Ineffable's Path to Superintelligence Without Human Data

The dominant field took a different direction. Silver was surprised by how systems trained on massive datasets could be fine-tuned on small datasets to change behavior completely—a finding that taught him a crucial lesson about scale and learning speed.

At the right scale, slow processes can become very fast once systems experience enough data. This observation shapes Ineffable's entire approach: instead of following the path of language models trained on human data, the company pursues learning systems that discover knowledge on their own.

The alternative is not a retreat into older methods but a different wager about what scale and experience can unlock—capabilities that remain locked behind the constraint of learning from human-curated datasets.

A superlearner must possess incredible capacity in two dimensions: the depth of knowledge it can attain and the breadth of domains it can master. In depth, it must begin knowing nothing at all and advance to discovering things far beyond human knowledge—a quality humans themselves embody.

Humans collectively started with empty minds and discovered all knowledge today through language, technology, science, and mathematics. They received no blueprint and no trillion-element dataset. They figured things out themselves through interaction with the world.

This self-discovery capability is the essence of the path to superintelligence but remains uncracked in AI systems today. Every system built so far has relied on human guidance—either through data, reward functions, or feedback loops that embed human judgment.

Cracking this self-discovery problem means building agents that, like humans, can begin with nothing and pull knowledge from the environment without human engineers specifying what to learn or how to learn it.

Rather than engineer reward functions that specify what agents should learn—science, mathematics, strategic knowledge—Silver proposes giving agents straightforward objectives within complex environments. Nature provides the model: evolution optimized for a single quantity, survival and fitness, and generated all the complexity visible in life today.

Humans followed a similar trajectory. Starting with simple goals like coordination and communication, they developed trade, discovered arithmetic, recognized the utility of tools, learned basic science, built complex technology, and ultimately created advanced mathematics. None of this complexity was predetermined in a curriculum; it emerged from relatively simple starting incentives.

The lesson is not that complexity requires complexity to seed it. Rather, simple-to-describe quantities pursued in sufficiently rich environments give rise to incredibly hard-to-describe complexity and emergence.

This principle guides Ineffable's approach: instead of hand-crafting what agents should discover, the company constructs complex environments and simple optimization targets, then lets emergent knowledge surface on its own.

Silver believes agents will discover far more sophisticated approaches to reasoning—parallelizable, rich, and complex methods that may resemble neural computation or take forms humans have not yet conceived.

These approaches could be far more expressive and far more capable than reasoning based on human language. Human language embeds historical biases: flat-earth beliefs, the theory of four humors, countless false explanations that seemed logical at their time.

Starting from human language as the foundation for agent reasoning constrains the system with these embedded biases and with the limitations of a substrate optimized for human communication, not for the deepest forms of thought. A system that discovers its own reasoning substrate is free from this constraint.

Humans fine-tune models, select examples, provide feedback, and guide the direction of learning. This approach works—it has generated impressive capabilities—but it is fundamentally constrained.

The bottleneck is not capability but dependency. As long as thousands of humans remain in the learning process, the system is not learning to discover on its own. It is learning to do what humans teach it, faster and at scale.

Stepping back from human guidance requires discovering the process that learns for itself—the algorithm and architecture that can operate without human hands in the loop. Only once that discovery is made can Ineffable's model of true superintelligence be approached.

Silver does not dismiss the current approach as unimpressive but as fundamentally limited: it will not reach the destination of a superlearner, no matter how large the human team or how sophisticated the fine-tuning becomes.

Simulation can be infinitely paralyzed, and agents within it achieve perfect reproducibility—two properties that accelerate learning at scales impossible in physical reality.

Simulation is Turing-complete. All computation, all economics, all language, and all mathematics are expressible within a sufficiently complex simulated environment. Agents do not need the real world to discover science or laws; they can discover their own science within simulation, starting from nothing.

Simulation also offers safety advantages for developing superintelligence. The system can be contained, paused, and studied before any capability leaves the bounded environment. Silver disagrees with researchers like Rich Sutton who dismiss simulation as inadequate; he believes the science to discover the superlearner can and should happen in simulation first.

Real-world transfer requires making simulations more high-fidelity, but once the algorithm is discovered in simulation, the same approach should apply to the real world. The bottleneck is discovering the algorithm, not adapting it to physical reality.

The gate is not arbitrary; it is calibrated to confidence.

The company will only move to the next level of real-world capability once absolutely confident that outcomes will be good. This reverses the typical startup approach of building in public and iterating rapidly; instead, Ineffable builds in the safest environment first, accumulates evidence of safety and capability, and only then crosses from simulation to the physical world.

This commitment shapes every technical decision the company makes and every system it deploys.

GPUs have been tailored over many years toward training and, more recently, toward inference. Simulation remains an afterthought in hardware design. Model-based reinforcement learning, which depends on fast simulation, makes the model itself more expensive to compute than the simulation itself—a backwards economics that slows scaling.

Ineffable addresses this constraint through replay buffers and data reuse techniques, squeezing efficiency from hardware that was not optimized for the task. The team is actively optimizing simulations to make the best use of available GPUs, a form of research that requires both algorithmic innovation and deep familiarity with hardware limits.

The mismatch between hardware and Ineffable's needs is not insurmountable, but it means the company is working against the grain of the market. GPUs are abundant because they serve the massive market for LLM training; simulation has a much smaller audience, and the hardware industry has not caught up.

Silver's reasoning is not financial but strategic: a small number of exceptionally talented people can do an enormous amount of work with modern tools, and adding more people creates coordination costs and resource competition that slow research.

Selection criteria are rigorous: exceptional talent, collaborative nature, mission alignment, and the specific skills the research requires. Four co-founders come from the AlphaStar team; three hires came from successful startups; all new hires are vetted at a high bar.

The high compute-to-person ratio means each researcher can run many experiments simultaneously without waiting for shared resources. This structure mirrors deep research labs—not startups optimized for growth—and reflects Silver's view that superintelligence research is fundamentally a science problem, not a business-scaling problem.