Scaling the Next Paradigm of Heterogeneous Intelligence — Adrian Bertagnoli, Callosum

Adrian Bertagnoli, founding engineer at Callosum, presents a compelling case that the era of homogeneous AI scaling — training ever-larger models on fleets of identical GPU clusters — is giving way to a new paradigm he calls heterogeneous intelligence. The core thesis is that real-world problems are complex, multi-step, and open-ended, and that routing subtasks to architecturally and computationally distinct models and chips is both mathematically provable and empirically demonstrated to outperform monolithic frontier-model approaches on cost, speed, and accuracy.

Bertagnoli outlines a three-stage evolution of heterogeneity. The current state of 'mild heterogeneity' already includes mixture-of-experts (MoE) architectures replacing dense models, multi-agent pipelines replacing single LLM calls, and prefill-decode disaggregation replacing single-chip inference. The next stage introduces cross-chip routing — placing different LLMs on different GPU or accelerator types — alongside state-space models and diffusion models acting alongside transformer-based systems. The final stage envisions full hardware-software co-evolution and vertical integration of intelligence and silicon.

To formalize the benefit of heterogeneity, Bertagnoli references a mathematical production function argument: a homogeneous system can only optimize for a single peak of skill demand, whereas a heterogeneous multi-agent system can match complex, multi-peaked demand functions. He notes that this principle holds empirically across neuroscience, economics, and ecology, and under any reasonable constraints heterogeneous systems outperform homogeneous ones.

Two concrete benchmark results anchor the talk. The first concerns recursive long-context reasoning. Callosum extended the MIT recursive language model architecture — which uses a file-based environment and Python REPL interactions instead of stuffing full context into a prompt — by routing generated sub-contexts to different chips and models rather than a single model on a single chip. Evaluated on the Ulong benchmark, their system running on Cerebras is 7x cheaper and 5x faster than GPT-5.2 (approximately $0.54 vs $3.75 per task, with ~400 seconds vs ~2,000 seconds runtime). Running on SambaNova achieves 12x cost reduction and 3x speed improvement at the cost of some additional latency.

The second result is a visual web navigation benchmark on Video Web Arena. By decomposing the web navigation task into heterogeneous subtasks — zooming and visual parsing assigned to Qwen3 VL8B-Instruct, higher-level reasoning handled by Kimi K2.5 — Callosum's mixture outperformed GPT-5.2 by 18% and Gemini 2.5 by 25% on the benchmark while being 3.7x cheaper and 3x faster than GPT-5.2 alone. Critically, offloading only the zoom and visual parsing subtasks to the smaller Qwen model yields 11x speed improvement and 43x cost reduction on those specific steps, illustrating that routing granularity is key to the efficiency gains.

On task routing automation, Bertagnoli notes that initial implementations used bespoke human decisions to map subtasks to models, but Callosum has since built an automation layer that detects task complexity and automatically predicts the optimal model and hardware pairing. Callosum is also partnering with the UK's Aria Institute (formerly ARIA) and has received a £3 million grant to operate the first heterogeneous collocated cluster in the UK. Bertagnoli closes by stating that heterogeneous intelligence — where models, workflows, and silicon co-evolve — represents the next scaling paradigm, and that current infrastructure is the worst it will ever be.