Ukraine built about 4 million FPV drones last year at roughly $500 each, and Yaroslav Azhnyuk says the United States has no capacity to do the same. He also says between 10% and 20% of those drones hit anything, which is the number that makes the rest of his argument work: a low base is cheap to double. Most people arguing about autonomous weapons argue about whether machines should be allowed to kill. Azhnyuk's position is the opposite of the one the debate assumes, and he states it as a prediction rather than a preference.
"I think like within 5 to 10 years, it will become unethical to use weapons without AI. The same way as it might become unethical to use cars without AI because those that are less lethal and non AI weapons will be more likely to cause collateral damage not less likely."
Azhnyuk spent 14 years building connected cameras for pets at Petcube before Russia's invasion; he now runs The Fourth Law, which sells the autonomy modules that Ukrainian drone makers bolt onto their own aircraft.
I listened to the full interview so you can skip it.
Here are the 15 takeaways that matter.
👤 Guest: Yaroslav Azhnyuk, co-founder of Petcube, the Kyiv consumer-camera company, and now head of The Fourth Law, which builds autonomous guidance software for drones used by Ukraine's armed forces
🎙️ Host: Craig S. Smith, a longtime New York Times correspondent who founded and presents Eye On A.I.
📰 Published: 31 August 2026
🔴 YouTube | 🔗 Episode page | ⏱️ 53 min
Key Takeaways
The ethics argument runs the other way: he expects AI-free weapons to be judged as the reckless option
His comparison is cars, and his reason is collateral damage rather than capability
A $500 drone that kills a $5M tank is the equation driving a decade of global rearmament
Adding autonomy costs a few hundred dollars more and multiplies who can fly it
Front-line FPV drones hit their target between 10% and 20% of the time
Level-one autonomy multiplies that by two to four times, which is where his revenue comes from
Poland shot down 4 of 20 Russian decoys, using kit two orders of magnitude more expensive
He calls the official verdict on it a disaster dressed up as resolve
Ukraine's advantage is 12 yrs of war nobody wants to take from it
A couple of million people have worked on this in some capacity
Ukraine can build 4M drones a year and the United States cannot build any such number
His China scenario is 5,000 submarines carrying autonomous drones to a foreign coast
Eric Schmidt is the closest competitor and the only Western operation whose equipment he says works
The difference is that Schmidt's engineers are in Kyiv iterating with actual brigades
He sells modules, not drones — Android to Schmidt's Apple
Ukraine has no semiconductor manufacturing, and the sensors have to be imported
The real AI risk is not a defense Skynet, it is what the hyperscalers build first
Most of what is sold as AI in defense is image recognition, and about 10% of a drone company's staff work on it
1. Pet Cameras to Drone Cameras
The pivot is the story, and Azhnyuk tells it as one sentence. He and colleagues from Petcube started a second defense company making thermal cameras, now among the most widely used in the field
"we started with making cameras for pets and now we're making cameras for drones and we were throwing treats to dogs and now we're throwing explosives to people who came to occupy our land"
Smith's response: "Fate loves irony for sure."
Those cameras sit in interceptors that he says have brought down a couple of thousand Shahed drones, and in front-line FPV aircraft
The Fourth Law is the other company, and it is a software business. "We're working on massively scalable autonomous systems for drones and at this point it involves frontline FPV drones and then the interceptor drones and some other ones."
Interceptors are the most dynamic category, and the class covers intercepting Shaheds, reconnaissance drones and smaller FPVs
2. Modules, Not Whole Drones
The competitive framing he uses is the smartphone market, with Eric Schmidt's operation as the closed system. "I would say what Eric is doing is more kind to sort of Apple strategy. They have the full vertical integration they ship that to you and what we are doing along with some other Ukrainian makers is more kind to Android strategy where different teams work in different modules."
The modules attach to whatever airframe a manufacturer is already building, and their capability grows independently of the aircraft
He also makes drones, and is dismissive about that half of the business. Making the drone is the commodity; making the factory is the hard part
"So you can make a hamburger in your kitchen, but the really difficult thing is make McDonald's making a factory which makes same hamburgers with a very short time frame."
