Significant opportunity exists in re-architecting how AI models work. Instead of building ever-larger single models, the focus is shifting to creating networks of smaller, specialized models that collaborate, which can drastically reduce the cost per token produced.

Instead of focusing on making Transformers cheaper, researchers should identify their inherent weaknesses. Jerry Tworek argues the current architectural bottleneck, not just scale or algorithms, is what's holding back progress toward smarter AI systems.

The argument that LLMs are just "stochastic parrots" is outdated. Current frontier models are trained via Reinforcement Learning, where the signal is not "did you predict the right token?" but "did you get the right answer?" This is based on complex, often qualitative criteria, pushing models beyond simple statistical correlation.

The emergence of specialized models like JEV signals a shift away from a "one model fits all" approach. Instead of forcing a single, expensive LLM to perform all tasks, companies will build complex architectures using a "model stack." This involves using fast judgment models for routing and then invoking generative models only when necessary.

Today's AI models are static once trained. The next architectural shift will be to 'continuous learning' models that can adapt and evolve post-deployment. This change will be so fundamental that it will render all existing models, from open-source to frontier, obsolete within the next decade.

LLMs operate autoregressively, making one decision (token) at a time without seeing the full problem space. This can lead to hallucinations or dead ends. EBMs are non-autoregressive, allowing them to see all possible routes simultaneously and select an optimal path, much like having a bird's-eye view of a map to avoid a hole in the road.

Traditional video models process an entire clip at once, causing delays. Descartes' Mirage model is autoregressive, predicting only the next frame based on the input stream and previously generated frames. This LLM-like approach is what enables its real-time, low-latency performance.

Instead of just expanding context windows, the next architectural shift is toward models that learn to manage their own context. Inspired by Recursive Language Models (RLMs), these agents will actively retrieve, transform, and store information in a persistent state, enabling more effective long-horizon reasoning.

A fundamental constraint today is that the model architecture used for training must be the same as the one used for inference. Future breakthroughs could come from lifting this constraint. This would allow for specialized models: one optimized for compute-intensive training and another for memory-intensive serving.