Useful Output Across Three Modes of Work

I care about useful output per hour of my attention. To compare how I work, I distinguish three modes by where my attention goes.

In mode 1, I directly produce the work. In mode 2, the assistant can accelerate production, but I still participate throughout the process. In mode 3, I move my attention to the beginning and end. The machine can make attempts while I do something else.

Mode 3 removes continuous supervision, but final review still counts as work.

Comparing the three modes with maths

For each mode $i$, I define $U_i$ as useful, verified output and $A_i$ as total hours of my attention:

I count equally valuable units of accepted work, not lines of code or raw generations. For uncertain output, I use expected useful output. A batch of $N$ attempts with probability $p$ of a verified, retainable result has expected output $Np$.

Here is an illustrative comparison across ten attempts. I assume manual and assisted work both succeed 100% of the time, while the automated loop succeeds 30%. These are modelling assumptions, not measured performance claims. I also assume the assistant halves my attention per attempt.

The loop produces fewer useful results from those ten attempts. Yet it produces six times as much useful output per attention hour as manual work, and three times as much as interactive assistance.

At equal attempt duration, the 100% approach produces 3.33 times as much useful output per execution hour as the 30% approach. Both observations can be true: one comparison values execution efficiency; the other values my attention. This table does not assume the three modes take equal elapsed time.

Including compute and failure costs

Attention efficiency is only part of the economics. I also define $K_i$ as total compute cost, $w$ as the value of an attention hour, and $D_i$ as expected failure damage in money:

I exclude damage already counted as repair time or compute, so I do not count the same cost twice.

Using the table above, suppose I value my attention at £60 per hour. I assume negligible incremental compute cost for manual work, £10 per batch in each AI mode, and £5 expected damage for the loop:

Mode 1: I type, £60 per useful result.

Ten hours of my attention cost £600, spread across ten useful results.

Mode 2: I type and view assistant output, £31 per useful result.

I spend £10 on compute and five hours of attention (£300), producing ten useful results.

Mode 3: Automated loop, then final review, £15 per useful result.

I spend £10 on compute and half an hour of attention (£30), with £5 of expected failure damage, producing three useful results.

Under these assumptions, the loop has the lowest cost per useful result. If its total review and repair time rises to two hours, its cost becomes £45 per useful result. Interactive assistance then wins at £31.

The condition that makes the loop worthwhile

The 30% must be useful results I can identify and safely keep. If only two results survive verification, the denominator becomes two. Passing weak tests does not establish usefulness.

I include investigation, filtering, final review, and repairs in my attention budget. I also count the cost of building and maintaining automated checks. Escaped failures belong in the damage estimate. Three successes are an expectation, not a guarantee for every batch.

My preference for mode 3 is therefore conditional: low success rates can be economical when failures are cheap, filtering requires little attention, and the attention saved is worth more than the extra compute and damage.

Related: Attention Arbitrage, The RALPH Loop, and Verification Ladder.