liveBook · Manning

Diminishing returns

As explained in Vibe Engineering

Diminishing returns is a pattern in which each additional increase in an input produces a smaller incremental improvement in outcome.

Diminishing returns occur when each additional increase in an input produces a smaller incremental improvement in an outcome. More input may still help, but the next increment is less valuable than the previous one—and it can sometimes make a result worse. section 6.4.1

For model development, repeated passes over the same corpus still help, but their gains diminish. More compute and scraped data are not a reliable substitute for improvements in data curation, post-training, reasoning, tool use, architecture, or method. The engineering question becomes what to build around the model available now, rather than assuming that scale alone will fix its limitations. section 1.2

More query history

In a query benchmark, moderate-query accuracy rose from about 0.39 with no prior queries to 0.56 with 10, then only to 0.61 with 20. Challenging queries showed a slight p90 regression at 20 examples. Ten prior queries was a sweet spot in this benchmark, not a universal setting: compare accuracy across difficulty levels and percentiles, and include model-running cost in the decision. section 6.4.1

Figure 6.1 Comparison of the p50 vs the number of last queries submitted.

Compare p50 similarity across query-history sizes to see where additional context yields smaller gains. section 6.4.1

When to stop optimizing costs

An optimization sequence cut monthly cost from $8,754 to $5,100 with a 30-minute meeting, then to $2,500 after four steps. Later work reduced the bill further, but the final step saved $493 per month and took a week. Compare the effort cost with the recurring savings over the relevant period before taking another step; a smaller bill alone does not prove the work is worthwhile. section 9.2.8

Table 9.4 Cumulative optimization impact
Problem Solved
Monthly Cost
Savings
Effort
Baseline (Sonnet everything)
$8,754
 
-
-
Wrong tool for the task
$5,100
−$3,654
30-min meeting
Context waste
$4,200
−$900
2-hour audit
One model for everything
$3,300
 
−$900
1 afternoon
Output verbosity
$2,500
 
−$800
1 hour
Everything synchronous
$1,900
−$600
1 day
One tool for all workflows
$1,407
 
−$493
1 week
Total reduction
$8,754 → $1,407
84%
~2 weeks

Inspect the cumulative cost, savings, and effort for each optimization step. section 9.2.8

How Vibe Engineering covers it

How deep it goes
Explains declining marginal gains through benchmark comparisons and an optimization ladder, without deriving a mathematical model.
Examples
Compares query-history benchmarks and a sequence of AI-assisted coding cost reductions.
What you can do after
Evaluate whether additional context or optimization effort is justified by its incremental benefit and cost.
How it connects
Connects diminishing returns to model scaling, context size, accuracy, and cost optimization.

Also explained in

Compare how 4 books explain diminishing returns →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Vibe Engineering. The section links point to the parts of the book it draws on.