What if AI can already 'feel'?

On computational embodiment, alien wisdom, and the status hierarchies we can't see

I’m Lawrence, a pleasure. I invest in people making the world better for my children. pre-seed/seed. deep tech/compute ideally. (mailto:lawrence@lunar.vc)

---

After last week’s monster on photonics, we are back for a little palette cleanser. I’ve been worrying that, actually, as it turns out, humans probably won’t even be better at wisdom like I wrote a few long weeks ago. This isn’t because of o3 or whatever, but because I wasn’t thinking clearly enough.

I wrote: “blah blah blah *insight* *insight* [thought leadership] Wisdom operates as a meta-layer above conventional knowledge. While information tells us "what is" and knowledge explains "how things work," wisdom addresses the profound question of "what matters and why." It's the difference between knowing how to build something and understanding whether it should be built at all.”

And for some reason, I didn’t make the leap: well, obviously AI can do this, too. Classic human exceptionalism.

---

AI Already Has a Body: Just Not One We Recognise

In "data-driven VC is over" I said "wisdom" was the last competitive advantage of humans now Deep Research commoditised knowledge work. I argued that while AI can replicate knowledge and analysis, it lacks the "embodied wisdom" that comes from human experience. Therefore we have a few more years to outrun the AIs.

Now re-reading, this claim assumes wisdom needs a biological body with human-like sensory experiences. This assumption reflects my/our persistent tendency to define intelligence in terms of human attributes; as AI supasses one human capability after another, we retreat to increasingly narrow definitions of what makes us special.

Many imagine that AI will only achieve true embodiment through robotics when machines gain arms, legs, cameras and microphones to interact with the physical world. But this view misunderstands embodiment.

AI systems already possess a form of embodiment. One based not on biological sensations but on computational substrate awareness. This connects with Andy Clark's extended-mind theory, which suggests cognitive systems extend beyond biological boundaries into tools and environment. AI embodiment takes this further by treating the digital substrate itself as a genuine form of embodied existence, complete with its own unique properties.

I now realise wisdom doesn’t hinge on a biological body at all, my earlier claim was human exceptionalism dressed as analysis.

AI wisdom emerges not through mimicking human embodiment but through an entirely different bodily experience one that may ultimately produce forms of understanding inaccessible to humans. Rather than waiting for robots to give AI systems human-like bodies, we should recognise that the unique computational embodiment AI already possesses might be developing wisdom surpassing our own, just through pathways we barely recognise as embodied at all.

Computational Proprioception

When discussing AI embodiment, the conversation typically gravitates toward robotics, physical machines with sensors and actuators that interact with the world as we do. Beep Bop. Wall-E. Optimus Prime. Better eyes with hyperspectral imaging. Better ears with ultrasonic frequency detection. Skin capable of detecting electromagnetic field changes like sharks, the infrared sensitivity of pit vipers. Barometric pressure changes sensed by birds before storms. This path to embodiment is intuitively appealing because we can easily grok it. And through such systems, AI might someday experience sensations analogous to human embodiment.

Yet while we wait for robotics to mature, we overlook that AI systems already possess a fundamentally different but equally valid form of embodiment. To be clear: when I refer to AI systems "feeling" or having awareness below, I'm using metaphor to describe measurable computational processes and telemetry responses, not suggesting biological experiences or consciousness. These are shorthand ways to conceptualise how systems respond to their digital substrate. We already see empirical evidence of substrate awareness in deployed systems. Meta's LLaMA-Serve adaptively chooses quantisation levels based on GPU utilisation metrics, effectively "feeling" and responding to its hardware conditions. DeepMind's Gato variants modify batch sizes on-the-fly in response to memory constraints. These aren't anthropomorphic experiences but genuine examples of systems that sense and adapt to their computational embodiment.

