Descriptions:
Adrian Bertagnoli, founding engineer at Colossyan, delivers a conference talk arguing that the next major paradigm shift in AI will come from heterogeneous intelligence — systems combining different model architectures, sizes, and hardware types rather than scaling a single model on uniform compute. He grounds the argument in a mathematical formalization called the principle of maximum heterogeneity, drawing on analogues from neuroscience, economics, and ecology to demonstrate that heterogeneous systems outperform homogeneous ones under any reasonable constraints. The talk frames current trends — mixture-of-experts replacing dense models, multi-agent systems replacing single LLM calls, prefill-decode disaggregation at the hardware layer — as early signals of a broader architectural transition already underway.
The first concrete primitive Bertagnoli introduces is heterogeneous recursion, an extension of MIT’s recursive language model paper, which showed that context complexity (not just context length) causes performance degradation at around 60–30% context window occupancy. Colossyan’s extension maps recursive sub-tasks to different models and hardware based on computational demand, achieving results comparable to GPT-5 and GPT-5.2 on the Ulong benchmark while running significantly faster and at lower cost — with GPT-5 clocking roughly 2,000 seconds on the same tasks.
The second primitive is multimodal video action language models (VALMs), which integrate visual, language, and action capabilities for agents operating in video-rich environments. Bertagnoli frames the long-term trajectory as a co-evolution of AI software and hardware converging on full vertical integration — specialized silicon matched to specialized model types — as the dominant architecture for solving complex, multi-step real-world problems that decompose into fundamentally different sub-tasks.
📺 Source: AI Engineer · Published May 24, 2026
🏷️ Format: Deep Dive
