Greedy-but-Smart Compute Allocation
Throw hundreds of dollars of inference at high-value problems — the value-to-cost ratio is absurd in your favor.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 3
- Confidence
- 85%
A mental model for spending on AI inference. People instinctively treat model-call costs as expense to minimize, but for high-value knowledge work the value-to-cost ratio is wildly favorable: hundreds of dollars of inference can deliver what a consulting firm would charge millions for. The rule is to apply AI greedily but smartly — aggressively throw compute cycles at your highest-value problems, using patterns like LLM map-reduce to process corpora larger than a single context window.
Origin
Howie Liu's framing, illustrated by his own habit of being (by his account) the single most expensive inference-cost user of Airtable AI. He describes being 'intentionally wasteful' on high-value analysis.
Core principles
- 01The value versus actual cost of AI, applied greedily but smartly, is a crazy ratio — more people should throw compute at high-value problems.
- 02'Intentionally wasteful' spending (hundreds of dollars per exercise) is trivial next to the strategic value of better insight.
- 03Cost objections ('that's so expensive') usually ignore the human hours or consulting fees the compute replaces.
- 04Use map-reduce to exceed context limits: chunk the corpus, run an LLM call per chunk, then an aggregation call over the results.
How to run it
- 1
Identify a high-value, high-effort knowledge problem
Target work where better insight has large strategic value and where the human alternative is slow or expensive — e.g. reading every sales-call transcript from the past year to extract product, marketing, positioning, and segmentation insights.
Pro tip Frame the comparison against the human/consulting cost, not against zero — that's what makes the spend obviously worth it.
Watch out 'Greedy' is qualified by 'smart' — aim compute at genuinely high-value problems, not at everything indiscriminately.
- 2
Apply LLM map-reduce for large corpora
When the content exceeds a single context window, map across it: break it into chunks, run an expensive-model call on each chunk to extract what you need, then run an aggregation call over those chunk outputs.
Pro tip This is expensive by design — a strong model runs over a lot of data and then again over the aggregates — but that's the point when the output is decision-grade.
- 3
Reframe the cost against the alternative
Judge the spend by the value and the substitute cost. Hundreds of dollars of inference that produces chief-of-staff-grade insight is trivial next to the consulting fees or human weeks it replaces; a $1+ deep-research call that saves a human hours is a bargain.
Pro tip A researcher charging $400–500 for guest background is undercut by a stack of ~$1 deep-research calls — the human was probably using the same tools anyway.
Watch out Don't let sticker shock ('$50/month of research calls') override the far larger human-hour savings it buys.
In the wild
Liu runs many LLM calls (and LLM map-reduce) over a year of sales-call transcripts to extract product, marketing, positioning, and segmentation insights — sometimes hundreds of dollars of inference per exercise, deliberately 'wasteful.'
→ He describes the result as if a brilliant chief of staff had read every transcript and returned astute insights — work a consulting firm might charge millions for — making the inference cost trivial by comparison.
Common mistakes
Minimizing inference cost on high-value work
Treating model calls as an expense to cut leaves enormous strategic value on the table when the compute would replace far costlier human hours or consulting engagements.
Being greedy without being smart
Throwing expensive compute at low-value or trivial tasks wastes money without the payoff; the ratio only works when aimed at genuinely high-value problems.
Is it for you?
Best for
Founders, executives, and knowledge workers deciding whether to spend real money on AI inference for strategic analysis.
Not ideal for
Low-value, high-volume tasks where cheap models suffice, or cost-constrained settings where per-call spend must be tightly capped.
From the transcript
“the like the value versus the actual cost of AI when applied greedily but smartly like it's just it's it's it's a crazy ratio and…”
“we also have now a capability that's basically like an LLM map reduce”
“It's as if like a really really smart chief of staff has gone through and read every single sales call like transcript that we've had…”
From the episode
How we restructured Airtable’s entire org for AI
Howie Liu (co-founder and CEO)