Vamsi Ramakrishnan

On the Trinity of Latency, Price, and Quality—and the Point Where Optimization Becomes Theology

Epistemic Status

• High confidence in the classical HCI timing bands (0.1 / 1 / 10 seconds) as durable psychophysics, not tech folklore.
• Medium-high confidence in conversational delay findings clustering around few-second breakpoints. Low confidence in the popular “TTFT < 200ms” rule as applied to streaming LLMs: it is a plausible extrapolation, not an experimentally established threshold.
• Speculative on how streaming, fillers, and the three-knob problem interact with perception—this is the key open question.

“In preparing for battle I have always found that plans are useless, but planning is indispensable.” — Dwight D. Eisenhower

Prologue: The Cartographer’s Dilemma

In 1817, West Point - America’s oldest military service academy - hired a civilian artist named Walter Weir to teach drawing. Not because the generals wanted prettier sketches—but because Napoleon had just demonstrated an unromantic truth: the army with better maps wins the war. Weir’s job was to train officers to see terrain: to translate the continuous chaos of valleys and ridgelines into discrete, usable intelligence. His students learned that the map is always a lie, but some lies are more useful than others.

Every AI product today faces Weir’s dilemma. The terrain is a trinity of constraints—latency, price, quality—and the map must decide what to emphasize and what to omit. The officer who obsesses over every creek bed will never finish his map. The one who ignores the canyon will lose his army.

<aside>

The quality of a map is not how much it includes, but how wisely it excludes.

</aside>

This essay is a map of the GenAI market’s terrain. It is incomplete by design. The thesis is simple and unfashionable: every use case has an asymptote—a point where further improvement in one dimension stops changing user behavior. Below the asymptote, you can buy love with optimization. At the asymptote, you’re paying for applause that no one hears. Above it, you’re overserving—marginal cost exceeds marginal value by definition.

The market confusion comes from a category error: we talk as if “quality” is a scalar, a single rising number, a ladder you climb until you reach the top. It isn’t. Quality is a family of curves, each indexed by the task, the user, the stakes, and the surrounding workflow. Most businesses quietly span several curves at once—then act surprised when the pricing math turns into sociology.