Average Is All You Need

A short essay-plus-product-pitch from the rawquery team. The essay argues that LLMs have made average output cheap across creative fields, software, and anything built on "descriptive textual semantics" — IP work, translation, marketing copy, Marvel movies. Before, average cost real time and effort, so most people settled for sub-par. Now average is the floor, and people who care can go further from there. The author's framing: "I can't draw. But now, I still can't draw, but better."

See llm-as-average-democratizer for the generalised concept.

Data as the canonical case

The pitch starts from a claim about analytics: users have high intuition about their own data but low skill at extracting from it. Writing SQL, syncing sources, making charts — those are average-skill tasks that LLMs handle averagely well, which is exactly enough. The slogan: "It deals with the average. You deal with the thinking."

This is a sharper version of the same argument behind ai-assisted-workflow and vamp-ai-frontend: pair LLM with explicit, machine-friendly tools and get past the plumbing faster.

Attribution modelling is wankery

The essay's strongest claim is about a specific ritual. Attribution modelling — deciding whether to weight last-click, first-click, linear, or time-decay — is what the author calls "wankery," a ceremony performed to justify marketing decisions to managers. Most teams give up, pick a default, and dashboard it.

The alternative: describe the join you want in English, get a table, make a decision.

The walkthrough

The post demonstrates with a Stripe + HubSpot scenario:

1. Connect sources. Tell Claude Code to connect Stripe and HubSpot to rawquery. Two minutes later, ten tables land across two schemas (Stripe customers/charges/invoices/subscriptions/products/prices, HubSpot contacts/deals/email_events/campaigns).
2. Ask the question. "Did the 'spring-sale-2026' campaign move the needle?" The agent writes SQL, returns a table: email cohort has 46% higher average basket (89.41 vs 61.27) on fewer customers.
3. Follow up. "Break it down by week. I want to see if the effect wore off." The agent returns weekly data: +50% lift week 1, decaying to +15% by week 4.
4. Publish. "Save the chart and give me a link for my manager." Agent publishes to a public URL, no login required.

The SQL is average. The joins are average. The chart is average. Five minutes, one English prompt, no data engineer, no attribution-model meeting.

What's actually new

Not the analysis — anyone with a week of SQL could produce that breakdown. What's new is that the user didn't need that week of SQL, and didn't need someone else who had it. The essay's closing line: "Honestly, average is clearly magic; prove me wrong."

Caveats the piece doesn't address

This is a marketing post, so the framing is optimistic. A few tensions it glosses over:

• The "average is fine" argument breaks down where correctness matters. LLM-average SQL against production data can silently return wrong joins; the author assumes the human "confirmed it made sense," which is exactly the skill the target user allegedly lacks. Compare cult-of-vibe-coding on why audit discipline matters.
• Attribution ceremony exists for a reason sometimes. Regulated reporting, media spend, and multi-touch incentive plans need defensible methodologies. "Wankery" dismisses that, fairly or not.
• Averaging outputs at scale creates peril-of-laziness-lost conditions. When average is free, more average gets produced, and the scarce resource shifts to whatever filters the flood.

Toolbox

See rawquery for the product itself — a data platform with a CLI designed for LLM-agent operation, with source connectors, SQL execution, chart generation, and public sharing URLs.