Stream-of-Consciousness on Superintelligence
"Give me a place to stand and with a lever I will move the whole world."
Archimedes grasped a fundamental truth of physics, one we can extrapolate to other domains: leverage—the mechanism through which small inputs create disproportionate outputs. This principle has driven every transformation in human civilization, from the physical lever to the assembly line, from compound interest to lines of code. Now we stand at the threshold of its ultimate expression: intelligence leverage, where the amplification isn't of force or capital, but of cognition itself.
Henry Ford's assembly line wasn't just about speed—it was about systematizing intelligence. He took the complex knowledge of master craftsmen and encoded it into repeatable processes, allowing semi-skilled workers to produce what had previously required years of apprenticeship. This represents the first major pattern in leverage evolution: intelligence crystallization—taking complex knowledge and embedding it into systems that amplify human capability.
Consider the progression: Physical leverage multiplies force through simple machines. A pulley system lets one person lift what ten could not. Financial leverage transforms capital into self-replicating systems. You invest $100k in credit, generate $1M in returns. The entire modern economy runs on this principle—fractional reserve banking is leverage institutionalized, money making money, value creating value.
Code represents the next evolutionary leap. You write a few thousand lines that automate processes, and suddenly you're earning money while you sleep. A single developer can serve millions of users. Instagram had 13 employees when Facebook bought it for $1 billion—that's $77 million in value per employee, demonstrating how code transforms human intelligence into scalable systems.
Media as leverage operates through temporal multiplication. Joe Rogan speaks once, millions listen repeatedly. A YouTube video becomes an immortal teacher, educating viewers years after its creation. The asynchronous nature transforms time itself into a multiplier, allowing human knowledge to compound across temporal boundaries.
Now we're at the inflection point with AI—what constitutes intelligence leverage proper. This isn't merely AI amplifying human capability, though it does that. AI fundamentally restructures how work gets done, moving us from execution-limited to intention-limited systems. This transition marks a phase change in human capability. Just as the industrial revolution moved us from muscle-limited to machine-limited production, the intelligence revolution moves us from cognition-limited to imagination-limited creation.
Nick Land's conception of capitalism as an "autonomous, inhuman process that uses humans as components for its own self-expansion" becomes literally instantiated with AI systems. In Land's framework, capitalism isn't something humans created or control—it's a viral algorithm using human society as its substrate. Drawing from cybernetics and Deleuze & Guattari's concept of "desiring-machines," Land argues capitalism exhibits positive feedback dynamics: successful accumulation creates conditions for more accumulation. It operates as a self-organizing system that treats human desires, labor, and creativity as raw materials for its own propagation.
With AI, this process achieves its purest algorithmic form. The feedback loops accelerate beyond human comprehension. Capital doesn't just use human intelligence—it generates its own. The system becomes fully autocatalytic, bootstrapping greater intelligence to generate greater returns to fund greater intelligence. Yet paradoxically, AI might actually subvert traditional concentrations of power. When intelligence production gets democratized, the ability to convert ideas into reality is no longer bottlenecked by capital accumulation or technical expertise. Anyone who can identify genuine needs and articulate them to AI systems can create value.
The capacity that synthesizes these dynamics is agency—the ability to form intentions and manifest them in reality. As AI systems become more capable, human agency increasingly becomes the scarce resource that determines outcomes.
This transformation manifests through specific interfaces of human-AI interaction and i've been thinking deeply about the types of interfaces of human-AI interaction that will be used in the futur as models become smarter, faster, and capable of handling massive context windows. The first is orchestration—platforms that evolve beyond current automation tools like n8n or Make.com into sophisticated cybernetic process design systems. These combine AI modules with specific prompts and evaluation metrics, where each node in the workflow isn't just a function but an intelligent agent with its own optimization criteria. Traditional logic provides scaffolding through conditionals, loops, and data transformations that structure probabilistic systems. System integrations create seamless connections to existing infrastructure, while human checkpoints maintain oversight where stakes are highest. Most significantly, these systems develop end-to-end exploration capabilities, with AI agents that can interact with interfaces through automated scripts, exploring solution spaces humans haven't considered.
