Defining Intelligence by Efficiency: Current AI Lags Behind Humans by 6 Orders of Magnitude

A Different Definition of Intelligence: Efficiency in Converting Experience into Capability

Debates about "what is intelligence" never truly end, but one perspective is drawing increasing attention from researchers: intelligence can be defined as the efficiency with which experience is converted into capability. In other words, measuring the level of intelligence isn't just about the final capability achieved — it's about how much experience, compute, and energy were required to get there.

Under this definition, a perspective that sparked discussion on social media argues that current AI lags far behind humans on the efficiency dimension — by roughly 6 orders of magnitude (OOM). This isn't an emotional dismissal, but a conclusion drawn from two quantifiable dimensions: learning efficiency and inference cost.

Learning Efficiency: Humans Take Hundreds of Hours vs. AI's Billion Hours

The first massive gap shows up in data efficiency.

Human evolutionary history made absolutely no preparation for "writing Python code" — our ancestors never encountered programming languages, and there's no compiler in our DNA. Yet an ordinary person only needs to invest a few hundred hours of learning to reach a level of competency sufficient for programming work.

By contrast, a large reasoning model (LRM) requires the equivalent of roughly 1 billion hours of training data to reach a comparable level of programming ability — and that's on top of having already absorbed vast amounts of other training data unrelated to programming.

Placing "a few hundred hours" next to "1 billion hours" makes the gap in learning efficiency immediately apparent. Humans can rapidly generalize from very few examples, while current models remain heavily dependent on large-scale data accumulation. Under the definition of "intelligence = efficiency of converting experience," this is AI's most glaring weakness.

Background: "Data efficiency" and "few-shot learning" are the core concepts for measuring this gap. When humans learn a new skill, they rely on transfer of prior knowledge, causal reasoning, and highly compressed conceptual representations — enabling rapid generalization from very few examples. Cognitive science calls this capacity "inductive bias." Current mainstream large language models and reasoning models are fundamentally statistical pattern-matching systems whose generalization ability depends heavily on the density of training data coverage. The "roughly 1 billion hours of training data" figure is not a literal measure of single-task training time, but rather a conversion of the total token count consumed during model pre-training into an equivalent value based on human reading speed — offering an intuitive sense of the order-of-magnitude difference in data scale. This conversion is inherently an approximation, but the order-of-magnitude comparison has considerable consensus among researchers.

Inference Cost: A Human Spends Pennies Per Game, AI Spends Hundreds of Dollars

The second dimension is test-time compute efficiency and energy efficiency — and the gap here is equally staggering.

Take reasoning games like ARC-AGI-3 as an example: the energy consumed by a human player completing one game, converted at retail electricity prices, costs less than $0.10. Yet according to the data cited in the original post, a system called Astra completing the same game costs $300 to $400.

This means that for solving the same class of problems, there is a 3 to 4 order-of-magnitude gap in cost and energy consumption between AI and humans. The human brain performs complex reasoning at roughly 20 watts of power — an energy efficiency that no AI system has come close to matching.

Looked at another way: even if future models match or surpass human accuracy, as long as the cost of a single inference remains in the range of "hundreds of dollars per game," such systems still don't qualify as truly efficient intelligence under this definition.

Background: ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) is a reasoning benchmark designed by AI researcher François Chollet to measure a system's abstract reasoning and generalization ability — not memorization or pattern matching. Each problem presents a small number of input-output examples, requiring the system to induce a pattern and apply it to new samples, deliberately avoiding the possibility of "memorizing answers" through large-scale training data. ARC-AGI-3 is the latest version in this series, with further increased difficulty. Because its design philosophy aligns closely with "efficiency as intelligence" — both emphasizing the ability to generalize from very few examples — it is frequently used as a reference scenario for measuring the reasoning efficiency gap between AI and humans. Notably, the average human score on ARC-AGI is around 85%, while top AI systems have only recently begun approaching this level on some versions, yet the computational cost they consume remains vastly disproportionate compared to humans.

Why This Perspective Deserves Attention

Mainstream AI capability evaluations tend to focus on whether something can be done — can it pass an exam, write code, win a game. "Efficiency as intelligence" offers a complementary and more demanding benchmark: at what cost is it done.

This perspective offers several insights:

• Data efficiency is a core unsolved problem. The human capacity for few-shot learning remains a target that current architectures cannot approach.
• Energy consumption and cost are not engineering details — they are fundamental indicators of intelligence. True general intelligence should be efficient, not reliant on piling on unlimited compute to achieve performance.
• Benchmark scores are not equivalent to intelligence. When a system requires a million times more data and thousands of times more energy than a human to complete a task, claims of "approaching human-level intelligence" need to be heavily qualified.

Background: The human brain consumes roughly 20 watts of power, with annual electricity consumption of less than 200 kWh — an energy efficiency forged by hundreds of millions of years of evolutionary pressure into an extremely optimized neural computing architecture. The corresponding concept is "neuromorphic computing," which attempts to dramatically reduce energy consumption by mimicking the brain's sparse activation and event-driven mechanisms at the hardware level. Representative projects include Intel's Loihi chip and IBM's TrueNorth. Current mainstream AI training and inference relies on GPU/TPU clusters that consume megawatts of power, and the energy cost of a single large model inference call can be thousands of times greater than a human completing an equivalent cognitive task. The energy gap is therefore not merely an economic issue — it is a deep-level indicator of whether a computing paradigm is architecturally approaching "real intelligence."

Beyond the Gap: Questions Left for the Industry

To be clear: "lagging by 6 orders of magnitude" is an estimate based on a specific definition and specific cases (learning to program, ARC-AGI-3 games). It is inherently discussion-oriented in nature, not a rigorous academic measurement. The efficiency gap will vary significantly across different tasks and different models.

But it does point to a directional question: if we genuinely understand intelligence as "efficiency in converting experience into capability," then current AI progress is reflected more in scale than in efficiency. The truly transformative breakthroughs ahead may not lie in making models larger or feeding them more data — but in enabling machines to learn like humans: acquiring more from less experience, at lower energy cost.

And that may be the critical step toward intelligence that more closely resembles our own.