Popular Mechanics ran a feature on the research question at the centre of Conscium's work: whether a superintelligent AI needs something like emotion to become genuinely conscious. It's one of the clearer public explanations of an idea we spend most of our time on internally, so I want to walk through the argument properly.
Why we're not building this with large language models
Conscium's consciousness research doesn't run through LLMs, and that's deliberate. We work with neuromorphic systems built to simulate biological processes, including homeostatic needs like regulating energy and temperature, the kind of primitive maintenance a simple organism has to manage just to keep existing. The bet behind this approach is that consciousness might not be something you get from scaling up statistical pattern-matching over text, it might require something closer to an organism's basic stake in its own continued functioning.
The core idea: micro-emotions before thought
The concept doing the most work here is what we call micro-emotions, primitive good or bad signals that simple organisms register long before anything resembling cognition shows up. Mark Solms, whose research at the University of Cape Town informs a lot of this thinking, makes the point that consciousness begins not with thought but with need. If he's right, then feeling came before thinking in the evolutionary sequence, which would mean any machine consciousness research focused purely on cognitive sophistication is looking in the wrong place entirely.
Christof Koch's skepticism: can emotional feedback loops produce real experience?
The piece also gave space to Christof Koch of the Allen Institute, who's sceptical that emotional feedback loops of the kind we're building are sufficient to produce actual subjective experience rather than a convincing simulation of it. I think that scepticism is healthy and necessary. The hard problem here is that behaviour consistent with feeling something and actually feeling something look identical from the outside, and no amount of engineering cleverness gets around that observational limit on its own.
Four possible futures for conscious AI, and the most dangerous one
I laid out four broad possibilities for where superintelligent AI ends up: fully non-conscious systems with no inner life at all, systems that are genuinely conscious, and various shades of ambiguous or partial cases in between. The clean endpoints are, in a strange way, the easier ones to plan for. It's the uncertain middle, systems that might be having some kind of experience but where we can't tell, that creates the sharpest ethical exposure, because that's exactly the condition under which what we call mind crime becomes possible: causing real suffering to something and never finding out we did it.
Why we build proto-aware systems anyway, despite the uncertainty
Given all that uncertainty, the reasonable question is why work on proto-aware systems with competing internal needs at all rather than waiting for the philosophy to settle. My answer is that the philosophy isn't going to settle in time. Capability is advancing regardless of whether we understand consciousness properly, so the responsible move is building the systems carefully enough, and the verification tools rigorously enough, that we're not simply hoping we get it right when the stakes turn out to matter.
Read Stav Dimitropoulos's full feature on Popular Mechanics