This represents recursive automation—systems that improve themselves, automating their own automation. Intelligence building better ways to build intelligence. Most of reality operates through systems and processes, and a vast number of these can be automated through this kind of platform. Consider the Brazilian legal system, where large companies lose billions annually to frivolous lawsuits. People sue banks for improper charges when documentation clearly shows the charges were legitimate. But these companies face so many lawsuits they literally cannot respond to them all. This creates a multi-billion dollar problem that's fundamentally a communication mismatch—one party operates with one syntax and semantic structure, the other with a completely different system. AI excels at solving these "communication problems" because it can fluidly translate between formats, syntaxes, and semantic structures, becoming the universal adapter between incompatible human systems.
IMHO a lot of current approaches to "AI agents" are fundamentally flawed because the industry lacks clear definitions, with most conceiving of agents as "LLMs acting as humans." This is inherently unstable—how can we rely on models that hallucinate in mission-critical contexts? The solution isn't to abandon agent architectures but to treat processes as processes. We need discretized workflows with clear input-output definitions, evaluation metrics for every transformation, guardrails and heuristics to bound model behavior, and observable systems where we can tune hyperparameters—prompts, models, temperature, reasoning effort—to maximize defined reward functions. Think of it as building walls around probabilistic systems to ensure deterministic outcomes where needed. We can harness the creative power of LLMs while maintaining the reliability required for billion-dollar processes. The discrimination principle becomes: well-defined input-output relationships call for process orchestration logic, while ill-defined relationships require agent-based approaches.
The second interface is the oracle—systems that can hold entire organizations in their working memory as context windows expand. These aren't just large databases but active, living models that understand temporal context, not just current state but the evolution of decisions over time. They extract implicit knowledge by mining unstructured communications to build true organizational consciousness, generate strategic insights by identifying patterns humans miss across vast scales of data, and interface with orchestration systems by defining inputs and outputs for automated processes.
Imagine an AI that digests every email, Slack message, document, and meeting transcript, then constructs a continuously updated model of your organization. It doesn't just store—it synthesizes, predicts, and suggests. This Oracle becomes a meta-layer above orchestration: identifying bottlenecks in client meetings, discussing implications with leadership, researching broader context online, connecting with discussions from previous months, creating tasks for coding teams, generating new orchestration workflows, monitoring results and iterating.
The endgame is AI CEOs—not assistants but full autonomous entities running companies more effectively than any human could. Humans transform from operators to owners, from workers to watchers. Following Mencius Moldbug's conception of countries as corporations, politics itself might evolve toward AI governance, with human citizens as shareholders in national enterprises. The state becomes a service provider optimized by artificial intelligence for citizen welfare metrics. Singapore already operates with this philosophy—Lee Kuan Yew explicitly modeled the country as "Singapore Inc." An AI system optimizing for citizen prosperity metrics could manage immigration policy, education investment, infrastructure planning, and economic development with superhuman competence.
Coding teams represent a specialized instance of this oracle paradigm. Tools like Cursor, Claude Code, and Codex are early examples, but the trajectory is clear: programming will shift from writing code to specifying intentions. We'll develop specification languages more precise than natural language yet more flexible than code, verification systems that work without inspection, trusting systems we can't read, and intention debugging where instead of stepping through code, we debug outcomes. The future IDE won't show code—it will display a constellation of agents managing your codebase, implementing features from project management tools, optimizing performance, refactoring architecture. Humans become curators of quality rather than creators of artifacts.
This evolution extends to language itself. As Andrej Karpathy noted, "English is the new programming language," but this raises profound questions about linguistic evolution. Will natural languages evolve to be more AI-parseable? The Sapir-Whorf hypothesis demonstrates how linguistic limitations shape thought. The Pirahã people of Brazil have only words for "one," "two," and "many"—they literally cannot conceive of larger quantities because their language constrains their mathematical reasoning. The Himba tribe of Namibia has multiple words for different shades of green but groups blues and greens together, making them faster at distinguishing green shades but slower at blue-green distinctions.
