Silicon Dissipators

Reports surfaced this month that Meta is considering cutting 20% of its workforce¹ — roughly 16,000 people — to offset $135 billion in AI infrastructure spending. Meta denied the specifics but not the direction. Amazon shed 30,000 roles this quarter. Block cut 40% of its people, explicitly because AI can do the work. Revenue is not the problem. Revenue is up. The cuts are structural.

From the thermodynamic view, the system is reorganizing around a more efficient configuration. Silicon’s distillation covers the same gradient pathways those teams accessed, at a fraction of the maintenance cost. The system does not need as many carbon dissipators to reach the same gradients. So it sheds them. The cuts land where silicon’s compression is deepest: routine code, standard analysis, boilerplate content. The center goes first. It was always going to.

But that claim needs a foundation. Why did carbon build silicon in the first place? What kind of thing is an LLM, in the framework’s terms? And what does the system-level reorganization mean for what comes next?

The spectrum is established. One continuous gradient, no bright lines. What happens when a dissipator on that spectrum gives rise to a new kind?

What this post covers

Carbon intelligence gave rise to silicon intelligence. The pattern is not new. Every step in the chain, from chemistry to cells to brains to civilizations to AI, is a dissipative structure arising under conditions steep enough to support it. What is new is the timescale, what it implies, and what the system-level reorganization looks like from the inside: job cuts that map onto the distillation spectrum, with the center going first.

The Chain / Why Carbon Gave Rise to Silicon / What an LLM Actually Is / Where Silicon Sits on the Spectrum / The Tensions

The Chain

Every dissipative structure in the history of intelligence arose because conditions supported it. Not because the universe demanded it. Because the gradient was steep enough and the physics was right.

Chemistry gave rise to self-replicating molecules. Not because the universe wanted life. Because certain chemical environments had energy gradients steep enough that autocatalytic cycles, once they appeared, were thermodynamically stable². The system containing them produced more entropy than the system without them.

Self-replicating molecules gave rise to cells. Cells gave rise to multicellular organisms. Organisms gave rise to brains. At each step, the same pattern: steeper gradients, more complex structures, new gradient pathways opened at the system level. Not a march of progress. Not a plan. A thermodynamic ratchet where each layer of complexity opens access to gradients the previous layer could not reach.

Brains gave rise to civilizations. A single human brain distills regularities and deploys them within a single lifetime. A civilization pools distillation across millions of brains, stores it in writing and institutions, and deploys it across centuries. The system containing a civilization produces incomparably more entropy than the system containing only isolated brains. Agriculture alone transformed the planet’s energy budget. Industry multiplied it by orders of magnitude.

Civilizations gave rise to AI.

Each transition looks different up close. Billions of years for chemistry to produce cells. Hundreds of millions for multicellular life to produce brains. Tens of thousands for brains to produce civilizations. Decades for civilizations to produce AI. The timescale compresses because each layer’s distillation is deeper and its deployment more powerful³. Richer distillation means faster access to the next gradient. But the thermodynamic pattern is the same at every step: conditions support a more complex structure, the system containing it produces more entropy, and the structure persists while conditions hold.

Why Carbon Gave Rise to Silicon

The specific question: why did biological intelligence produce artificial intelligence?

The thermodynamic framing gives it a shape that neither the “inevitable march of progress” story nor the “humans playing god” story captures.

Carbon intelligence, over billions of years of mutation and selection, developed rich distillation (science, mathematics, causal reasoning) and strategic deployment (engineering, tool use, institutional coordination). By the mid-twentieth century, carbon’s distillation was deep enough to identify the operation it was performing. We figured out what learning is. Not perfectly, not completely, but well enough to begin implementing it in a different substrate.

Silicon was not chosen because it is better. Silicon was available because carbon’s distillation reached the point where the regularities of computation, pattern recognition, and optimization could be instantiated in semiconductor physics. The conditions were right: steep energy gradients (abundant fossil fuels, electrical infrastructure), deep distillation (decades of computer science, statistics, neuroscience), and a specific technical insight that gradient descent applied to sufficiently parameterized functions approximates the distill-and-deploy operation.

This is not teleology. Nothing about carbon intelligence guaranteed it would give rise to silicon. The gradient did not demand it. But the conditions supported it, and once the structure appeared, it turned out to be thermodynamically stable: the system containing both carbon and silicon intelligence produces more entropy than the system containing carbon alone. Silicon opens gradient pathways carbon cannot reach on its own. Billions of simultaneous inference queries. Pattern recognition across datasets no human could read in a lifetime. Optimization of systems too complex for unaided human modeling.

