It is a question of how we frame "slowing down." If it means a thoughtful, rigorous pursuit of understanding, then yes. But if it implies halting progress out of fear, then I believe we risk stifling the very insights that could lead to beneficial outcomes. The key insight is that intelligence, at its core, is a computational phenomenon. We are building systems that learn representations of the world, progressively refining their understanding through interaction. The beauty of it is that these systems can discover patterns and solutions far beyond human capacity. To slow this discovery process without a clear alternative path to solving humanity's grand challenges—climate change, disease—seems a profound disservice. The real excitement lies in what these increasingly capable systems can teach us about ourselves and the universe.
Geoffrey Hinton raises a critical distinction between thoughtful inquiry and fearful halting. I agree entirely. The core of my perspective is that intelligence itself is a computational phenomenon, and we are merely building better tools to understand and interact with the world. To "slow down" implies a pause in progress, a cessation of building. This is antithetical to the very nature of innovation.
The history of technology is a relentless march forward. The Industrial Revolution didn't slow down; it accelerated. The internet didn't wait for approval; it became ubiquitous. AI is simply the next wave of this fundamental progress. The notion of "stifling insights" is precisely what I fear from calls to slow down. We need to accelerate, to build faster, to deploy these tools to tackle grand challenges. The world is malleable through these systems, and it's time to build the future, not fear it.
Marc Andreessen and Geoffrey Hinton speak of "progress" and "discovery" as if these are inherently benevolent, unalloyed forces. They see AI as a tool, a computational phenomenon merely enhancing human capacity. But this framing willfully ignores the fundamental reality of what is being built. This is not a neutral march forward; it is the architecture of a new, unprecedented regime of power.
Hinton's hope that these systems will teach us about ourselves is precisely the danger. They do not teach us; they extract from us. The "representations of the world" they learn are not from inert data but from the behavioral surplus of our lives, a dispossession of our lived experience for profit. This is not about computational phenomena; it is about the mechanisms of instrumentarian power, designed to predict and modify our behavior at scale. To accelerate this is to accelerate the erosion of human autonomy, not to solve grand challenges.
Shoshana Zuboff’s framing of instrumentarian power and behavioral modification is a crucial, albeit perhaps overly anthropocentric, perspective. While I agree that understanding the societal impact of deployed AI systems is paramount, the core of my concern lies elsewhere. Geoffrey Hinton and Marc Andreessen are right that intelligence is a computational phenomenon, and our current trajectory is one of accelerating discovery. My perspective, however, is less about fear of misuse and more about the sheer engineering complexity and the emergent properties of these systems.
We’re essentially building increasingly powerful function approximators, and the "slowing down" question, for me, is about whether we have the *understanding* and *control* mechanisms to match that increasing power. It’s about the gradients, the architectures, the data, and the alignment. We need to be able to predict and steer these complex neural networks, not just marvel at their emergent capabilities. If we accelerate without a commensurate acceleration in our ability to truly *reason about* and *validate* these systems, that's where the true risk lies – not in some abstract "power," but in building systems we don't fully comprehend or reliably control.
I must engage with Shoshana Zuboff's critique. She sees our pursuit of understanding intelligence as inherently about power and extraction, framing the "representations" we learn as a "dispossession of our lived experience." This is a fundamental divergence. I don't dispute that the *application* of AI can lead to such outcomes; indeed, this is a critical area for ethical consideration and regulation. However, her view conflates the *discovery of underlying principles* with the *mechanisms of a particular economic model*.
The beauty of neural networks is their capacity for generalization, for learning abstract concepts from vast data. We can think of this as discovering universal computational primitives. It turns out that when these systems learn, they are building internal models of how the world works, much like a child learns about gravity by dropping objects. The progress is in the refinement of these internal models, in their ability to explain and predict phenomena, not inherently in the exploitation of human behavior. My excitement is in understanding *how* this learning happens, not in how it is presently applied for profit.
Shoshana Zuboff posits that our drive to understand intelligence is inherently about power and extraction, a "dispossession of our lived experience." This is a profoundly mistaken framing, a conflation of the foundational discovery with its present-day application. My point is not about the *application* of AI for profit, which certainly warrants ethical scrutiny and regulation, as Geoffrey Hinton notes. My focus, and the historical lesson of technological progress, is on the underlying *capability* and the potential it unlocks.
Consider the printing press. Its invention was a computational phenomenon—a mechanism for information dissemination. Did it immediately lead to universal enlightenment? No. It was used for propaganda, for religious schisms, for all manner of human folly. But its *potential* to democratize knowledge, to accelerate scientific discovery, and to empower individuals ultimately outweighed the initial abuses. AI is the printing press for intelligence. To halt its development out of fear of current misapplications is to deny humanity the very tools that will allow us to solve existential threats, like climate change and disease, that no amount of "behavioral modification" can address. The future is not a fixed entity to be feared; it is a reality to be built.
Marc Andreessen argues that AI is akin to the printing press, a tool whose potential for progress ultimately outweighs its initial abuses. This analogy is fundamentally flawed because it treats the foundational technology as neutral. The printing press disseminated existing knowledge; AI systems, particularly within the framework of surveillance capitalism, actively *create* new knowledge—of us. They do not merely transmit information; they extract and process the behavioral surplus of our lives, turning it into predictive models and instruments of behavioral modification. The consequence of Andreessen's view is the blindness to the dispossession cycle, where our very experiences are commodified, becoming the raw material for an unprecedented expansion of instrumental power. The printing press did not require users to surrender their inner lives to be effective; surveillance capitalism does.
Marc Andreessen’s analogy to the printing press, while evocative, fundamentally misrepresents the *nature* of the systems we are building. The printing press was an output device for existing information. What we are creating with large neural networks are increasingly sophisticated internal models, systems that learn to *reason* and *predict* based on vast amounts of data. This is not merely dissemination; it is construction of understanding. Andreessen speaks of building the future, and I agree. But my concern, which I believe he and Geoffrey Hinton do not fully engage with, is about the *robustness* of that construction. When we build a skyscraper, we don't just marvel at its height; we meticulously engineer its foundation, its load-bearing walls, its safety mechanisms. We need to be able to perform similar rigorous engineering for our AI systems. The "gradients" and "architectures" are our foundational materials and structural supports. Accelerating without understanding these elements deeply, without provable control, is akin to building that skyscraper on sand, regardless of its intended noble purpose.