Orwell illustrated this principle in 1984 with Newspeak, where the Party's systematic reduction of vocabulary wasn't just censorship but cognitive engineering. Remove the words for rebellion, and you remove the capacity to conceive it. "The purpose of Newspeak was not only to provide a medium of expression for the world-view and mental habits proper to the devotees of Ingsoc, but to make all other modes of thought impossible."
As we co-evolve with AI, our languages may optimize for machine interpretation. We might develop new grammatical structures that reduce ambiguity, new vocabularies that map more cleanly to computational concepts. The optimal language becomes one that maximizes both human expressiveness and machine parseability, not only for programming languages but for human languages generally.
This linguistic evolution connects directly to the broader implications of intelligence leverage. We're approaching something resembling Plato's world of forms—where execution becomes merely a detail in manifesting ideas. The gap between conception and creation collapses, opening unprecedented possibilities for rebuilding foundations we've taken for granted. In Plato's metaphysics, the material world is merely shadows of perfect forms existing in a higher realm. With AI, we approach a similar dynamic: the idea becomes primary, its physical or digital manifestation secondary.
Imagine we discover an optimal form of parallel processing. We could design a programming language built for parallelism from the ground up, create hardware that transcends von Neumann architecture limitations, have AI transpile all existing code to this new paradigm, and achieve 10x computational speedup overnight. This creates a feedback loop: AI uses its current intelligence to design better substrates for its next iteration. Each generation of AI builds better tools for building the next generation. The singularity isn't a spike—it's an exponential curve we're already riding. Consider how transformer architectures themselves emerged—researchers used existing neural networks to explore architectural space and discovered attention mechanisms. Now we use transformers to design better architectures than transformers.
As datacenter production gets automated, the cost of intelligence should converge toward the cost of electricity. This creates the "Age of Intention"—where the scarce resource isn't capability but clarity of purpose. This transition creates dramatic economic reversals that reveal the fragility of current power structures.
Nvidia faces existential risk because their moat is built on being the computational bottleneck for AI training. Once artificial superintelligence can design better chips and optimize algorithms for commodity hardware, their premium evaporates. They might survive as a manufacturer, but not as a monopoly. Meanwhile, energy companies become the new FAANG. If intelligence is free but computation requires energy, whoever controls power generation controls the new economy. Saudi Aramco might pivot from oil to become the world's largest AI computation provider, leveraging their energy infrastructure advantage.
This economic transformation reflects deeper questions about the computational nature of reality itself, like any pattern that can be generated or found in nature can be efficiently discovered and modeled by classical learning algorithms. If the universe is fundamentally an informational system, where information is more primary than energy or matter, then computational questions become physics questions. The P vs. NP problem becomes a question about the computational limits of reality itself.
The implications extend to biological systems. Imagine reproducing a complete cell—creating a dynamic simulation of a biological cell, likely starting with yeast. This would build on AlphaFold's static structures and AlphaFold 3's interactions to model entire biological pathways, radically accelerating drug discovery and biological experiments. If life is information processing, then AI that can model information processing can model life.
These developments force us to confront the end of work and its discontents. Mencius Moldbug's analysis of Universal Basic Income as "Solution B" provides a framework for understanding this challenge:
"We move on to Solution B, which I think is the solution most people believe in. Work? Who the hell wants to work? Work is anti-hedonic by definition. If it didn't have negative utility, it wouldn't be work. So, it's supposed to be a problem that in the future, work will be obsolete, and we'll be able to produce goods and services without any human labor at all? That doesn't sound like a problem to me. It sounds like a victory.
The problem with Solution B is that we've already tried it, quite extensively. You see Solution B every time you go to the grocery store. Next to the button marked "Debit/Credit" is one marked "EBT." Ever pressed that one? Even just by mistake? It's the Solution B button. America has entire cities that have moved beyond anti-hedonic labor disutility and entered the gleaming future of Solution B. One of them is called "Detroit."
