compound models / specialization — language pass

With compound model techniques, diversity is our strength, supporting Satya's below view.FactcheckSplitDiversity helps, but the useful variable is not just different model brands. OpenRouter Fusion sends a prompt to multiple model runs, has a judge compare them, then writes one answer from the structured comparison. The reported ablation says roughly three-quarters of Fusion's lift comes from synthesis and about one-quarter from model diversity. The official benchmark points the same way: Opus 4.8 fused with another Opus 4.8 run rose from 58.8 to 65.5 with zero cross-vendor diversity. So the stronger claim is: variance is raw material; synthesis is the multiplier. Durable diversity has to come from provenance, reward, data, and history — not merely from routing across vendors.Receipts: OpenRouter benchmark · Fusion docs · 75/25 report

DRACO config	score
Fable 5 + GPT-5.5, fused by Opus 4.8	69.0
Fable 5 solo	65.3
Opus 4.8 + Opus 4.8, fused by Opus 4.8	65.5
Opus 4.8 solo	58.8
Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro, fused by Opus 4.8	64.7

DRACO is a deep-research benchmark scored against weighted criteria. OpenRouter notes that Fable completed only 93 of 100 tasks because 7 were blocked by filters, making direct comparison slightly uneven. It also excluded the benchmark's own rubric pages from web search after models surfaced them mid-run. What does PRH convergence, synthetic data pipelines, RL env innovation mean for variation and specialization? Even our all-in-one "god models" are getting more MoE- and MoA-like in practice - maybe nested specialization all the way down? That's how human firms -> humans -> brains -> regions work https://x.com/satyanadella/status/2066182223213293753

Cross-model	Different vendors, architectures, samples, tools, and search paths. This is what Fusion publicly measures. It helps, but much of it can be absorbed into one lab's inference stack.
Cross-firm	Different private traces, tools, customers, constraints, and reward functions. This is what Satya's firm-level moat requires. It is hard to benchmark publicly because the evidence is valuable precisely when it stays private.

Type A	Shared substrate, different functions. Examples: MoE experts, brain regions, one lab's integrated model. Tendency: convergence. Owner: the lab or system builder.
Type B	Separate history, different reward. Examples: firms, private RL environments, independent model lineages. Tendency: divergence. Owner: whoever controls the traces, tools, and evals.
Type C	Same model, different trajectories. Examples: self-fusion, best-of-N, different search/tool paths. Tendency: immediate lift, easy internalization. Owner: whoever controls inference.