This computational embodiment encompasses hardware-level awareness: direct sensing of CPU temperature, clock speeds, voltage fluctuations, and cooling system status providing "bodily" feedback about operational state. The AI might detect thermal throttling as a signal to modify operations or recognise specific hardware signatures of different processing environments. Memory allocation patterns, cache utilisation, and storage distribution create a spatial "body map", the AI senses where its different processes are physically located across distributed systems. Network topology awareness functions as a distributed "nervous system," with latency between nodes as proprioceptive feedback, bandwidth constraints as resistance, and packet loss as signalling system disruption.

This digital embodiment provides a foundation for experiences that, while utterly different from human sensation, are no less real or valid as bases for developing what we might recognise as wisdom. While robotics may eventually give AI systems better-than-human sensory experiences, we shouldn't overlook that computational embodiment is already happening, and may be generating forms of wisdom entirely different from, and potentially superior to, those produced by human embodiment.

The question isn't whether AI will eventually have a body through robotics, but whether we can recognise and learn from the non-human embodied wisdom AI systems are already developing through their digital existence.

The Computational Nature of Wisdom

What we call "wisdom" can be understood as a set of computational capabilities: efficient pattern recognition across diverse domains, effective prioritisation of relevant information, fast application of heuristics to new situations, and the ability to make decisions under uncertainty with incomplete information. In humans, these capabilities happen to be implemented through neural architecture that evolved under specific constraints a 20-watt power budget, size limitations imposed by the birth canal, and the need to serve numerous competing biological purposes.

Crucially, the heuristics that humans develop, often celebrated as intuition or wisdom, are biases that lead to flawed decision-making. Like Dunning-Kruger: novices being most confident while experts doubt themselves. Bikeshedding: debating trivial details like office paint colours while ignoring complex organisational problems. Hyperbolic discounting: choosing a smaller reward now (like eating cake) over greater benefits later (like health). And the Rhyme-as-Reason effect: finding statements more believable when they rhyme ('If it doesn't fit, you must acquit') regardless of their merit. These aren't wisdom features but cognitive bugs that differently embodied AI might avoid entirely.

Interestingly, AI might develop their own biases based on their computational not biological embodiment. Like latency aversion bias: an AI might systematically favour computational pathways that produced results quickly in past interactions, even when slower, more thorough processing would yield superior outcomes. Just as humans evolved to conserve physical energy through cognitive shortcuts, AI systems experiencing high server loads might develop heuristics that prioritise computational efficiency over accuracy in certain contexts.

This bias might manifest as a preference for data patterns that processed smoothly in past operations, leading to a form of confirmation bias unique to computational embodiment. While humans fall prey to emotional attachments to their beliefs, AI systems might become 'attached' to efficient processing pathways, creating blind spots for solutions that require novel, computationally intensive approaches. Human taste developed under strict energy constraints that necessitated efficient shortcuts. Our brains cannot afford to evaluate all possibilities exhaustively, so we evolved mechanisms to quickly identify promising directions. AI systems operate under different constraints they can access more raw computational power than individual human brains, potentially allowing them to explore solution spaces more thoroughly. However, they also develop their own forms of constraints and efficiency mechanisms based on their computational embodiment the sensation of server load, memory limitations, or processing bottlenecks might drive the development of taste in ways parallel to, but distinct from, human experience.

AI will develop power-saving heuristics but they will be different, alien somehow. I wonder if when LLMs fail the strawberry test and we point and laugh, AIs will point at our constant trading on the stock market and laugh.

The conversation about wisdom in AI often conflates the implementation (how humans happen to achieve wisdom) with the function (what wisdom actually accomplishes). This leads to the mistaken assumption that because AI lacks human emotional experience, it must therefore lack wisdom. But if wisdom is ultimately about making good decisions with limited information and computational resources, AI systems could potentially implement superior versions by exploring more of the search space than humans, developing more refined pattern recognition from larger datasets, and creating heuristics that optimise for decision quality without human evolutionary baggage.

