hawkes.aiKey ideas
- The one idea — Routing is a rich-get-richer loop — favored experts improve and get favored more — and training must actively enforce the committee's democracy.
- How it is shown — Collapse in time-lapse: two experts hoarding all traffic while neighbors atrophy; then balancing forces restoring the spread.
- The trap to avoid — Judging MoE by its paper elegance — the delicate part is invisible: the balance machinery that keeps trained capacity alive.
- What it sets up — The vaults are reformed — but attention has its own bills.
Expert models have a rich-get-richer problem: whichever expert gets picked improves, so it gets picked more — until two hoard all the work and the rest atrophy. Unused. Wasted.
The one idea
Routing is a rich-get-richer loop — favored experts improve and get favored more — and training must actively enforce the committee's democracy.
The failure mode last episode was hiding: left alone, expert committees rot. Traffic collapses onto a favored few; the rest fade into expensive, dormant glass. It's called routing collapse, it emerges from training itself — and fighting it is some of modern AI's most delicate engineering. The mechanism has no villain — only feedback. The router slightly favors an expert; it sees more tokens, trains more, genuinely improves — so the router, scoring honestly, favors it more. Around and around, the favorite compounding, until two cells do everything. Rich-get-richer isn't a bug. It's what learning systems do when winners train on the spoils. And collapse ruins the entire bargain. A collapsed committee is a small dense model wearing a trillion-parameter costume — the label says huge, the effective capacity is the few hot cells, and everything last episode celebrated quietly dies. The shelves still bill rent.
How it works — the demo
Collapse in time-lapse: two experts hoarding all traffic while neighbors atrophy; then balancing forces restoring the spread.
Capacity you can't route to is capacity you don't have. So training enforces the democracy. Classic fix: a balance term in the loss — skewed routing taxed, spreading rewarded. Structural fix: capacity limits; hot experts overflow to others. Modern refinement: small auto-tuning bias dials steering traffic toward the neglected without taxing the objective. The committee stays a committee because training polices it, every step. But enforcement walks a tightrope. Push balance too hard and you override honest judgment — tokens shipped to mediocre experts for fairness, quality paying. Too soft, the duopoly returns. Every MoE run tunes this tension live; wrong in either direction has sunk real runs. Balance isn't a checkbox.
The trap to avoid
Judging MoE by its paper elegance — the delicate part is invisible: the balance machinery that keeps trained capacity alive.
Why it matters — and what’s next
The vaults are reformed — but attention has its own bills.
It's a wire. The trap: judging the architecture by its blueprint. On paper, MoE is three elegant sentences. In practice, the decisive engineering is everything around those sentences — balancing forces, overflow plumbing, monitoring — invisible in diagrams, decisive in runs. The old lesson again: ideas are cheap; keeping them alive at scale is the craft. The vaults got their revolution — committees, cops, enforced democracy. But the zoo's other great reform happened in attention, where act two's conversation-memory cache groans under modern context lengths. Three generations of fixes await: sharing, compressing, and finally replacing attention itself. Next group: attention, evolved.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
The failure mode last episode was hiding: left alone, expert committees rot. Traffic collapses onto a favored few; the rest fade into expensive, dormant glass. It's called routing collapse, it emerges from training itself — and fighting it is some of modern AI's most delicate engineering.
The mechanism has no villain — only feedback. The router slightly favors an expert; it sees more tokens, trains more, genuinely improves — so the router, scoring honestly, favors it more. Around and around, the favorite compounding, until two cells do everything. Rich-get-richer isn't a bug. It's what learning systems do when winners train on the spoils.
And collapse ruins the entire bargain. A collapsed committee is a small dense model wearing a trillion-parameter costume — the label says huge, the effective capacity is the few hot cells, and everything last episode celebrated quietly dies. The shelves still bill rent. Capacity you can't route to is capacity you don't have.
So training enforces the democracy. Classic fix: a balance term in the loss — skewed routing taxed, spreading rewarded. Structural fix: capacity limits; hot experts overflow to others. Modern refinement: small auto-tuning bias dials steering traffic toward the neglected without taxing the objective. The committee stays a committee because training polices it, every step.
But enforcement walks a tightrope. Push balance too hard and you override honest judgment — tokens shipped to mediocre experts for fairness, quality paying. Too soft, the duopoly returns. Every MoE run tunes this tension live; wrong in either direction has sunk real runs. Balance isn't a checkbox. It's a wire.
The trap: judging the architecture by its blueprint. On paper, MoE is three elegant sentences. In practice, the decisive engineering is everything around those sentences — balancing forces, overflow plumbing, monitoring — invisible in diagrams, decisive in runs. The old lesson again: ideas are cheap; keeping them alive at scale is the craft.
The vaults got their revolution — committees, cops, enforced democracy. But the zoo's other great reform happened in attention, where act two's conversation-memory cache groans under modern context lengths. Three generations of fixes await: sharing, compressing, and finally replacing attention itself. Next group: attention, evolved.