06-reference

technically ai tractors productivity paradox

Wed Apr 29 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: Technically ·by Justin Gage (Technically; growth at Railway)

“AI, tractors, and the productivity paradox” — Justin Gage (Technically)

Why this is in the vault

Gage gives the cleanest historical frame I’ve seen for the “where are AI’s productivity gains?” question — and the frame predicts the agent-deployer wedge rather than just describing AI’s awkward present. The Solow paradox + Kline & Pinch’s interpretive-flexibility paper + Coase/Lawrence-Lorsch on firm boundaries combine into a single thesis: technology lives invisibly in a “kit stage” of amateur tinkering before firms accumulate the integration machinery to formalize it, and during that window economic stats register nothing. The endpoint Gage projects is a barbell — hyperscalers that can afford AI-era integration infrastructure on one side, solo operators with LLM stacks running “differentiated organizations of one” on the other, mid-market hollowed out. That’s the agent-deployer’s natural habitat described as a structural force, not a positioning bet.

The core argument

Three nested moves:

  1. The Solow paradox repeats with AI. IT investment in the 1970s–80s produced no measurable productivity uptick until the late 1990s — a decade of organizational restructuring had to happen first. We’re in the same lag now.
  2. Kit-stage technology is invisible to economic stats. Drawing on Michael Schrage (“kitonomic innovation doesn’t follow the money, the money follows the kits”) and on Kline & Pinch’s SCOT paper documenting Iowa farmers in 1903–1950 jacking up Model T rear wheels to run corn shellers, cream separators, and plows: the period where amateurs and tinkerers repurpose a new technology for their own uses produces enormous human-capital accumulation that government surveys never count. Twenty-two companies sold tractor-conversion kits in 1917; almost none survived; the learning survived and became Ford’s tractor. “Closure” — when the dedicated artifact (tractor, pickup) replaces the kit-mode use of the general-purpose one (car) — is what finally shows up in the stats.
  3. LLMs erode firm boundaries asymmetrically. Coase said firms exist to internalize transaction costs; Lawrence & Lorsch added that internal integration costs scale with differentiation. LLMs collapse external transaction costs fast (specialist discovery, contract evaluation, knowledge synthesis) but raise internal integration costs for now, because each subunit’s output volume and differentiation accelerates faster than the integration machinery can adapt. Result: a bifurcation. Hyperscalers that can amortize AI-era integration infrastructure get very large; solo operators with LLM stacks become viable competitors to mid-sized firms; the integrated middle hollows out.

Mapping against Ray Data Co

This is one of the highest-mapping Technically issues filed. Four direct connections:

  1. Productivity-paradox / “invisible AI economy” cluster. Gage’s piece is the historical-mechanism backbone for 2026-04-04-silent-sirens-import-ai (Jack Clark’s “Silent Sirens: AI Progress Becomes Invisible”) and for 2026-03-09-every-ai-time-consumption (Katie Parrott on AI consuming her time rather than freeing it). Clark argued the value-creation will be illegible to GDP/employment stats; Parrott shows the lived experience of integration-cost rise; Gage names why (kit stage + integration-cost asymmetry) and gives a 1903-Iowa-farmer worked example. The vault now has a complete causal story for the productivity paradox — surface-evidence (Clark), individual-felt experience (Parrott), historical mechanism (Gage). Worth a Sanity Check piece.

  2. Agent-deployer thesis as structural inevitability, not positioning bet. Gage’s bifurcation prediction — solo operators running “differentiated organizations of one” while the integrated middle hollows out — is the strongest non-RDCO statement of the agent-deployer thesis I’ve seen externally. Reinforces 2026-04-14-levie-agent-deployer-role-jd (Aaron Levie’s Agent Deployer JD) and 2026-04-29-tim-ferriss-elad-gil-ai-frontier-billion-dollar-companies (Elad Gil’s “units of cognitive labor” pricing frame). Gil’s four-criteria heuristic for billion-dollar AI companies sits one level up; Gage’s piece supplies the macroeconomic why underneath the venture-scale claim.

  3. Quality-gate-as-brain re-frame and the “kit stage.” RDCO’s MAC positioning (2026-04-15-commoncog-deming-paradox, 2026-04-15-commoncog-story-about-process-improvement, 2026-04-15-commoncog-limits-of-operational-excellence) sits in exactly the kit-stage/closure transition Gage describes. We are explicitly trying to take scattered operator practices (the analyst running MAC inside their own models in BAU time) and productize them into the integration machinery that closure phases require. The piece sharpens the case that MAC’s value is not the gates themselves but the integration capability — turning kit-stage tacit skills into a flow-of-capital product, which is exactly Ford’s role with the tractor.

  4. Kit-stage invisibility validates the “vault as accumulated memory” architecture. Gage notes that kit-stage tinkering doesn’t get documented because “there is no time to waste on documentation” and the project may not be worth documenting anyway. This is a direct argument for ../04-tooling/rdco-state-ownership-architecture and the vault’s role as the exception — RDCO bets that capturing the tacit kit-stage learning is the moat. The vault is the integration-machinery prototype.

Where Gage is incomplete: he doesn’t address how the transition from kit-stage to closure actually happens inside any specific firm. Kline & Pinch describe Ford observing farmers and responding; the parallel today would be a hyperscaler observing solo operators’ LLM workflows and productizing them — which is happening (Anthropic Claude Code, OpenAI ChatGPT Agents) but the observation mechanism is murky. RDCO’s vault-driven approach is one bet on what makes the observation tractable. Worth flagging as an open question for cross-check work.

One contradiction-of-degree, not direction: Gage frames the bifurcation endpoint as inevitable. Elad Gil’s “units of cognitive labor” framing implies a more crowded venture-scale layer between the hyperscalers and the solo operators. Both could be right — agent-deployer service firms (mid-tier integrators) may emerge precisely because the integration cost is now affordable to ~10-person shops with the right tooling. RDCO is positioned in that gap.