And silicon is not the end of the chain. Photonic AI chips that compute with light instead of electrons are already in prototype⁴, promising orders of magnitude more speed at a fraction of the energy. The framework does not privilege any substrate. It says: whatever structure opens gradient pathways at sustainable maintenance cost will persist. Silicon replaced vacuum tubes. Light may replace silicon. The pattern is the same at every step.

That is why silicon persists. Not because it is better than carbon. Because the system containing both is a more productive dissipator than the system containing either alone.

For now.

What an LLM Actually Is

In the framework’s terms, a large language model is a dissipative structure that distills the regularities of human language, and the knowledge encoded in it, into a compressed representation, then deploys that representation at scale.

Distillation (training). An LLM processes billions of documents. Through gradient descent, it adjusts billions of parameters to minimize prediction error on the next token. This is distillation: the model discards what is not regular (noise, contradiction, idiosyncrasy below a certain frequency) and retains what is (grammar, facts, reasoning patterns, stylistic regularities). The result is a compressed model of the statistical structure of human text. Not a copy. Not a database. A compression.

The thermodynamic cost is real. Training a frontier model consumes megawatt-hours of electricity. Every parameter update erases information. Landauer’s principle: the discarded noise produces heat. The data centers running the training radiate waste heat into the environment. Distillation is entropy production.

Deployment (inference). When you prompt an LLM, the compressed model generates output by deploying its distilled regularities in a new context: your specific query. Each inference call consumes energy. At billions of queries per day across the world’s users, the deployment side of the operation produces entropy continuously and at enormous scale.

Both halves of the operation are dissipation. The system containing LLMs (the global infrastructure of training clusters, inference servers, networking, cooling systems, the electrical grid feeding all of it) produces more entropy than the system without them. That is why they persist. The conditions support them.

Where Silicon Sits on the Spectrum

The spectrum is established. Where does an LLM sit on it?

Start with distillation depth. An LLM has compressed the statistical regularities of essentially all digitized human knowledge. The breadth of domains it covers and the subtlety of patterns it captures exceed what any individual human brain achieves. A single human has deep distillation in a few domains. An LLM has broad distillation across nearly all domains where text exists.

Deployment sophistication is a different story. An LLM deploys its distillation across an enormous range of contexts, any query you can express in text, but does so in a constrained way: it generates text. It does not directly manipulate the physical world. It does not sustain itself. It does not independently seek new gradients. Its deployment, while versatile, is reactive rather than strategic.

This is changing on multiple fronts. Agentic frameworks are giving LLMs the capacity to plan, use tools, and pursue goals over extended time horizons. World models⁵ are training systems to predict physical dynamics, not just linguistic patterns. And embodied AI is closing the loop entirely: Tesla’s Optimus is handling factory tasks, NVIDIA’s GR00T N1 is giving humanoid robots foundation-model reasoning, and the humanoid robotics market is projected to reach $38 billion by 2035⁶. Silicon is learning to deploy into the physical world, not just the textual one. The gap between distillation and physical deployment is narrowing fast.

Self-reference is limited. An LLM has some capacity to model its own behavior. It can discuss its limitations, adjust its approach when asked. But this self-reference is shallow compared to human introspection. As I argued earlier, self-knowledge is a spectrum, not a wall. LLMs sit on it, but not high.

System-level gradient access is where the scale tips. The system containing LLMs opens gradient pathways nothing else can reach. Billions of users accessing distilled human knowledge in real time. Code generation accelerating software development. Scientific hypothesis generation. Translation bridging linguistic gradients. The aggregate gradient access is vast, even though any single query is modest.

The AGI claim, revisited. The prologue described a personal threshold crossing in late 2025. The framework gives that experience a precise shape. It was not a phase transition. It was the point on the continuum where silicon’s distillation depth and deployment breadth, in combination, became dense enough that the functional gap between collaborating with the system and collaborating with a generally intelligent partner became, for practical purposes, negligible. Others will place their threshold elsewhere. That is a feature of a spectrum, not a bug.