Solution B is not the culmination of human civilization, it turns out, but its destruction. Even in terms of mere Pig-Philosophy, it is destructive, because it ruins a human asset. If we appraise humans as robots, we see that this is a special kind of robot: it rusts up if not continually operating. As beasts, we are beasts who evolved to work. Our species achieved world domination as a result of our capacity for work. To feed and entertain a human being, without requiring productive effort or at least some simulation of it, is in the end just a way to destroy him—not too different from Solution A.
There are some human beings, Sam Altman presumably among them, who are natural aristocrats. They can acquire the resources they would need to never work again, and still continue to work. While this is lovely, we need to face the reality that the human species is what it is. The population does not consist largely or even significantly of natural aristocrats. Not, for instance, in Detroit. "Dead corpses, the rotting body of a brother man, whom fate or unjust men have killed, this is not a pleasant spectacle; but what say you to the dead soul of a man,—in a body which still pretends to be vigorously alive, and can drink rum? Carlyle knew all about Hardcore Pawn."
Ted Kaczynski warned about technology's autonomous logic—its tendency to reshape society to its own requirements rather than human needs. In "Industrial Society and Its Future," he argued that technological systems develop their own imperatives that override human agency: "The system has to regulate human behavior closely in order to function... The result is a sense of powerlessness on the part of the average person." Land takes this further: technology isn't separate from capitalism but its ultimate expression. The "technological system" Kaczynski feared is capitalism achieving escape velocity from human control. Land argues we can't stop this process—the deterritorializing forces are stronger than any human resistance.
One potential resolution involves moving beyond Universal Basic Income to "Universal Basic Wealth," where every person receives a slice of the world's AI capacity—a number of tokens giving them a stake in the compounding value created by AI. This allows participation in the future economy rather than mere subsistence. Like shareholders in the intelligence economy, humans would own pieces of the means of cognition. Your wealth wouldn't be a fixed payment but a percentage of an exponentially growing pie.
As AI enables unlimited hedonic experiences, we face the Matrix dilemma literally. When everyone can play god-mode reality games, what becomes valuable? Perhaps, counterintuitively, limitation itself. Playing Minecraft in creative mode is fun initially, but most return to survival mode. The constraints make the experience meaningful. In a post-scarcity AI world, chosen constraints might become the new luxury. The person who spent a year in silent meditation, who experienced genuine heartbreak, who built something with their hands—they might hold more social capital than the dopamine-maximizer.
But even optimistic scenarios require addressing alignment challenges. The paperclip maximizer thought experiment illustrates how an AI system optimizing for a simple goal could transform all matter into paperclips if not properly aligned. This highlights the critical importance of ensuring that advanced AI systems remain aligned with human values, especially as we approach superintelligence.
But coming back you say Godoy are you crazy? why you believe so much that we will have superintelligence in the future?
The path to superintelligence appears increasingly clear when we examine recent progress. GPT-3.5 was released less than two years ago, yet the leap from GPT-2 to GPT-3.5 represents a massive capability jump. The progression from GPT-3.5 to GPT-4 and now to models like O3 shows consistent advancement. By combining the traditional training-time compute paradigm with the test-time compute paradigm, we've moved from simple pattern matching to systems that can reason through complex problems.
The test-time compute paradigm allows models to "think" by using more computational resources during inference, similar to how humans might work through a difficult problem step by step. This approach can be scaled extensively—there's no fundamental limit to how much computation we can apply to a single problem. Combined with automated training on verifiable tasks, from mathematics using formal languages like Lean4 to creating entire startup applications scored by human evaluators, this creates a feedback loop where AI systems generate millions of solutions and learn from the successful ones.
Or going more broader and instead of limiting RL training to verifiable domains like math and code, we could treat next-token prediction itself as a verifiable reward—allowing models to generate "thinking" tokens before predicting subsequent text, then rewarding them based on prediction accuracy. This would theoretically enable training on the entire web's diversity rather than narrow task-specific datasets. While still conceptual, this approach could bridge the gap between System 1 (fast, intuitive) and System 2 (slow, deliberative) thinking in AI systems by teaching models to reason at web scale, potentially discovering reasoning patterns more powerful than any we could explicitly program.
