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

Key Concepts

• Heterogeneous Intelligence: A paradigm where AI systems utilize a diverse mix of models, architectures, and hardware, rather than relying on a single model type on identical chips.
• Homogeneous Intelligence: The current standard of scaling single, dense models across uniform hardware clusters.
• Heterogeneous Recursion: A methodology where context is treated as an environment, allowing agents to programmatically extract sub-contexts and delegate tasks to specialized, smaller models.
• Production Function: A mathematical framework used to model how different skill distributions (agents) meet specific task demands.
• Vertical Integration: The future state where software, hardware, and intelligence co-evolve to optimize performance.

---

1. The Shift from Homogeneous to Heterogeneous Intelligence

Adrian Bertagnoli argues that while "neural scaling laws" (more data/parameters = better models) defined the era of homogeneous intelligence, this approach is becoming inefficient for complex, real-world inference tasks.

• Current State: We are in a phase of "mild heterogeneity," characterized by Mixture of Experts (MoE) architectures, multi-agent workflows, and disaggregated hardware (pre-fill/decode systems).
• The Goal: To move toward a paradigm where intelligence, software, and silicon are vertically integrated, allowing for the co-evolution of systems.

2. Theoretical Foundation: The Principle of Maximum Heterogeneity

Bertagnoli presents a mathematical argument for why heterogeneous systems outperform homogeneous ones:

• The Problem: Real-world tasks are open-ended and multi-step. A single "generalist" model is often inefficient because it lacks the specialized "peak" performance required for specific sub-tasks.
• The Evidence: By mapping a distribution of skills (agents) to a distribution of demands (tasks), heterogeneous systems can cover a broader "skill space" than a single, dense model. This trend is supported by research in neuroscience, economics, and ecology.

3. Methodologies and Practical Applications

A. Heterogeneous Recursion

This approach addresses the "context rot" observed in large language models (LLMs) when dealing with high-information-complexity tasks.

• Process: Instead of loading all context into a prompt, the system treats the context as a file. A coding agent uses Python REPL (regex, keyword search) to extract relevant sub-contexts.
• Delegation: These sub-contexts are passed to smaller, specialized recursive agents.
• Performance: On the Ulong benchmark, this method proved 7x cheaper and 5x faster than GPT-5.2 when using Cerebras hardware, and 12x cheaper and 3x faster using SambaNova.

B. Visual Web Navigation

The team applied a mixture of open and closed models to beat state-of-the-art benchmarks (Video Web Arena).

• Strategy: The system decomposes visual navigation into sub-tasks (e.g., zooming, textual reasoning).
• Optimization: Simple tasks like "zooming" are offloaded to smaller, less intelligent models, while complex reasoning is reserved for frontier models.
• Results: This heterogeneous approach outperformed GPT-5.2 and Gemini 2.5 by 18% and 25% respectively, while being significantly more cost-effective.

4. Automation and Orchestration

When asked how the system decides which model to use for a specific task, Bertagnoli explained:

• Evolution of Logic: Initially, the team used "bespoke decisions" (hardcoded rules).
• Current State: They have implemented an automation layer that analyzes task complexity and dynamically predicts the optimal model and hardware combination for the specific sub-task.

5. The Future of Compute

Bertagnoli outlines three eras of compute:

1. CPU Era: Focused on raw speed.
2. Nvidia/Parallel Era: Focused on massive parallelism.
3. Heterogeneous Era: Focused on mapping multi-agentic workloads onto optimal, diverse silicon.

"The era of homogeneous scaled delivered extraordinary progress... What comes next is heterogeneous intelligence where models, workflows, and silicon co-evolve and every new source of diversity makes the whole system smarter, faster, and cheaper." — Adrian Bertagnoli

Synthesis

The transition to heterogeneous intelligence represents a fundamental shift in AI efficiency. By moving away from the "one-size-fits-all" model approach, developers can achieve superior performance at a fraction of the cost. The key takeaway is that intelligence should be modular and task-specific, with an orchestration layer that intelligently routes sub-tasks to the most efficient hardware and model architecture available.