In the breathless race to build ever-larger language models, a growing chorus of researchers and entrepreneurs is asking whether the entire field of artificial intelligence has taken a wrong turn — or at least missed several promising exits. A provocative new conversation reported by TechCrunch frames the debate with an arresting metaphor: the history of aviation, and the long, failed pursuit of flight through flapping wings.
The analogy is simple but powerful. For centuries, humans tried to fly by imitating birds — strapping on wings and flapping furiously. It wasn’t until the Wright brothers abandoned biomimicry in favor of fixed-wing aerodynamics that powered flight became possible. The implication for AI is pointed: today’s dominant paradigm, built on scaling transformer architectures and training on oceans of text data, may be the modern equivalent of flapping. It works, after a fashion. But it may not be the path to truly intelligent machines.
The Metaphor That Has Silicon Valley Talking
The “flapping airplanes” metaphor, as explored in the TechCrunch piece, is not merely a rhetorical flourish. It encapsulates a genuine strategic and scientific disagreement at the heart of the AI industry. The prevailing approach — championed by OpenAI, Google DeepMind, Anthropic, and others — is to continue scaling up large language models (LLMs), pouring more data and compute into architectures that have delivered astonishing results over the past several years. GPT-4, Gemini, and Claude have demonstrated capabilities that seemed like science fiction a decade ago, from passing bar exams to writing functional code to engaging in nuanced conversation.
But critics argue that these systems, for all their fluency, remain fundamentally limited. They do not truly understand the world. They hallucinate facts. They struggle with reasoning tasks that a child could handle. And the cost of training and running them — measured in billions of dollars and staggering energy consumption — is becoming increasingly difficult to justify if the returns are plateauing. As the TechCrunch article details, a new generation of AI researchers wants to “try really radically different things,” exploring architectures and approaches that break decisively from the transformer-and-scale orthodoxy.
Beyond Scaling: The Search for New Foundations
The debate over whether scaling alone can deliver artificial general intelligence (AGI) has intensified markedly in recent months. Prominent voices including Yann LeCun, Meta’s chief AI scientist, have argued publicly that autoregressive language models — the backbone of ChatGPT and its competitors — are a dead end for achieving genuine understanding. LeCun has advocated for what he calls “world models,” systems that build internal representations of physical reality rather than merely predicting the next word in a sequence.
This is the intellectual context in which the “flapping airplanes” conversation takes place. The researchers profiled by TechCrunch are not dismissing the achievements of LLMs. Rather, they are arguing that the field needs to diversify its bets. The analogy to aviation history is instructive: the people who tried to build flapping-wing aircraft were not stupid. They were observing nature and drawing reasonable conclusions. But the breakthrough came from a completely different direction — from studying aerodynamics, lift, and thrust as abstract principles rather than trying to copy the biological solution directly.
What “Radically Different” Might Actually Look Like
So what are the alternatives? The conversation around post-transformer AI is rich and varied, encompassing several distinct research programs. One promising direction involves neuromorphic computing, which attempts to replicate the structure and function of biological neural networks at the hardware level, using spiking neurons and event-driven processing rather than the matrix multiplications that dominate current GPU-based training. Companies like Intel (with its Loihi chip) and a host of startups are exploring this space, arguing that it could deliver orders-of-magnitude improvements in energy efficiency.
Another avenue involves hybrid neurosymbolic systems, which combine the pattern-recognition strengths of neural networks with the logical reasoning capabilities of classical symbolic AI. This approach has gained traction in fields like drug discovery and scientific modeling, where the ability to reason about causal relationships — rather than merely identifying statistical correlations — is essential. Researchers at MIT, Stanford, and several European institutions have published work suggesting that neurosymbolic architectures can outperform pure neural networks on tasks requiring compositional generalization, the ability to combine known concepts in novel ways.
