In an industry intoxicated by the promise of infinite growth, Arvind Narayanan delivers a sobering reality check: the belief that simply making AI models bigger will inevitably lead to artificial general intelligence is a dangerous myth. While the tech sector races to pour more compute into larger datasets, Narayanan argues that we are hitting hard physical and economic walls that no amount of hype can scale past. This piece is essential listening for anyone trying to separate genuine technological progress from the speculative fever dreams of venture capital.
The Illusion of Endless Scaling
The prevailing narrative suggests that AI capabilities will improve predictably as models grow, eventually crossing a threshold into human-level intelligence. Narayanan dismantles this by pointing out a fundamental misunderstanding of what "better" actually means in machine learning. He writes, "Scaling laws only quantify the decrease in perplexity, that is, improvement in how well models can predict the next word in a sequence." This is a crucial distinction that often gets lost in the noise. Perplexity measures statistical prediction, not genuine understanding or the ability to solve novel problems.
The author argues that while larger models have historically shown new "emergent abilities," there is no law guaranteeing this trend will continue. Narayanan notes, "Emergence is not governed by any law-like behavior." This challenges the core assumption of many AI boosters who treat current trends as inevitable physics rather than contingent engineering outcomes. The evidence suggests these models struggle with extrapolation, failing to solve tasks they haven't seen in their training data. As Narayanan puts it, "If LLMs can't do much beyond what's seen in training, at some point, having more data no longer helps because all the tasks that are ever going to be represented in it are already represented."
Critics might argue that we simply haven't found the right metric to measure these emergent abilities yet, or that the models are just under-trained. However, Narayanan's skepticism is grounded in the observation that traditional machine learning models always plateau, and there is no empirical reason to believe large language models are exempt from this rule.
"Scaling alone will lead to AGI is a view that rests on a series of myths and misconceptions."
The Data Wall and the Cost of Growth
Even if the theoretical desire for larger models existed, the practical reality of data availability is rapidly closing the door. The industry is already consuming the vast majority of high-quality human-generated text available on the internet. Narayanan challenges the optimistic view that scraping the entire internet, including platforms like YouTube, will provide a new endless well of data. He writes, "People sometimes assume that new data sources, such as transcribing all of YouTube, will increase the available data volume by another order of magnitude or two." But after filtering for noise, music, and unusable content, the actual gain is negligible compared to the trillions of tokens already used.
Furthermore, the economic landscape is shifting. Copyright holders are demanding compensation, and regulatory pushback is growing. Narayanan warns that "getting it will cost more and more. And now that copyright holders have wised up and want to be compensated, the cost might be especially steep." This isn't just a technical bottleneck; it's a business one. The article draws a parallel to the history of CPU clock speeds, which stopped increasing not because of a lack of ambition, but because the industry decided further increases were too costly and pointless. Similarly, the flight speed of airplanes plateaued when fuel efficiency became a higher priority than raw velocity.
The author suggests that the industry may have already reached the limit of what scaling can achieve. "With LLMs, we may have a couple of orders of magnitude of scaling left, or we may already be done," Narayanan writes. This uncertainty is often ignored in favor of linear extrapolation, but history shows that tech trends can stall suddenly rather than gradually.
The Shift to Efficiency and Synthetic Data
Perhaps the most surprising shift Narayanan identifies is the industry's move away from massive models toward smaller, more efficient ones. The market is no longer driven by the need for raw capability, which is already sufficient for many applications, but by the need to reduce costs. Narayanan observes, "In the past year, much of the development effort has gone into producing smaller models at a given capability level." He points to recent releases from major developers where newer, smaller models are both cheaper and more capable than their predecessors. This indicates that the race is no longer about who can build the biggest model, but who can build the most efficient one.
The article also debunks the idea that synthetic data—AI generating data for AI—can solve the data shortage. While useful for specific tasks like math or code, Narayanan argues it cannot replace the diversity of human data for general learning. "In short, it's unlikely that mindless generation of synthetic training data will have the same effect as having more high-quality human data," he writes. The success of self-play in games like Go is a specific case that doesn't easily translate to the open-ended complexity of human language.
"The trendline itself contains no clue that it is about to plateau."
The AGI Definition Game
Finally, Narayanan addresses the shifting goalposts of Artificial General Intelligence. As predictions of imminent AGI fail to materialize, industry leaders are quietly redefining the term to save face. Instead of admitting their timelines were wrong, they are watering down the definition until it becomes meaningless. Narayanan suggests we view generality as a spectrum rather than a binary switch. He writes, "Instead of viewing generality as a binary, we can view it as a spectrum." This reframing acknowledges that while AI is becoming more capable, the leap to performing any economically valuable job as effectively as a human is not a matter of just scaling up current models.
The author concludes that the AI research community has historically been terrible at predicting the limits of current paradigms. "Historically, standing on each step of the ladder, the AI research community has been terrible at predicting how much farther you can go with the current paradigm," Narayanan notes. This humility is a stark contrast to the confident timelines often presented by CEOs and investors.
Bottom Line
Arvind Narayanan's argument is a necessary corrective to the industry's unchecked optimism, grounding the conversation in the hard constraints of data availability, economic reality, and the limitations of current machine learning architectures. While the piece may underestimate the potential for breakthrough algorithmic efficiency, its core thesis—that scaling alone is not a magic bullet for AGI—is a robust and timely reminder that physical and economic laws still apply to artificial intelligence. Readers should watch for how quickly the industry pivots from scaling narratives to efficiency narratives as the data wall becomes undeniable.