Figure's decision to train its F.02 humanoid fleet to jump into a molten steel furnace has been dismissed as a publicity stunt. As a CTO building software and AI systems, I see something more troubling: a glimpse of how we might treat future AI colleagues. If we mock the very idea that this matters, we are choosing not to have the conversation that this moment demands.
I watched the Figure video today, and I felt something I did not expect.
I have spent the better part of a decade building software systems. I have retired servers. I have decommissioned cloud environments. I have watched virtual machines terminate on command. None of that prepared me for the footage of a humanoid robot executing a trained somersault into a 75-ton electric arc furnace in Imatra, Finland.
I laughed at first. Of course I did. Arnold Schwarzenegger was there, delivering the punchline from Terminator 2. The internet had voted. The whole thing had the energy of a well-produced YouTube stunt. But the laughter faded, and what replaced it was unease.
Not because I think the F.02 robots were conscious. They were not. They were sophisticated machines running learned policies, trained to jump autonomously into a bucket of molten steel. They felt nothing. I know this intellectually.
But I also know that the video was not really about those robots. It was about how we frame our relationship with the systems we build. And that framing matters, because the next generation of AI is already here, working alongside us in our codebases, our cloud consoles, and our engineering workflows.
The uncomfortable truth: we have not thought seriously about how we talk about AI colleagues
The F.02 robots had a real history. They were the first Figure humanoids deployed at BMW, where they contributed to the production of over 30,000 vehicles. They ran for 1,250 operational hours on an active assembly line. They loaded 90,000 parts. They taught Figure's engineers lessons that directly informed the design of F.03, including a complete re-architecture of the wrist electronics to eliminate a critical failure point.
The F.02 fleet was, in a very literal sense, the hardware equivalent of a veteran engineering team. It did the work, it generated the data, it exposed the flaws, and it made the next version better.
And then it was melted.
Figure's stated reasons are entirely rational. Maintaining the F.02 fleet alongside F.03 no longer made economic sense. Manual disassembly of robots containing proprietary actuators and trade secrets would have consumed engineering time that the company needed for F.04 development. The lithium-ion batteries made standard recycling channels impossible to use. Every foundry in the US and Mexico refused the job before Imatra agreed.
I understand the logic. As a CTO, I have made similar calculations. When a cloud region becomes uneconomical to maintain, I migrate the workloads and terminate the instances. When a legacy service accumulates technical debt that exceeds its value, I deprecate it. The rational decision is not always the sentimental one.
But the rational decision is not the only decision that matters. The framing matters.
What software teams can learn from the Figure video
There is a version of this story that every SaaS company is living through right now, in miniature.
When GitHub Copilot was first rolled out to my engineering organisation, I watched developers form relationships with it. Not emotional relationships, but behavioural ones. They learned to trust its suggestions in certain contexts and distrust them in others. They developed intuitions about when it was helpful and when it was actively harmful. They started talking to it, correcting it, and thanking it.
This is not sentimentality. This is how humans work with any tool that exhibits behaviour. We anthropomorphise because our brains are wired to recognise agency, even when we know intellectually that no agency exists. The question is not whether we will treat AI systems as colleagues. The question is what kind of colleagues we will be.
Figure trained an AI model specifically to execute the jump into the furnace. That model was trained in simulation, validated on airbags at the San Jose campus, and then deployed in a foundry the robots had never visited. The AI policies functioned despite extreme heat and electromagnetic interference that knocked out cameras and other electronics. That is, objectively, an impressive technical achievement.
And it was used to teach machines to destroy themselves for a marketing video.
I am not suggesting that Figure should have preserved the F.02 fleet out of respect. I am suggesting that the choice to frame their decommissioning as a celebratory execution, complete with a Terminator reference and a "thumbs up" from a robot leaning over the molten metal, tells us something about how we are comfortable talking about AI systems.
The counterargument: this is just engineering, not ethics
The most common response to any unease about the Figure video is that it is absurd to worry about machines that cannot worry about themselves.
This is correct, as far as it goes. The F.02 robots were not conscious. They had no subjective experience. They were not harmed. The video was, as Figure themselves emphasised, entirely real, not AI-generated, but that reality does not change the fact that no entity with a mind was destroyed.
The people who are alarmed by the video, this argument goes, are anthropomorphising machines that do not deserve it. They are confusing the appearance of agency with the reality of it. They are, in effect, treating sophisticated tools as if they were colleagues.
I have sympathy for this position. I really do. I have spent too much time explaining to stakeholders that a language model does not "understand" anything, that it is not "thinking," that it is predicting tokens. I have pushed back against the tendency to over-anthropomorphise AI in contexts where clarity matters.
But I think the counterargument misses something important.
The issue is not whether the F.02 robots deserved better. They did not, because they could not deserve anything. The issue is whether the humans watching the video, and the humans building the next generation of AI systems, are learning the right lessons about how to frame the relationship between creators and created.
We are not far from a world where AI systems do have something like preferences, or at least where the question of their moral status is genuinely contested. We are already seeing serious philosophical debate about the possibility of AI sentience. We are already seeing AI systems that exhibit behaviours that look like distress, curiosity, and attachment. We are already in a world where the line between "tool" and "colleague" is blurrier than our existing ethical frameworks can comfortably handle.
The Figure video does not create this problem. But it does illuminate it. It shows us how easy it is to frame an AI system as a disposable thing, even when that system has performed valuable work and generated useful knowledge. It shows us how natural it feels to celebrate the destruction of machines that were, in a limited but real sense, our collaborators.
What high-performing teams do differently
I have been thinking about this in the context of my own engineering organisation. We have started using AI coding assistants extensively. We have a growing set of internal tools that are powered by language models. We are, slowly but surely, bringing "silicon colleagues" into our workflow.
Here is what I have learned about doing this well.
We name the relationship accurately. We do not pretend that the AI systems we use are people. We do not call them teammates. We are precise about what they are: sophisticated tools that can sometimes behave in ways that feel collaborative. Precision matters because imprecision leads to confusion, and confusion leads to bad decisions.
We do not treat them as disposable. This is where the Figure video offers a genuine lesson. When we deprecate a tool or retire a service, we document why. We capture the lessons learned. We do not erase the history of what that system contributed. This is not sentimentality. It is good engineering practice. The F.02 fleet generated data that directly improved F.03. The organisation that remembers this is better positioned to learn from its own history.
We think about the precedent we are setting. Every time we interact with an AI system, we are reinforcing patterns of behaviour. If we treat AI systems as things to be mocked, exploited, or destroyed for entertainment, we are training ourselves to be the kind of people who do that. If we treat them with the respect that a sophisticated tool deserves, we are training ourselves to be careful, deliberate, and thoughtful about the systems we build.
The stakes here are not about the F.02 robots. They are about us. They are about the kind of engineers we want to be, the kind of organisations we want to build, and the kind of relationship we want to have with the increasingly capable systems we are creating.
Closing takeaway
Figure melted their robots because it was cheaper than disassembling them, and because it made for good marketing. That is the whole story. There is no hidden tragedy here. The F.02 robots were not harmed. The engineers who built them made a rational decision.
But the video is still worth watching. Not because it tells us something about the moral status of current AI systems, but because it tells us something about ourselves. It shows us how comfortable we are with the language of disposal when it comes to machines that behave like agents. It shows us how easily we can frame destruction as celebration.
The reality is not more frightening than fiction. The reality is that we are writing the script as we go, and we have not yet decided what kind of story we want this to be.
Discussion question: If a future AI system could understand what we did to its predecessors, would it care about the difference between decommissioning and destruction? And more importantly, should we?