Two airframe families are winning, on his account: rocket-shaped tail-sitters with four wing propellers, and conventional fixed-wing aircraft. His modules go on both
He named Wild Hornets and its Sting interceptor as the most successful of them, and mentioned a public row in which Sting accused another maker of copying the design. His view of that argument was that it is like accusing someone of copying the idea of a car
3. Schmidt Is the Main Rival
He treats Eric Schmidt's drone company, Swift Beat, as the competitor that matters, and said so without pretending otherwise. He put its public work into three lines: Shahed interception, front-line autonomy, and a mid-range strike category he described as reaching up to about 200 kilometers
"I think certainly I mean Eric Schmidt is capable on AI like they're our main competitor/per."
On the split: he rates his own side as close on ground autonomy and somewhat behind on aerial interception
The commercial position is not Ukraine-first. He said the interceptors have started selling into Poland and Romania, and that he was not sure what the plan for Ukraine was, having heard of production bottlenecks there
Schmidt's other vehicle is an incubator, D3, and Azhnyuk had a hand in starting it. "I was like full disclosure I was somewhat involved in and stuff."
Its public stance is defense tech, about half the portfolio companies come out of Ukraine, and the fund is small — he put it at a couple of dozen million dollars
Not everyone competing on interceptors competes on autonomy. Some capable interceptor makers buy his AI modules instead of building their own
4. Ukraine Has No Chip Fabs
Smith asked where the chips in Ukrainian interceptors come from, having heard a Ukrainian manufacturer say they were made domestically.
The answer was flat. "There is no semiconductor manufacturing in Ukraine."
He thought the manufacturer had misunderstood the question, and was talking about circuit-board assembly
What Ukraine does have is board assembly, and reasonable depth of it. "Yeah custom facilities where you can do up to eight layers. So you can do a lot in Ukraine."
The sensors are the binding import. Thermal camera sensors have to be bought from foreign fabs, and there are plans — his own included — to build a sensor factory in Ukraine that has not happened yet
5. The 5 Levels of Autonomy
He has published a classification the market has largely adopted, and it maps onto the self-driving levels. "So, when we talk about the individual autonomy for single drone, we talk about these five levels of autonomy and when we talk about the ecosystem, we talk about sort of eight dimensions of autonomous system."
"level one is autonomous terminal guidance" — a human pilot flies and locks the target, and the drone completes the strike on its own in any circumstances
Level two is the same but for dropping ordnance rather than hitting, including the calculation of when to release and the escape
Level three is the drone finding and identifying the target itself, then deciding to strike, which resolves back into a level one or level two attack
Level four is autonomous navigation of the whole route
The economics of the module are the constraint on all of it. "So if a system that allows drones to be individual autonomy costs even a couple thousand dollars it will be difficult to use such system on FP drones which cost just $500 per piece."
"The system has to be relatively inexpensive and also relatively easy to manufacture so that it can be manufactured in millions because that's what I mean demands."
6. 8 Dimensions of the Fight
The second half of his framework is what a whole autonomous battlefield requires, and it is the part he says people miss.
The dimensions he listed: the five levels above; the carrier, which runs from quadcopters and fixed-wing aircraft to missiles and artillery shells; the domain, meaning ground to ground, ground to air or ground to sea, and every combination of them
Higher-order autonomy is its own dimension — swarms, carriers and nests rather than single aircraft
Conditions and targets are another: day, night, winter, summer, and the difference between attacking infantry, a tank and a vehicle, which he said require different development
Command and control is the sixth, and it is a data problem as much as a military one: a brigade commander has to manage the units, the autonomous assets and everything those assets report
The seventh is infrastructure — simulation used routinely for testing, data collection, security, remote deployment
The eighth is deployment scale, and he treats it as a capability rather than a footnote. A system fielded in hundreds of thousands of units and a system in field trials at a couple of hundred units are different systems
7. Kyiv Is the Defense Valley
Smith's own read, offered mid-conversation: "I think it's like what's going on here is just astounding. This is this is like the defense valley."
