Recursive self-improvement, or RSI, is the idea currently dominating conversations in AI labs — the notion that once an AI system gets good enough to meaningfully contribute to AI research itself, it can help build a better successor, which in turn builds a better one still, and so on until progress starts moving faster than humans can track. But long before RSI became an industry buzzword, a philosopher was already laying out the argument on television — in 1996.
Philosopher David Chalmers, best known for his work on the “hard problem” of consciousness, recently resurfaced an old clip of himself making this exact case nearly three decades ago. The footage appears to be from a 1996 discussion on the Australian TV show Lateline, where Chalmers appeared alongside AI researchers Rod Brooks and Doug Lenat — a fact Chalmers himself has referenced in his later academic writing, noting that he “advocated the possibility vigorously” in that very discussion. What’s striking watching it back today is how closely his reasoning tracks the argument AI researchers are making right now, more than two decades before the term “recursive self-improvement” entered common use in the industry.
In the clip, Chalmers lays out the core mechanism in strikingly simple terms. Say you have a machine which is as smart as us, then maybe ten years later we get a little bit better at programming, the machine is just a little bit smarter than us. Then that machine in turn is probably going to be a better programmer than us — it’s going to be able to build a better machine than we can, so it’ll build a smarter machine yet. This gets into an ongoing spiral, so his feeling was that once we get to the point where a machine is as smart as us, the point where machines are thousands of times smarter than us is not going to be so far off in the future.
That is, almost word for word, the loop AI labs are describing today when they talk about an “automated AI researcher” — a system capable of designing the next system, which designs the one after that. It’s also the same basic logic that statistician I.J. Good first sketched out in 1965 with his idea of an “ultraintelligent machine”. But Chalmers wasn’t just describing a mechanism — he was already grappling with what it would mean for humanity to live alongside it, decades before questions about AI safety and alignment became a fixture of every major lab’s roadmap.
The interviewer pushed him on the human side of the equation: does this mean we’re all looking at having a big identity crisis in fifty years, feeling very insecure? Chalmers’ response was measured rather than alarmist. Maybe it’s not fifty years, he said — maybe it’s two hundred. But he argued that at some point we’re going to have to sit down as a species and ask whether we want this to happen. Maybe we’re happy to be the next step in evolution, arriving at a race of machines a thousand times smarter than us. Maybe we don’t want that to happen, in which case there’s a real Pandora’s box that we’re just going to have to try to keep closed.
Thirty years later, that framing feels less like a hypothetical thought experiment and more like a preview of the actual debate now playing out between AI labs, safety researchers, and governments. Anthropic and OpenAI have both started citing automated AI research directly in their safety frameworks, alongside risks like cybersecurity and bioweapons. Startups are now raising hundreds of millions of dollars specifically to build self-improving AI, and researchers at labs like Google DeepMind have said the industry has quietly moved past the point of treating this as a distant, speculative idea. Tech leaders have also floated their own timelines for when machines might overtake collective human intelligence altogether, with some, like Elon Musk, predicting AI could exceed the combined intelligence of every human on Earth by 2030.
Not everyone agrees the spiral will play out as cleanly or as quickly as Chalmers described it in 1996. Some researchers argue that the messiness of the real world — chip fabrication timelines, physical infrastructure, compute constraints — means a “hard takeoff” is less likely than the pure feedback-loop argument suggests. Even so, the core question Chalmers posed to a TV audience thirty years ago — whether humanity actually wants to build something that spirals past its own intelligence, and whether that decision is even still ours to make — is now sitting squarely in front of an entire industry, rather than a late-night philosophy segment.