“Extending our Perception” | February 2026
The Advanced Research and Invention Agency (ARIA) [the UK’s version of DARPA] was set up 3 years ago to help find and develop commercial uses for frontier research.
I’ve been thinking about how one might approach the “Extending our Perception” pillar, which is founded on the notion that “our AI models are dominated by human-centric data.” The document outlining the research area begins by asking whether the goal of AI is “to mimic humans or extend our capabilities”, and moves on to propose that by “co-designing AI with dynamic sensor systems [we may] reveal what remains invisible, catalysing breakthroughs in health, scientific discovery, and agriculture.” [The pillars are intentionally vague to allow various interpretations that would surface a breadth of ideas and solutions.]
It probably makes sense to start at Q4.
Q1: Are signal relationships objective or species dependent, and do we need to model human perception first?
Signals in the physical world often occur together but their co-occurrence is context dependent rather than a fixed universal relationship. For example, fire produces both light and sound, but an electromagnetic wave at a given frequency does not necessarily accompany a pressure wave at any particular frequency. These are distinct and the fact that they occur together in a given event is not a hard rule for how those signals are bound. But can this still tell us something useful about their relationship i.e. if you reliably find signal X in a given context, will you always also find signal Y (and is that relationship proportional--not necessarily in a linear sense, but in a way that is still measurable and predictable)? If so, is that because a rule exists or does it just tell you something about the actual context, or just about the kinds of events that happen to produce them both?
Do we even need a pre-existing model to make sense of those relationships? If so, would human psychophysics be a good place to start, with the idea that building out a formal model of how humans perceive the signals around them would eventually enable us to layer on additional signals that we don’t natively perceive (like ultraviolet)? Would anchoring to human perception be helpful for generating useful associations, separating correlation and causation, and reducing overall noise--or would it introduce bias and cause us to miss relationships that matter to other species but not to us?
Are the associations humans draw between co-occurring signals objectively true or are they species dependent? As in, when a species interprets a pressure wave logarithmically, does that indicate that sound (pressure) behaves logarithmically in the physical world, or is that scaling imposed by the perceptual system of each specific organism? And if the latter is true, might two separate organisms experience the same event by applying fundamentally different mathematical transformations to the same physical reality such that the relationship between signals doesn’t just _feel_ different, but is structurally modeled differently? And what would be the implications of this, if found to be true?
Q2: Do different species experience different physics?
Humans are highly visual and relatively less sensitive to sound. Does the particular bucket of signals an organism perceives--and its relative sensitivity to each--mean that it is effectively operating in a different physical world?
This question, in part, arises from the idea that signals are physical in nature. Sound, for instance, is a pressure wave. So, in a way, we literally ingest and interpret raw physics. Which makes it especially interesting to ask whether physically grounded signals (as opposed to chemical ones--though maybe physics and chemistry are ultimately the same thing. ah!) might be encoded in structurally different ways in across different organisms. And not just with different intensities, but with different underlying shapes e.g. logarithmically (like how humans do with loudness and brightness as per the Weber-Fechner law--as gestured towards above in Q1) while other organisms may encode them differently (the shape of Stevens’ Power law.) If two organisms are both responding to the same pressure wave but one encodes it logarithmically and another with a different exponent or scaling, then:
a/ are they doing different mathematics on the same physical world, or
b/ is the physical world not actually experienced as universal across biological systems?
Are there organisms for whom gravity feels weaker, or for whom the proportional relationships between forces are experienced differently? And if so, is this connected to psychophysics, as in, is the way a species sensory system is tuned a kind of implicit model of the physics most relevant to its survival such that different species are, in a subjective yet meaningful sense, living in a world governed by different physical realities (not just perceptual ones)?
Q3: Could we model physical environments in the same way we model perception (like with the Weber-Fechner law and Stevens’ Power law for humans)?
Psychophysics gives us a formal model of how humans transduce physical signals into perception (not completely accurate ones, but at least they give us something relative to go off of). These change based on the environmental context e.g., cold, still air allows sound to travel further and more clearly than warm air, so, in the cold, we may hear sounds that would otherwise be too quiet for us to perceive when it is warmer. Could we do the same thing as we do for perception/psychophysics/organisms, for environments, and place different organisms into different environments to better understand their interpretation of “reality”? If we modeled “underwater” and “above water”, for example, we may better understand the acoustic reality that whales experience.
Q4: Question about the questions...
It seems like there are (at least) two ways to approach the ARIA pillar. My question is whether they’re actually different--does it make more sense to approach them individually (as I have) or to put them together?
The first approach is expanding the range of signals we are sensitive to e.g. incorporating an ultraviolet reality into our model of the world. This approach can be summarised as: adding a new channel of information.
The second approach would be to focus on improving the associations we draw between the signals we already receive e.g. using a haptic belt to learn where magnetic north is. In this case, we are not actually gaining access to a new signal (the information was always there), but rather, we are redirecting our attention toward the signals we already receive so that we may begin to form associations that enable us to perceive them in a meaningful way.
Are “the acquisition and incorporation of previously undetected signals” and “perceptual association-building” actually two different things--such that, tackling each task should be a separate matter--or are they part of the same underlying process, but merely at different stages?
Q5: Can we mathematically model threat perception? In humans, what proportion of those thresholds are universal vs personal (and learned).
When an object is tilting toward us or accelerating in our direction, this can trigger our “looming” response. At what angle, and at what velocity, does something that is neutral begin to feel threatening? Is there a mathematical relationship between those variables that predicts when a person will perceive an event as a threat? And if you were to model that relationship, would the model hold for everyone, or does it change based on an individual’s previous experiences--and by how much? Are differences in how an event is perceived traceable to specific experience, like PTSD triggers, or have they not yet been explained?
If you want to talk about these, please reach out. :)