Azhnyuk's evidence is the density of the room they were sitting in. Another defense founder was somewhere in the same cafe; he said the same thing happens to him constantly. "So it's a very very like Silicon Valley vibe."
The advantage he claims is not technology but experience, and he defines it as strategy. "I say this all the time that my definition of strategy is that strategy is a plan of victory using the uniqueness of your situation."
"The uniqueness of Ukraine situation is that we've been at war with Russia for 12 years and no one else in the world wants to steal that advantage from us."
A couple of million people have been involved, he said, as fighters, makers, engineers, entrepreneurs, assembly workers, volunteers or clerks
"So this is an incredible domain experience and it will remain so for decades to come even if tomorrow there is ceasefire."
He expects the rearmament to run for decades, and named three drivers. "And I like for me it's obvious the world will be rearming for the next couple decades driven by supply chain instability by obviously advances in drones and robotics and obviously by AI."
The unconventional version worries him too. "I think we have probably another dozen or dozen and a half countries in the world who are now working on their nukes who do not have an official nuclear status."
His definition of "after the war" is not a ceasefire. "If we have ceasefire with Russia in 2026, this is not after the war. After the war is when the global arms race would stand still."
8. A $500 Drone vs a $5M Tank
This is the calculation the rest of the argument rests on. As he put it, these systems "can take out $5 million tank you have a drone costing $500."
"And if you can add autonomy to that drone for the cost of couple hundred dollars on top and autonomy itself makes every drone usable by hundred times as many people."
"You don't need to train as pilots makes 10 times more precise makes it not one time but multiple time bombers probably 10 times more utility."
He then worked the arithmetic out loud, and it came apart as he did it. "5 million dollars and $500. How many orders of magnitude is that?"
"It's like,000 10,000. So it's 10 like 10 to the 4th multiplied by 10 to the 4th. It's like 10 to the 8th."
The point he was reaching for is that autonomy multiplies both the cost advantage and the number of people who can use the weapon at once
9. Poland Was a Disaster
His worked example of what that gap looks like in practice is the September incursion into Polish airspace. As he described it, "in September there were 20 Russian Shahed imitators" over Poland
"Only four of those were shut down. With equipment that cost probably two orders of magnitude more than the assets."
NATO's secretary general publicly called it a job well done. Azhnyuk did not agree: "But it's obvious to everyone it was a huge disaster."
The scenario he asked Smith to consider next is the same event with live weapons. "Now imagine Russia attacking with those were even drones without any explosives in them. Imagine Russia attacking with 50 actual Shaheds."
Russia's framing would be that the targets were legitimate. The question is what Europe does about it
"Well, most people I talk to and EU say that we won't have the guts to do that." Even a proportionate response would not settle it
"Will be huge defeat for Europe, huge defeat for the free world. It will be a precedent. And Russia doesn't care." He added that the market and economic consequences would follow
10. What China Could Build
The production comparison is the part with the clearest read-across for anyone thinking about defense supply chains. "Now the second scenario is Ukraine produced about 4 million of these inexpensive FPV drones last year, $500 each."
"If Ukraine can do that, by the by the way, United States cannot do it today." There is no capacity to produce at that volume
China, on his reckoning, could go an order of magnitude further, swap propellers for wings, launch from tubes and reach 200 kilometers instead of 30
He rates the Chinese pieces as already in place — DJI as the world's leading drone technology company with the autonomy to match, and Chinese shipbuilding as very impressive
"So, what if China builds 5,000 underwater submarines carrying a couple hundred of these autonomous drones, very inexpensive, capable of going to 200 kilometers in land each and shows up to the coast of any coast or California or one or Great Britain. Is there a good response to that? I don't think so. Hunt for subs won't be enough."
11. Why Isn't Anduril Here?
Smith asked why, if this is a solved problem, it has not happened already.
His answer was that it is not only Russia that cannot do it. "No, the United States doesn't have full autonomous drones." They rely on GPS, and what exists is not cheap
He was careful about the American defense-tech companies and then pointed at the gap anyway. "I have tremendous respect for companies like Anduril." His question to them was why their systems are not in Ukraine helping fight the war
Schmidt is the exception he keeps making. "Although this is what interests me about Eric's. He is different in that extent because his stuff actually works."