The Economics of Exploration vs. Exploitation
The tension between exploring radically new approaches and exploiting the proven transformer paradigm is not merely scientific — it is deeply economic. The major AI labs have invested tens of billions of dollars in infrastructure optimized for training and serving large language models. NVIDIA’s market capitalization, which has soared past $3 trillion, is built substantially on the demand for GPUs used in transformer training. The entire ecosystem — from cloud providers to chip designers to the venture capitalists funding AI startups — has enormous sunk costs in the current approach.
This creates a powerful institutional bias toward incrementalism. When you have spent $100 billion building data centers optimized for a particular kind of computation, the incentive to declare that computation obsolete is approximately zero. The researchers advocating for radical alternatives are, in many cases, working with budgets that are a rounding error compared to what OpenAI or Google can deploy. As the TechCrunch article makes clear, the desire to “try really radically different things” often runs headlong into the reality that the industry’s capital allocation is overwhelmingly concentrated in a single paradigm.
Lessons From the History of Technology Transitions
History suggests that the incumbents in a technological paradigm are rarely the ones who pioneer the next one. The transition from mainframes to personal computers was led not by IBM but by Apple and Microsoft. The shift from on-premises software to cloud computing was driven by Amazon Web Services, not by Oracle or SAP. The pattern is consistent: the companies that have optimized for the current paradigm are structurally disadvantaged when the paradigm shifts, because their organizations, incentives, and infrastructure are all aligned with the old way of doing things.
If the “flapping airplanes” thesis is correct — if the transformer-scaling approach is a local optimum rather than the global one — then the most important AI research being done today may be happening in small labs and university departments, not in the gleaming data centers of Big Tech. This is a sobering thought for investors who have poured hundreds of billions into companies whose valuations depend on the continued dominance of the current approach. It is also an exciting one for the researchers who believe they are working on the fixed wing of artificial intelligence, even if the rest of the world is still focused on making the flapping work better.
The Talent Pipeline and the Problem of Orthodoxy
One underappreciated dimension of this debate concerns the talent pipeline. The overwhelming majority of AI researchers trained in the past five years have been educated in the transformer paradigm. They know how to design attention mechanisms, optimize training runs, and fine-tune large models. Relatively few have deep expertise in alternative approaches like spiking neural networks, probabilistic programming, or formal verification. This creates a self-reinforcing cycle: the dominant paradigm attracts the most talent, which produces the most results, which attracts more funding, which draws in more talent.
Breaking this cycle requires deliberate effort — and, crucially, funding. Some governments have recognized this. The European Union’s AI research programs have explicitly allocated resources to non-mainstream approaches, and Japan’s RIKEN institute has invested heavily in neuromorphic and quantum computing for AI applications. In the United States, DARPA has historically played the role of funding high-risk, high-reward research that the private sector won’t touch, and several of its current programs are aimed at post-deep-learning architectures.
Why the Next Five Years Will Be Decisive
The AI industry stands at an inflection point. The scaling laws that have driven progress for the past several years are showing signs of diminishing returns. Training costs are rising faster than capabilities are improving. And the limitations of current systems — their inability to reason reliably, their tendency to fabricate information, their opacity — are becoming harder to paper over with engineering workarounds. If these problems prove to be fundamental rather than incidental, the case for exploring radically different approaches becomes not just intellectually compelling but economically urgent.
The “flapping airplanes” metaphor, for all its simplicity, captures something essential about this moment. The AI industry has achieved remarkable things by scaling a single architectural paradigm. But the history of technology teaches us that the most important breakthroughs often come from unexpected directions — from people who are willing to abandon the conventional wisdom and try something genuinely new. Whether the current generation of AI dissidents will produce the equivalent of the Wright Flyer remains to be seen. But the fact that serious researchers are asking the question — and that outlets like TechCrunch are amplifying the conversation — suggests that the era of uncritical enthusiasm for scaling may be drawing to a close. What comes next could be far more interesting than anything we have seen so far.