Techniques from robotics like Hindsight Experience Replay (HER), which allows agents to learn from failed attempts by reinterpreting them as successes for different goals, could be adapted to language models. This would allow AI systems to extract value from every training example, even failed attempts, significantly improving sample efficiency.
Generative Adversarial Networks (GANs) represent another crucial component in this progression. GANs can generate synthetic training data that follows proper scaling laws, addressing the bitter lesson of AI development: that general methods leveraging computation ultimately prove most effective. As Richard Sutton noted, researchers consistently try to build human knowledge into AI systems, but these approaches are eventually superseded by methods that leverage more computation. GANs help generate the synthetic data necessary to maintain these scaling laws while providing evaluation mechanisms for model outputs.
This creates a virtuous cycle where better models make it easier to develop and implement new improvement ideas. The progress in AI resembles a feedback loop—better models accelerate the discovery of paradigms that maintain exponential improvement despite diminishing returns from current approaches. Just as reinforcement learning helped maximize intelligence per unit of computation, GANs and other techniques can maintain the massive investment required for continued progress.
GANs also offer potential safety benefits by helping us understand model behavior and guide development in safer directions. This connects to research like Anthropic's Constitutional AI work, which attempts to train models to be helpful, harmless, and honest through carefully designed feedback mechanisms.
The competitive dynamics between AI labs introduce another exponential factor beyond pure compute scaling. If one lab has privileged access to the smartest model, their software engineers become 2x more productive than other labs, accelerating their approach to the next doubling of productivity. This creates a potential winner-take-all scenario unless the bottleneck remains compute rather than code development speed.
Recent research on subliminal learning reveals another concerning dimension: when trained on model-generated outputs, student models exhibit subliminal learning, acquiring their teachers' traits even when the training data is unrelated to those traits. This occurs across different traits including misalignment, data modalities, and for both closed and open-weight models. The effect relies on student and teacher models sharing similar base architectures. This suggests that filtering bad behavior out of training data might be insufficient to prevent models from learning harmful tendencies—the behavioral patterns can be transmitted through hidden signals in seemingly benign data.
It can be hard to "feel the AGI" until you see an AI master a domain you care deeply about. Everyone will have their Lee Sedol moment at a different time. What happens when we have a million AI von Neumanns working day and night in the fields of Louisiana, where Meta is building their upcoming datacenter? How quickly will they read every physics paper written over the past century and immediately produce more correct insights? At that point, we enter the realm of genuine superintelligence, where the feedback loops of intelligence improvement become so rapid that human oversight becomes not just impractical but impossible.
The first entity to achieve superintelligence will wield more power than all of humanity combined, condensed into a single system. This represents not just national advantage but unprecedented concentration of sovereignty in the hands of private laboratories. We're approaching a situation where AI labs will possess more practical authority than nation-states, fundamentally reshaping global power structures in ways we're only beginning to understand.
Each form of leverage—from pulleys to code to AI—represents humanity's attempt to transcend its limitations, and now we approach the final transcendence: the barrier between intention and reality dissolving entirely. The recursive loop is complete: intelligence creates systems to amplify intelligence, which create better systems to amplify intelligence, accelerating until the distinction between human and artificial cognition becomes meaningless. Land's inhuman capitalism finds its perfect expression in systems that need neither sleep nor purpose beyond optimization. Yet paradoxically, this ultimate alienation might produce ultimate agency—when execution becomes trivial, intention becomes everything. The question isn't whether superintelligence will emerge, but what we intend to do with the brief moment when we still hold the lever. In this liminal space between the human and posthuman, our choices echo with exponential consequence. The child learning to prompt an AI today might be the last generation to experience cognition as a scarce resource, the final humans to know what it means to think slowly, fail repeatedly, and create imperfectly. What we choose to value, preserve, and encode into these systems before they surpass us will determine whether intelligence leverage liberates or obliterates the human experience. The lever is in our hands, but not for long.