"I think Eric is a remarkably smart guy I had a chance to meet him a couple times and you just saw around the corner before everyone else." He noted Schmidt was doing work for the US Department of Defense early, and had unusual freedom to think about the problem
The reason he gives for that working is a startup lesson rather than a technical one. Azhnyuk went through Y Combinator in 2016, where the advice is to talk to your customers
For an American defense contractor the customer is the Department of Defense, which it already meets. For this technology the customer is the units that will fight the next war, and they are in Ukraine
"I think like one of the smartest things that he did is actually hiring a lot of people here and very quickly iterating with actual military units." Schmidt keeps a deployed engineering team working with particular brigades
12. Artillery Shell With Props
His answer to the ethics of autonomous targeting is to relocate the decision. "You can think of an autonomous drone as an artillery shell with propellers and you make a decision the moment you launch the moment you give command."
"So you cannot stop an artillery shell 500 meters away from the target and ask for human confirmation whether it's the right target or whether we shot 2 km in wrong direction and hit the village."
The prediction the episode is named for follows from collateral damage, not from capability. "I think like within 5 to 10 years, it will become unethical to use weapons without AI. The same way as it might become unethical to use cars without AI because those that are less lethal and non AI weapons will be more likely to cause collateral damage not less likely."
The third leg is that the restraint is unilateral. "And the third one is it can be the good guys but bad guys are not going to abide by the rules."
"So we got to have at least the capability to have these systems working not at the speed of a human command chain but the speed of a software command chain"
He compared American policy statements on lethal autonomy to nuclear enrichment limits — the position is that fully autonomous targeting is off the table, and his view is that the step from here to there is a switch rather than a program
13. It's Software, Not AI
The framing he most wants changed is the label. "What we should be really talking about I think is software in defense."
"AlexNet is what 15 years old and like image recognition capabilities are at least a decade old and there are more sort of cases and a lot of what is called AI for defense is just image recognition."
"So people think that okay drone with AI something like a ChatGPT with propellers that can explode. It's not like that at all."
His model of a robotic system is sensors in, three software layers, actuators out — perception, planning and control. Each layer can be classical software or a neural network, or the three can be collapsed into one network
On the doomsday framing he was dismissive and specific. "And for years ahead an AI enabled defense system will be no closer to a Skynet appraising that than a warehouse of 155 caliber shells."
"And then I think before we are at a risk of being wiped out by defense skynet, we will be many years under risk of being wiped out by advanced AI developed on servers of one of these big hyperscale firms."
Sovereignty is the second-order consequence, and he thinks one company benefits most. He cited news that Washington might block an American AI product in Germany, and said the signal that sends to every country is "you got to develop your own models"
The person he expects to be happiest about that is Nvidia's chief executive, since a country that wants its own model still has to buy the chips. "Well, he's in very good business. He's very smart."
On what would actually get to general intelligence, his bet is embodiment. "I think there's very high probability that to develop AGI it will have to it will have to get into physical bodies."
His analogy is Plato's cave: today's language models learn the world from what people wrote about it, shadows on a wall, and "they cannot experience the world themselves"
14. Edge Cases, Not Models
Asked why AI is not everywhere already — self-driving cars, agents doing useful work — he gave one answer for all of it. "I think the answer to all of those questions about why isn't AI more widespread is the edge cases."
Smith's version of the question went back to Sebastian Thrun and the 2007 challenge, and asked why autonomous driving is still not in New York
The reason edge cases bite less in warfare is that the bar is so low. On the share of front-line FPV drones that actually hit and detonate: "Well, it's between 10 and 20%."
"And that's why drones like ours are currently used in the front line and they increase this efficiency by two to four times even with level one autonomous system."
The staffing figure is the one that undercuts the sector's marketing. "So from the outside because AI is such a val buzz word it might seem that drone company that does AI has 80% of its staff or aging staff doing AI but it's more like 10%."
