AI's Next Leap Requires Architectural Overhaul, Not Just Bigger Models
By Imani Sutton · Reporting from Atlanta ·
Experts argue the Transformer model is fundamentally limited and cannot achieve true continual learning in complex, messy environments.
The current technological race is not a contest of scale, but one of fundamental architectural failure. This was perhaps the most sobering takeaway from the latest episode of "Sequoia Capital," where Core Automation’s Jerry Tworek and Rohan Anil argued that the very foundation of modern AI—the Transformer model—is fundamentally incapable of true continual learning. They suggested that simply making models bigger is a dead end, pointing instead to the need for an entirely new architecture capable of meta-learning in the messy reality outside the lab.
On the podcast, Tworek and Anil spent considerable time dissecting the limitations inherent in our current computational methods. Tworek argued that while we have mastered large-scale pre-training and reinforcement learning (RL), the next bottleneck lies not in data or compute, but in the architecture itself. He noted that existing benchmarks fail because they cannot replicate real-world complexity, meaning models need to learn at test time with user data, rather than relying solely on controlled pre-training.
The discussion highlighted a deep tension: AI is trained in an idealized lab environment but must function in the "messy, murky real world." Tworek critiqued current methods like in-context learning for being too limited in scope and duration, while fine-tuning suffers from catastrophic forgetting. He posited that the breakthrough requires finding an algorithm that can meta-learn at the architectural layer to handle much longer learning horizons.
The Illusion of Scaling Tworek provided a historical context that is crucial for understanding current hype cycles: Transformers were economically valuable because their training cost was lower than the revenue they generated, allowing them to scale where earlier architectures like LSTMs could not. However, he quickly pivoted to point out the structural weakness of this scaling model. He argued that current computation focuses heavily on inference time and token generation—a process that is inherently inefficient, merely a "band-aid" solution. The core architectural problem, according to Tworek, is that Transformers have poor "computational depth," meaning their usefulness depends entirely on valuable information being present in the training data; if the world changes, they suffer and require controlled retraining.
This critique directly challenges the prevailing narrative of endless scaling. If a model cannot adapt when its operational environment shifts—if it requires human intervention or massive retraining cycles to keep up with new events—then its supposed "intelligence" is merely sophisticated mimicry, not true understanding.
The Human Element and Agency The most progressive point in the discussion was the definition of labor and intelligence itself. Core Automation’s founder defined automation not as removing humans from the loop entirely, but rather giving every human the maximum level of agency to maximize their time and iteration speed. This reframes AI's purpose from job replacement to augmentation—a necessary shift if these systems are to be deployed responsibly.
However, this talk of "agency" rings hollow when paired with the current economic reality: advanced research labs face a major hurdle due to the lack of long-term focus, as companies compete for short "release cycles," driven by the fact that tokens are not sticky. The relentless pursuit of immediate profitability seems destined to sideline the deeper, more foundational architectural work required to solve the actual problems of continuous learning.
Beyond Perplexity: End-to-End Outcomes The discussion concluded with a technical roadmap focused on solving this fundamental incompatibility between pre-training and RL. Anil argued that focusing solely on metrics like perplexity is insufficient because it only helps build the "prior." To achieve significant improvement, one must look at the end-to-end outcome, requiring a genuine combination of pre-training and RL as the necessary learning algorithm.
The implication for society is profound: if progress requires such fundamental architectural shifts—moving beyond mere scaling to true meta-learning—then the current race among venture-backed labs, which prioritize quick market wins over deep scientific breakthroughs, is fundamentally misdirected. The focus on achieving artificial general intelligence (AGI) must be tempered by a critical understanding of who controls the definition of "progress" and whose labor these systems are ultimately meant to support.
What we need is not another trillion-parameter model trained on more scraped data; we need a radical reorientation of research funding that values algorithmic robustness, true adaptability, and sustainable deployment over the spectacle of sheer scale. The path toward genuinely useful AI must be guided by a commitment to systemic efficiency—in energy use, in resource allocation, and in labor—rather than merely chasing the next impressive benchmark score.