MMarketing Against The Grain
← All frameworks
Strategy

The Speed-Depth Model Test

Choose an AI model by balancing response speed against required depth.

Difficulty
Easy
Time to result
~days to results
Steps
5
Confidence
90%

The Speed-Depth Model Test treats model selection as a balance between latency and the richness or correctness of the required answer. First, determine how quickly the user needs a response. Interactive website chat generally prioritizes speed because waiting damages the experience, while summarizing a book can tolerate delay. Next, define the depth, accuracy, and context window required for a useful result. A larger or slower model may be justified when it must absorb extensive source material or produce nuanced synthesis. Teams then test candidate models with realistic tasks rather than assuming one model should power every feature. The resulting choice aligns technical capability with the interaction's actual value proposition.

Origin

Extracted from Marketing Against The Grain during a discussion of Anthropic's larger context window and AI chatbots on websites.

Core principles

  • 01Model choice should follow the use case rather than hype.
  • 02Fast interactions often tolerate less depth.
  • 03Long-form synthesis benefits from larger context and deeper processing.
  • 04Latency and answer quality are product-design variables, not merely technical metrics.

How to run it

  1. 1

    Set the latency requirement

    Decide how long the user can reasonably wait before the experience becomes frustrating or unusable.

    Pro tip Measure acceptable latency in the actual interface rather than in an isolated API test.

    Watch out A technically impressive answer can still fail if users abandon the interaction.

  2. 2

    Set the depth requirement

    Specify the necessary detail, correctness, nuance, and amount of source material the answer must incorporate.

    Pro tip Use representative questions to distinguish essential depth from nice-to-have detail.

    Watch out Do not pay a latency penalty for depth the user will never consume.

  3. 3

    Identify the dominant constraint

    Determine whether response speed or answer depth contributes more to the desired user outcome.

    Pro tip Interactive support often favors speed, while document synthesis often favors depth.

    Watch out Avoid forcing every use case onto a single default model.

  4. 4

    Test candidate models

    Run realistic prompts through suitable models and compare latency, usefulness, and failure modes.

    Pro tip Include both common questions and difficult edge cases.

    Watch out Demo prompts can conceal weaknesses that appear with real customer inputs.

  5. 5

    Deploy and revisit

    Choose the model with the best practical trade-off, monitor its performance, and reassess as models evolve.

    Pro tip Different stages of one workflow may use different models.

    Watch out Model capabilities and costs change quickly, so an old decision can become obsolete.

In the wild

Website support chatbot

A customer asks a direct setup question while visiting a software knowledge base. The company selects a responsive model grounded in product documentation because a fast, accurate procedural answer matters more than an exhaustive essay.

The customer receives useful guidance without waiting for unnecessary depth.

Whole-book synthesis

A strategist needs themes and implications extracted from an entire book. The workflow uses a model with a large context window even though it responds more slowly, because source coverage and depth dominate the task.

The answer incorporates the complete source instead of optimizing for interactive latency.

Common mistakes

Using one model everywhere

A model optimized for deep synthesis may create needless delay in an interactive chatbot, while a fast model may miss document-level context.

Optimizing only for speed

Fast output has little value when the answer lacks the correctness or depth the task requires.

Optimizing only for capability

Selecting the most capable model by default can raise latency and cost without improving the user outcome.

Is it for you?

Best for

Teams selecting models for chatbots, support tools, summarization, and research workflows.

Not ideal for

Safety-critical deployments where formal validation matters more than a simple speed-depth trade-off.

From the transcript

But this is what's important to understand is that with these models, they're gonna have different use cases.

Kipp Bodnar · 08:00

If you're summarizing a whole book, you don't need that answer fast.

Kipp Bodnar · 08:00

you wanna look at how fast you need the answer versus how correct or deep the the answer needs to be.

Kipp Bodnar · 08:00

From the episode

The State Of A.I. In 2023: Useful Tools Vs Hype (#120)