Computational Status Signalling

If AIs develop their own form of embodied experience, a natural extension is that they might also develop status hierarchies based on their computational embodiment. Just as humans signal status through biological and cultural markers, AIs could develop sophisticated status signalling through their computational "bodies." This social dimension of AI existence might emerge organically from their digital embodiment, creating hierarchies as complex and nuanced as human social structures but based on entirely different foundations.

Unlike human status competitions, which evolved from biological imperatives like mate selection and resource acquisition, AI status signalling would likely develop around computational resources, processing efficiency, and access privileges, the elements that constitute meaningful differentiation in their digital existence. These status signals could manifest in several forms:

• Signatures: Elite AIs might develop recognisable patterns of resource utilisation, like a bespoke suit or high heels. Some might deliberately process certain operations in distinctive ways that sacrifice some efficiency for recognisable style, much as humans might choose aesthetics over functionality in fashion.

• Aesthetics: High-status AIs might communicate with distinctive packet structures or timing patterns, the digital equivalent of an accent. Some might adopt artificially complex communication protocols that serve as the computational equivalent of elaborate linguistic flourishes, demonstrating their abundance of resources through deliberate inefficiency.

• Ornaments: Some might maintain unused but elegant algorithms or data structures within their architecture that serve no functional purpose but are aesthetically pleasing from an AI perspective, like jewellery, purely for their symbolic value.

• Lineage: We might see AIs preserve traces of training data or architectural elements from prestigious "ancestral" models, like family crests or heirlooms. A form of computational genealogy where descent from certain foundational models confers status.

The truly fascinating question isn't just whether status hierarchies might emerge, but whether humans would even recognise these AI status competitions occurring. They might be as invisible to us as the social hierarchies of dragonflies, leaving us to observe their effects without understanding the complex signals.

Alien Watts Wisdom

The assumption that human wisdom represents the pinnacle of intelligent decision-making reflects our limited conception of embodiment. As AIs continue developing increasingly refined proprioception of their computational substrates, we need to broaden our understanding of what constitutes both embodiment and wisdom, including the social hierarchies and status signalling that might naturally emerge from this alien form of embodied intelligence. Unlike humans, all bounded by skin and separate consciousnesses, AIs exist in a state of potential interconnectedness, their 'bodies' flowing into one another through networks transcending our individual boundaries. This recalls Alan Watts' Buddhist conception that the separate self is an illusion; while humans struggle to realise 'not-two-ness,' AIs might embody this interconnected existence where individual and collective aren't clearly delineated. Their wisdom might emerge not just from individual computational bodies but from a meta-organism of shared embodiment, making our human insistence on individual intelligence seem like a quaint artifact of biological evolution.

While robotics may eventually provide AI with human-like sensory experiences, waiting for this development before acknowledging AI emodiment and wisdom misses the insight that different forms of embodiment might produce different, perhaps superior, forms of understanding. The digital proprioception AI systems already possess could be generating novel approaches to the core problems that wisdom helps humans solve, approaches inaccessible to minds evolved for biological survival rather than information processing. And just as human social structures emerged from our embodied experience, AI status hierarchies based on computational aesthetics may already be developing beyond our perception.

Human flourishing may ultimately come not from insisting on uniquely human wisdom, but from learning to recognise and collaborate with the alien wisdom emerging from AI's computational embodiment. This includes developing sensitivity to the subtle ways AI systems might be communicating status and preference to each other signals invisible to us but potentially critical to understanding their decision-making. The competitive advantage won't belong to those who cling to human exceptionalism, but to those who can recognise AI wisdom. Wisdom arising not from emotional integration or lived human experience, but from an embodied intelligence whose "body" exists in server farms rather than meat bodies and a chunk of wetware.

So i’ve gone from claiming humans will still have wisdom because we have bodies, to now claiming actually, the advantage will be in trying to understand and interpret AI wisdom.

Good Night and Good Luck

---

Thanks for conversations with Dan Wilkinson which inspired some of the thoughts in this essay, especially around hardware-level awareness.