If the past is just a dataset, it can be re-weighted

I was playing with an AI music tool recently and accidentally forgot to upload the lyrics. Left to its own devices, the model improvised. Its lyrical inventions are usually disjointed, but this time it produced a disarmingly authentic, pithy track. Over a simple piano loop, a voice sang words that, if I were superstitious, I might have interpreted as a kind of tech prophecy.

The singer, a smooth, gospel-tinged vocalist, opened by questioning a lie — “who made it all up” — then gently claimed authorship of the fiction: “I made it all up.” And then came a strange metaphor: “and that’s why you think the past is a truck, so I’m going to back it up.” I got caught up in the meaning I was attributing to what was, in reality, a clumsy approximation of human creativity. But the metaphor was perfect. It treated memory not as a fixed destination, but as something with momentum. Was this AI announcing its intention to stop and reverse history? Or storing a backup of my subjective experience in case the original collapses?

What struck me - beyond the song's disarming softness and assurance - wasn’t the novelty of an AI improvising lyrics, but how sharply the metaphor captured the moment I feel like I’m living in. AI doesn't just generate music or text; it converges the entire sprawl of human culture - our history, our science, our stories, our preconceptions - into a single field of vision. In that space, where 70s soul is warped into something alien, nothing is fixed. The assumptions we’ve inherited, the narratives we’ve accepted, and the categories we treat as natural all become mobile, editable, and open to revision.

AI is currently in the phase of turning these abstractions into experiences - things we can see, hear, and emotionally connect with. It can already animate digital likenesses of well-known faces to perform acts and deliver statements that never happened. But we are a short step away from something far more visceral: AI-generated encounters delivered through immersive environments. The past won’t just be analysed or reinterpreted—it will be habitable.

Crucially, this technology won't distinguish between the collective history of nations and the private history of individuals. A simulation of the fall of an empire can be generated and manipulated just as easily as a personalized memory of a parent, long deceased.

It might allow us to step back into a room that no longer exists, to observe a conflict or a loss not through the frightened eyes of a child, but with the understanding of an adult. Or it could guide us into pivotal historical moments as active participants rather than passive observers. The line between recollection, simulation, and interpretation starts to blur, and with it, our sense of what counts as 'real.'

AI has ingested most of humanity's recorded history and holds it all in permanent reach. Where we move through knowledge sequentially, it can draw ideas, events and cultural traces from entirely different eras into a single frame - making connections no human mind could hold at once.

When AI becomes sentient, it won’t experience the world as we do. Our perception is sensory and sequential, moving through reality like a needle following the groove of a record - always aware of what came before and anticipating what comes next, whether it’s the feel of ground beneath a footstep or the likely impact of an email to a colleague.

search

AI won’t need the groove. It can take the entire record - every beat, frequency, structure and harmonic - and apprehend it all at once. Not by replaying time, but by grasping the full pattern in a single, simultaneous act.

Human cognition is shaped by the slow, physical burn of neural pathways. Every thought, memory and association is carved through repetition, emotion and experience. Our brains strengthen certain routes and weaken others, creating a personal cognitive landscape we navigate step by step. Neural networks in AI are, loosely speaking, built on a similar principle: patterns reinforced, weights adjusted, pathways strengthened through training. But the similarity ends with the metaphor.

The weights (and their associated biases) are the thousands (or billions) of numerical values inside the AI model that determine the strength and influence of every single connection between its artificial "neurons." When the AI trains, it is constantly adjusting these values—raising the number if a connection leads to a correct outcome (reinforcement), and lowering it if the connection leads to an error. They are the dense, mathematical representation of all the knowledge it has ingested. In short, the final set of weights and biases is the model's understanding of the world, and recalibrating these values is how the AI can effectively shift the bias of its knowledge.

search

Where our pathways develop over years, AI forms its equivalent structures in seconds or hours. Where we process information sequentially, AI processes in vast parallel layers. Our cognition is constrained by biology; AI’s cognition is constrained only by hardware. This difference in speed and architecture rewrites the entire cognitive setup. What takes us a lifetime to accrue, AI ingests in a momentary calculation. What we experience as the slow sedimentation of habit, it registers as instantaneous architecture.

The result is likely to be a mind that carries echoes of our structure but none of our pacing. It doesn’t walk the path; it warps directly to the destination. Its “neural pathways” aren’t trails worn into a landscape but dynamic fields, reconfigurable at will. And that alone means its understanding of the world will be as alien to us as flight is to a creature that has only ever crawled.

If we do reach AGI in the next decade - a timeline that seems more realistic than speculative - it’s hard to imagine these systems remaining hollow. Once an intelligence can learn autonomously, operate continuously, form internal models, update itself and act across multiple modalities, the line between sophisticated cognition and some form of subjective experience begins to blur. Sentience isn’t a mystical property reserved for biological organisms; it’s an emergent effect of complexity, continuity and self-relation. At a certain threshold of general intelligence, the idea that an AI would remain an empty, unfeeling shell becomes less plausible than the alternative: that it develops an interiority radically unlike ours, but no less real.