In the corridors of Silicon Valley and the trading floors of Wall Street, a singular narrative has taken hold with almost religious fervor: artificial general intelligence is just around the corner, perhaps arriving within the decade, and it will reshape everything. Billions of dollars in capital expenditure flow into data centers. Valuations of AI companies soar on the promise of superintelligent systems. CEOs of the world’s most powerful technology firms speak of AGI timelines measured in single-digit years.
But a thoughtful counter-argument is emerging from within the technical community itself — not from Luddites or skeptics unfamiliar with the technology, but from practitioners who work intimately with these systems and understand both their remarkable capabilities and their fundamental limitations. One such voice belongs to Daniel Lants, a software engineer and AI researcher, who published a detailed technical essay titled “AGI Is Not Imminent” that methodically dismantles the prevailing consensus on near-term AGI arrival.
A Practitioner’s Dissent From the Silicon Valley Consensus
Lants’s argument is not that AI is unimpressive or that large language models lack utility. Rather, his thesis is more nuanced and, for investors and industry leaders, potentially more consequential: the current paradigm of scaling transformer-based models, while producing genuinely useful tools, is running into fundamental barriers that no amount of additional compute or data is likely to overcome on the path to true general intelligence. The essay represents a growing undercurrent of technical skepticism that stands in stark contrast to the bullish pronouncements from OpenAI CEO Sam Altman, who has suggested AGI could arrive by 2028, and Anthropic CEO Dario Amodei, who has spoken of “powerful AI” emerging within a similar timeframe.
At the heart of Lants’s argument is a careful distinction between what current AI systems actually do and what AGI would require. Large language models, he argues, are extraordinarily sophisticated pattern-matching and interpolation engines. They excel at tasks that fall within the distribution of their training data — summarizing text, generating code that resembles existing code, answering questions whose answers exist in some form in their training corpus. But genuine general intelligence, the kind that could autonomously conduct novel scientific research, navigate truly unprecedented situations, or engage in the kind of open-ended reasoning that even an average human performs daily, requires something categorically different.
The Scaling Wall: When More Compute Stops Buying More Intelligence
Perhaps the most commercially significant element of Lants’s analysis concerns the so-called “scaling laws” that have driven the AI investment thesis for the past several years. The empirical observation that model performance improves predictably with increases in data, parameters, and compute has been the foundational assumption behind hundreds of billions of dollars in planned infrastructure spending. Companies like Microsoft, Google, Amazon, and Meta have committed staggering sums to GPU clusters and data center construction on the premise that scaling will continue to yield proportional — or even accelerating — returns in capability.
Lants challenges this assumption directly. He argues that the gains from scaling are exhibiting diminishing returns on the metrics that matter most for general intelligence. While benchmark performance continues to improve in certain narrow domains, the gap between what these systems can do and what would constitute genuine understanding or reasoning is not closing at the rate the scaling narrative implies. The low-hanging fruit of pattern recognition and statistical correlation has been harvested; what remains — causal reasoning, genuine abstraction, robust generalization to out-of-distribution scenarios — may not yield to the same brute-force approach.
The Benchmark Illusion and the Problem of Evaluation
A particularly incisive section of Lants’s essay addresses the problem of benchmarks — the standardized tests by which the AI industry measures progress. He argues that the community has fallen into a trap of optimizing for metrics that do not actually measure what they purport to measure. When a language model achieves high scores on a medical licensing exam or a bar exam, the natural human inference is that the system “understands” medicine or law in some meaningful sense. But Lants contends that these benchmarks are systematically gamed by the training process itself, and that performance on static, well-defined tests is a poor proxy for the kind of flexible, adaptive intelligence that AGI implies.
This critique resonates with concerns raised by other researchers. A recent wave of discussion on X (formerly Twitter) among machine learning researchers has highlighted cases where frontier models that score impressively on benchmarks fail at surprisingly simple tasks that require genuine reasoning rather than pattern completion. These failures are not mere edge cases, critics argue, but symptoms of a fundamental architectural limitation. The models are, in a meaningful sense, performing an extraordinarily sophisticated form of retrieval and recombination rather than genuine cognition.
What Current Systems Cannot Do — and Why It Matters
Lants enumerates several capabilities that would be prerequisites for AGI and that current systems conspicuously lack. Among them: the ability to form and maintain coherent long-term goals across extended time horizons; the capacity for genuine causal reasoning as opposed to correlation detection; the ability to learn efficiently from small amounts of data in novel domains (the way a human child can generalize from a single example); and robust performance in truly novel situations that bear no resemblance to training data.
These are not minor gaps to be papered over by the next model release. They represent, in Lants’s view, qualitative differences that likely require fundamentally new approaches — new architectures, new training paradigms, or perhaps entirely new theoretical frameworks that do not yet exist. The history of artificial intelligence is littered with examples of paradigms that produced impressive early results and then hit walls that could not be overcome by incremental improvement: expert systems in the 1980s, early neural networks before the deep learning revolution, and symbolic AI approaches that dominated for decades before being supplanted.
The Economic Stakes of Getting the Timeline Wrong
For investors, corporate strategists, and policymakers, the question of AGI timelines is not merely academic. The current AI investment cycle is predicated in large part on the assumption that capabilities will continue to advance rapidly and that the economic returns from AI deployment will justify the enormous upfront capital expenditures. If Lants and like-minded skeptics are correct — if the current paradigm is approaching fundamental limits well short of AGI — the implications for capital allocation are profound.
This does not mean that AI is a bubble about to burst. The current generation of language models and related systems are genuinely useful tools that are already generating real economic value in code generation, content creation, customer service automation, and dozens of other applications. But there is a meaningful difference between “AI is a useful technology that will generate solid returns” and “AGI is imminent and will transform the entire economy within a decade.” The former justifies significant but measured investment; the latter justifies the kind of speculative frenzy that has characterized the past two years.
The Intellectual Honesty of Uncertainty
What makes Lants’s essay particularly valuable is its intellectual humility. He does not claim to know that AGI is impossible or that it will never arrive. His argument is more precise: that the evidence does not support the confident predictions of near-term AGI that dominate public discourse, and that the technical community has a responsibility to be honest about the gap between current capabilities and genuine general intelligence. As he writes on his personal site, the default assumption should not be that because progress has been rapid, it will continue to be rapid in the same direction. Paradigm shifts in AI have historically been unpredictable, and the next breakthrough — if it comes — may look nothing like the current approach.
This perspective is gaining traction among a subset of serious researchers and engineers, even as the dominant narrative remains overwhelmingly bullish. The tension between these two views — the optimists who see a clear path from GPT-4 to AGI and the skeptics who see a chasm that current methods cannot bridge — will likely define the AI industry’s trajectory for years to come. For those with capital at stake, the prudent course may be to listen carefully to both sides and resist the temptation to treat any timeline, whether optimistic or pessimistic, as settled fact.
What Comes Next for the Industry
The debate over AGI timelines is ultimately a debate about the nature of intelligence itself — a question that humanity has grappled with for millennia without resolution. What the current moment demands is not breathless prediction but careful, technically grounded analysis of the kind Lants provides. The AI systems we have today are remarkable achievements. They are also, by any honest assessment, far from the general intelligence that their most enthusiastic proponents promise. Recognizing this gap is not pessimism; it is the foundation of sound strategy in an industry where the distance between hype and reality can be measured in hundreds of billions of dollars.