The OpenAI Deep Research graph showing tool calls vs pass rate reveals something fascinating about how these models handle increasing amounts of information. The relationship follows a logistic curve that plateaus around 16% pass rate, even as we allow more tool calls.
This plateau behavior reflects something deeper about our current approach to AI. We've built transformer architectures partly inspired by simplified observations of human cognition - particularly how our brains use attention mechanisms to filter and process information. And like human attention, these models have inherent constraints: each attention layer normalizes scores to sum to 1, creating a fixed "attention budget" that must be distributed across all inputs.
A recent paper (https://arxiv.org/abs/2501.19399) explores this limitation, showing how standard attention becomes increasingly diffuse with longer contexts. Their proposed "Scalable-Softmax" helps maintain focused attention at longer ranges, but still shows diminishing returns - pushing the ceiling higher rather than eliminating it.
But here's the deeper question: As we push toward AGI and potentially superintelligent systems, should we remain bound by architectures modeled on our current understanding of human cognition? The human brain's limited attention mechanism evolved under specific constraints and for specific purposes. While it's remarkably effective for human-level intelligence, it might be fundamentally limiting for artificial systems that could theoretically process information in radically different ways.
Looking at the Deep Research results through this lens, the plateau might not just be a technical limitation to overcome, but a sign that we need to fundamentally rethink how artificial systems could process and integrate information. Instead of trying to stretch the capabilities of attention-based architectures, perhaps we need to explore entirely different paradigms of information processing that aren't constrained by biological analogues.
This isn't to dismiss the remarkable achievements of transformer architectures, but rather to suggest that the path to AGI might require breaking free from some of our biologically-inspired assumptions. What would an architecture that processes information in ways fundamentally different from human cognition look like? How might it integrate and reason about information without the constraints of normalized attention?
Would love to hear thoughts from others working on these problems, particularly around novel approaches that move beyond our current biological inspirations.