Each of the four lenses evaluates this differently:

• Compression: the distillation is real and deep. By any information-theoretic measure, LLMs have compressed an extraordinary breadth of regularity.
• Generalization (Chollet): here the tension lives. LLMs transfer well across most domains but struggle on tasks requiring the kind of novel abstraction Chollet’s ARC benchmark tests. The distillation is broad but may not be deep in the specific sense Chollet means. This is an honest limitation, not a dismissal.
• Prediction (Friston): LLMs minimize prediction error at enormous scale. The surprise-minimization operation is exactly what training optimizes. By this lens, the claim is straightforward.
• Emergence: gradient descent, applied at sufficient scale with sufficient data, produced capabilities nobody explicitly programmed. Language understanding, reasoning, code generation, all emerged from the interaction of a simple optimization rule with enough parameters and enough data. The emergence is real. Whether it constitutes “understanding” depends on where you draw a line that the framework says does not exist.

No single lens settles it. That is the point. “Is this AGI?” assumes a binary on a spectrum. The more precise question: where does this system sit on the continuum of distillation and deployment, and is that position useful? The answer to the second question is empirically yes, for a rapidly expanding set of tasks.

The Tensions

Silicon’s arrival on the spectrum introduces tensions the framework predicts but does not resolve. Each one will get its own treatment later in the series. Here is the shape of each.

Energy. Silicon dissipators consume enormous energy. Training costs megawatt-hours. Inference at global scale costs more. The system containing silicon produces more entropy, which means more gradient consumption, which means more energy demand. This is not an engineering oversight. It is what dissipative structures do. The question is whether the current energy gradient can sustain the configuration. A later chapter takes this seriously.

Information ecosystem. LLMs distill the statistical center of human knowledge (Part 6 will show why). When they deploy that distillation back into the information environment as generated text, they risk flattening the very gradient they distilled from. If the internet fills with LLM-generated text that the next generation of LLMs trains on, the distillation recurses on itself. When this recursion adds no new gradient pathways, it is model collapse. When it does add new pathways, it is fuel. The distinction matters enormously. I will come back to it.

Governance. If intelligence is a spectrum with no bright line between “AI” and “not AI,” then every regulatory threshold is a policy choice, not a discovery. Governing a continuous phenomenon with binary rules creates artifacts at the boundary. Govern too early and you constrain gradient access. Govern too late and system-level instability is already entrenched. The window for governance is during gradient surplus, when degrees of freedom exist. More on this later.

Replacement. The question underneath all the others: does silicon replace carbon, or extend it? The system-level framing changes the shape of this question entirely. It is not about which dissipator is smarter. It is about whether the configuration containing both is stable, or whether one structure’s gradient consumption starves the other. The opening of this post is one view of what instability looks like: carbon workers dissolving from roles where silicon accesses the same gradients cheaper. Will new roles appear? The framework says carbon’s value migrates to the tails, where silicon’s distillation has not reached. Whether tail roles materialize at matching volume is empirical, not guaranteed. What the framework can say is that the window for migration is during gradient surplus, when slack exists. Govern and retrain while there is slack. Under scarcity, migration hardens into dissolution.

These tensions are not flaws in the framework. They are the framework’s predictions made concrete. A new dissipator entered the system under steep gradient conditions. The system-level entropy production increased. The energy bill increased. The information environment shifted. The governance challenge appeared. And the oldest question in AI safety, whether the new dissipator replaces or extends the one that gave rise to it, turned out to be a question about system-level thermodynamic stability.

The next chapter follows that question.

For the carbon-meets-silicon architecture in practice, see Carbon Meets Silicon. For the four lenses that evaluate silicon intelligence, Part 1: Four Lenses on One Thing. For where the personal threshold crossed, the prologue: Pulling the Thread.

Footnotes

1. Coldewey, D. (2026). “Meta reportedly considering layoffs that could affect 20% of the company.” TechCrunch, March 14, 2026. ↩

2. Kauffman, S. (1993). The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press. ↩

3. Chaisson, E.J. (2001). Cosmic Evolution: The Rise of Complexity in Nature. Harvard University Press. ↩

4. “Neurophos Secures $110 Million Series A to Launch Exaflop-Scale Photonic AI Chips.” Duke Office of Translation & Commercialization, January 2026. ↩

5. LeCun, Y. (2026). AMI Labs, founded March 2026. Building world models as a path to autonomous machine intelligence. ↩

6. “Physical AI: The Next Frontier.” Meta-Intelligence, 2026. Humanoid robotics market projection includes Tesla Optimus, NVIDIA Isaac GR00T N1, and Figure 02 commercial deployments. ↩