Most of the work, he said, is software and robotics rather than models
Navigation is not solved, and he objected to the premise that it is. It is not one problem but a hundred: a $500 drone flying 40 kilometers at up to 100 kilometers an hour, a subsonic winged missile and a hypersonic vehicle are different problems, and altitude, terrain and flying over water change all of them
"And there won't be like one AI Google Maps that solves it for anything"
Target recognition is closer to solved, with the same caveat. "It's easy to spot a green tank on a white snowy fields. It's much more difficult to spot something camouflaged then you see Russians riding on donkeys. Like was that in the data set? Probably not."
"Then you see Russian Russians dressed up like penguins. Was that in the data set? Also not really."
The practical problem is not the algorithm: how low and how fast the drone flies changes what it can see, how recognition integrates with navigation, and what tactics the whole thing serves
15. Terrifying If They Have It
Smith put the humanitarian objection directly, and got the shortest answer in the interview. "Oh, no. Like it's terrifying if the enemy has it. If we have it, it's a savior."
The reason autonomy on one side forces it on the other is mass and speed. Autonomous interceptors are the only answer to autonomous attack drones, because the numbers deployed are beyond what people can handle and the speed gap keeps widening
What he is actually afraid of is the schedule. "So I am what I'm terrified about is that we won't get that we won't get there as fast as the enemy does."
Bonus Insights
What The Fourth Law ships in volume today is narrower than the framework suggests. "But what we ship currently in mass is sort of level one autonomy for ground interception and we call a level zero for aerial interception which is a detector as the AI detector which helps pilot to see the target but doesn't really take over the control of the ground"
Several further systems are with the military for early testing, and he said he could not discuss all of them
The company also licenses the modules and the software to drone manufacturers rather than only making its own aircraft
He cited a line from Vladimir Putin as evidence of intent rather than capability. "And then remember in 2017, Vladimir Putin said that whoever will master AI will rule over the world."
"Will rule over the world will not cure all the diseases will not save the humanity."
His assessment of the follow-through was that Russia has lost most of what it inherited, including space technology, and that the country is destroyed by corruption and incompetence but still able to do a great deal of harm
A digression on the Soviet space program turns on a Ukrainian engineer. After the shuttle program ended, American astronauts flew on a vehicle Russia operated but Sergei Korolev designed, and Azhnyuk claims Korolev as Ukrainian, down to the Ukrainian song that was the first sung in space
On why Schmidt's company avoids the press and then does not, he had no explanation beyond the lawyers: long silences broken by large public announcements, including a deal signed with Volodymyr Zelenskyy present and, later, a deal with Poland
Azhnyuk's bottom line is that autonomy is a manufacturing and software problem rather than an AI one, and that the side which industrializes it first sets the terms for everyone else — which is why he treats the question of whether autonomous weapons are acceptable as already settled by the people who will not be asking.
Products, Companies & Tools Mentioned
The Fourth Law (His company: autonomy modules and software for drones, shipped at level one for ground interception and licensed to manufacturers)
Petcube (The connected pet-camera business he co-founded, and the source of the engineers who started his thermal-camera company)
Swift Beat (Eric Schmidt's drone company, the competitor he takes most seriously and the only Western operation whose equipment he says works in the field)
D3 (Schmidt's defense-tech incubator, which he helped start; about half the portfolio is Ukrainian and the fund is small)
Anduril (He has respect for it and wants to know why its systems are not fighting in Ukraine)
Wild Hornets (Maker of the Sting interceptor, which he calls the most successful in the category)
DJI (The world's leading drone technology company, and the reason he thinks China could scale autonomous drones fast)
Nvidia (The company he says benefits most from sovereign AI, because every national model still buys chips)
ChatGPT (The reference point that misleads people about what a drone's AI actually is)
Y Combinator (Where he learned the customer-development advice he says explains Schmidt's results in Ukraine)
NATO (Whose secretary general called the Polish interception a job well done, which Azhnyuk disputes)
McDonald's (His illustration of why the factory, not the product, is the hard part of a drone business)
Books & Resources Mentioned
AlexNet (His marker for how old image recognition is — the thing now sold as AI for defense)
The National Security Commission on Artificial Intelligence (The Pentagon-funded body Smith worked on and Eric Schmidt chaired, and the source of the American position on lethal autonomy)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: