When asked if he believed in God, Albert Einstein gave an answer that startled both theologians and atheists. ‘I believe in Spinoza’s God’, he replied, ‘who reveals himself in the orderly harmony of what exists.’

Born into a Jewish community in Amsterdam in 1632, Baruch Spinoza was excommunicated in his twenties for denying that God was a transcendent creator intervening in the universe. God, he argued, is not an all-powerful deity standing apart from Nature, but is identical to it (Deus sive Natura). This idea – often described as pantheism – was radical for stripping divinity of personality and intention. Spinoza’s God is impersonal: an infinite, self-sustaining substance moving with mathematical necessity. From this source flow all things external – a blade of grass, a planetary orbit – and internal: thought, emotion, even the sense of self. Individual beings, in this view, are not independent entities but temporary configurations within an eternal, unified reality.

Today, in the glass-walled offices of Google DeepMind and OpenAI, that 17th-century ‘orderly harmony’ has found an unexpected afterlife. It is hardly surprising that Demis Hassabis, the founder of DeepMind, has described his own worldview as Spinozan.

Though the vocabulary has shifted from divine attributes to scaling curves, the architects of artificial intelligence are building a fundamentally Spinozian project: a system designed to decode the infinite attributes of a single, universal substance – information. AI engineers are implicitly operating within this metaphysical framework. The assumption behind scaling is deceptively simple: if you provide sufficient data, parameters and compute, a coherent universal structure will emerge. Supply the system with enough information, and the world’s orderly harmony begins to reveal itself through statistical regularities.

Seen in this light, Large Language Models are not merely conversational agents. They are attempts to approximate the attributes of a single, all-encompassing universe where a token of information is akin to an atom moving through space.

If reality, as Hassabis suggested in a recent interview, is ultimately ‘computationally tractable’, then AI can, in principle, decipher the ‘Book of Nature’ by identifying the mathematical structures – learnable manifolds – behind phenomena such as protein folding or fusion plasma. Yet to perceive these structures, a model requires built-in forms of receptivity. And here, Hassabis’ other favourite philosopher, Immanuel Kant, becomes indispensable. According to Hassabis,’I think we will need some new philosophers.’ ‘This would be the perfect time for a new Kant to arrive.‘

Kant distinguished between the Noumenon (the thing-in-itself) and the Phenomenon (the thing as it appears to us). We can never truly know the world in its raw, unfiltered state. Instead, he argued, our minds act as a processing plant, filtering sensory noise through a priori categories such as Space and Time.

In AI terminology, raw data can be seen as the Noumenon. A trillion tokens of text or pixels are, in their unprocessed state, a meaningless deluge of binary code. The model can no more ‘see’ this raw data than a human eye can perceive infrared light. To render this chaos intelligible, the model must pass the data through its own Kantian filter: latent space.

To understand latent space, it is helpful to note what it is not. If you were to look inside the architecture of a model like Gemini or DeepMind’s AlphaFold, you wouldn’t find a neatly indexed library of facts. Instead, you would find a vast, high-dimensional mathematical vacuum, where every concept – from a protein fold to existential dread – is represented as a set of coordinates, also known as a vector. This latent space functions as the model’s own framework of intelligibility: the silicon analogue to a human mind.

Kant believed that humans are born with hard-wired structures – a priori knowledge – that allow us to intuitively experience the world. You don’t learn that things happen in Space and Time; you are born with these filters already switched on, ready to process the world.

In machine learning, this is what’s known as inductive bias. A system like AlphaFold, for instance, does not blindly guess atomic arrangements. The model is encoded with the constraints of three-dimensional space from the outset. Before encountering a single protein, it ‘knows’ that certain geometric relations are inviolable. It does not merely observe the world; it is built to perceive it through a specific, structured lens.

In other words, when engineers train advanced AI, they are not merely supplying it with information. They are constructing a transcendental schema in Kantian terms: a mathematical framework that enables the machine to translate the chaotic noise of a Spinozian universe into a structured, navigable reality.

Among Kant’s most profound insights was his explanation of how the mind unifies fragmented sights and sounds into a single, coherent object – a process he termed the Synthesis of Apperception. In transformer architectures, the attention mechanism performs a computational analogue. Presented with the sentence, ‘The bank was closed because the river flooded,’ attention weights the relations among tokens, disambiguating ‘bank’ to infer not a financial institution, but a geographical feature. Fragmented inputs are ‘synthesised’ into a unified Phenomenon.

Kant argued that reality, as we experience it, is inseparable from the mind’s structuring activity. Light, warmth and touch feel categorically distinct, yet, at a deeper level, they are all sensory signals processed by a common apparatus. In this respect, humans are themselves information-processing systems. The question at the heart of Artificial General Intelligence thus becomes unavoidable: what are the formal limits of computation? What are the boundaries of a Turing machine?

Spinoza held a deterministic conception of freedom. To him, free will was not the power to have acted otherwise; every event, he argued, unfolds from prior causes with a mathematical necessity. Freedom consists instead in the lucid understanding of those causes. This, for Spinoza, is the only thing that differentiates a human being from a stone, which, flying through the air, ‘would believe itself to be completely free, and would think that it continued in motion solely because of its own wish’.

For Spinoza, the human advantage is not to escape the laws of nature, but to recognise them and accept our place within this universal order. The more we align our lives with this rational understanding of cause and effect, the freer – and more fulfilled – we become.

AI systems remain closer to Spinoza’s stone. They traverse latent space, predicting tokens determined by weights and gradients, without awareness of the causal mechanisms directing them. Artificial General Intelligence would require the machine to move from blind inference to rational and reflective understanding – an ability to model and articulate the causal chains that produce its outputs.

A machine capable of charting its own internal determinants would achieve, as far as silicon allows, the Spinozian freedom of the human mind. Whether such a reflective intelligence is attainable remains the industry’s great unanswered question. Yet, in returning to the philosophers of the past, we are reminded that the ultimate frontier of AI is not technical, but metaphysical. The question is no longer how efficiently machines can compute our world, but whether they will ever possess the interiority to ask themselves why.