four billion. The Chinese figure has no production period or supporting calculation; it expresses his concern about industrial scale rather than reported output.

His larger hypothetical combines mass manufacturing, autonomous aircraft, and maritime transport to illustrate how inexpensive incoming systems could overwhelm defenses. The proposed requirement is defensive mass at a comparable cost. Being able to intercept one aircraft does not establish the capacity to withstand a large attack.

The respondent nevertheless declines to infer overall Chinese military supremacy from drone capacity. He cites insufficient information about Chinese military performance and the continuing importance of artillery, naval aviation, air forces, missiles, electronic warfare, and satellites. Current factories also do not fix wartime output: another participant argues that urgency could accelerate new factories and allied production. The comparison with US mobilization during World War II draws a qualification—expansion was difficult, and factories not facing attack had an advantage.

The manufacturing bottlenecks extend beyond raw materials: - Thermal cameras: a major constraint identified by the guest and a focus of his organization. - Motors: suitable magnets must still be turned into high-quality motors. - Rare-earth processing: he identifies refining capacity and investment as the constraint, rather than deposits existing only in China. - Labor: production also depends on available workers and coordination across countries.

The guest also cautions against arrogance in assessing Chinese technology. Drawing on time spent living in and visiting China, he points to WeChat as an example of consumer software he considers more advanced than Western alternatives. His broader message is to take competitors’ technical capabilities seriously.

Readiness, endurance, and deterrence

Asked about US readiness, the respondent proposes closer exchange with Ukraine: learning from specialized companies, battlefield personnel at different command levels, and the government arrangements supporting defense businesses. A second proposal is procurement reform, including expanding the effort called Drone Dominance. These are the respondent’s recommendations.

He acknowledges that Ukraine’s theater differs substantially from potential US conflicts. He nevertheless considers its maritime warfare and long-range drone experience relevant, citing Shaheds with claimed ranges up to 2,000 km as an example with implications for Pacific islands and aircraft carriers.

Greater production does not guarantee victory, the respondent argues. He interprets Ukraine’s resistance to Russia and Iran’s resistance to the United States as examples of weaker defenders preventing stronger attackers from achieving all their aims. Another participant challenges the analogy because both defenders rely on Chinese supply chains. The response concedes partial reliance while asserting that Ukraine relied mainly on existing resources during the first six to twelve months after February 2022. The argument then broadens to military, economic, and political endurance: internal pressure can constrain a war, and willingness to bear its costs depends partly on whether the population sees its country as aggressor or victim.

The security-assurance debate turns to the Budapest Memorandum. One participant interprets Ukraine’s experience as undermining confidence in external assurances and predicts that more countries will pursue nuclear weapons over the next couple of decades. A different participant explicitly advocates nuclear weapons for Japan, South Korea, Poland, and Ukraine. The prediction and the advocacy are distinct positions. The speaker discussing Ukraine then presents nuclear weapons as potentially the only credible deterrent suggested by recent history and links European uncertainty about US support to Washington’s reluctance to confront a nuclear-armed Russia. These remain the participants’ interpretations and policy positions.

There is no single drone race

Asked whether Ukraine is ahead of Russia, the guest first challenges the single ranking. He describes at least 30 categories across aerial, ground, surface maritime, and undersea drones. Their missions include logistics, evacuation, reconnaissance, strikes, radio relaying, mining, and demining.

He contrasts Ukraine’s competing products with Russia’s centralized scaling: dozens, sometimes more than 100, Ukrainian companies and products in a category, versus typically one to three Russian products scaled centrally. Leadership must therefore be assessed category by category.

His contemporary assessment separates the categories: - Frontline strikes: Ukraine had recently moved ahead after scaling fiber-controlled drones; Russia previously held a slight lead. - Mid-range strikes, roughly 40–200 km: Russia had led, but he is uncertain whether Ukraine was moving ahead. - Deep strikes: he claims a recent Ukrainian advantage in damage inflicted, without comparative damage figures. - Sea and ground drones: he describes a consistent Ukrainian lead at sea and a probable lead on the ground. - Navigation components and aircraft-dropped winged bombs: he credits Russia with strengths in these areas. These are the guest’s assessments, not matched comparative evaluations.

Helicopter interception illustrates the gap between an incident and a repeatable capability. The guest characterizes reported Ukrainian drone interceptions of Russian helicopters as exceptional cases. A systematic capability requires identification, potentially using radar followed by visual detection, and an interceptor in the right place at the right time. Reaching the target is only part of interception. Meanwhile, he says both sides use helicopters against deep-strike drones.

Countermeasures depend on what the drone needs

The closing technical discussion challenges the premise that simple defenses are useless. The guest describes shotguns as widely used against drones, with success strongly dependent on operator skill. His secondhand example of a soldier reportedly shooting down about seven FPV drones before being killed is presented as exceptional, not a survival estimate.

Jamming depends on both frequency coverage and the drone’s reliance on radio. Both sides seek frequencies outside opposing countermeasures’ coverage. The guest’s categorical claim that matching-frequency jamming will stop an approach applies only to drones without fiber control or last-mile autonomy. It is not accompanied by measured protection or survival probabilities.

Other defenses act physically on the aircraft or its approach: - Mechanical barriers: fishing-net corridors over roads obstruct drones. - Vehicle additions: protruding structures and nets are intended to trigger detonation away from the hull. - Small interceptors: both sides are described as experimenting with onboard guidance to intercept approaching FPV drones. - Protective drone arrays: the guest imagines vehicles carrying many defensive drones as active armor. This remains a proposed future design.

Cost per effect includes simultaneous arrivals

Cost per effect broadens the comparison beyond ammunition price to what a defense can accomplish. The guest recalls an RTX exhibition conversation about a 10-kilowatt laser reportedly requiring about three seconds to disable an FPV drone, with that interval restarting after retargeting an evading aircraft. These are remembered vendor statements, not supplied test results.

He recalls a system price around $3 million, which he equates to 6,000 FPV drones, and doubts it could withstand even 600 arriving simultaneously. The numerical limit is hypothetical, but the distinction is substantive: sequential engagements and simultaneous arrivals are different workloads. Handling many aircraft one at a time does not establish enough throughput for a concentrated attack.

The operating context also changes the desired effect. Protecting a stadium against a few drones differs from battlefield defense against mass arrivals. Above spectators, falling debris could make net capture preferable to shooting an aircraft down.

When a host suggests a more powerful laser, the response favors achieving the same effect with less power. The objection concerns expense, weight, transport, manufacturing volume, and replacement after damage. Those constraints determine how much of a long front can be covered; they do not establish that higher power is inherently less effective.

The guest favors small systems that can be produced widely and improved quickly. His smartphone analogy returns to deployment breadth: individual capability matters alongside how many systems can be built, operated, and upgraded.

Technical ambition, human consequences

Near the end, the interviewee challenges the assumption that Russia is an unintelligent adversary. He asks what “smart” means and warns against relying on an opponent’s supposed incompetence. The questioner acknowledges that an information bubble may shape that impression.

The conversation returns to the human cost through an American volunteer friend’s account from near Chasiv Yar. According to the guest, repeated Russian infantry assaults were psychologically traumatic for the Ukrainian defenders firing on the advancing soldiers. This is a secondhand account of the people behind the discussion’s production volumes and technological comparisons.

A late StarCraft analogy compresses the technical ambition: Protoss technology with Zerg speed—advanced capabilities combined with rapid production and deployment.

The final questions complete the opening story. Azhnyuk and his fiancée did marry, but postponed the wedding celebration until the war is over. His closing appeal is for investment in defense and security, grounded in his belief that maintaining peace requires preparedness for war.