video youtu.be Watch ↗

You Can't Compete on Cheap Models Anymore

TL;DR

Nate B Jones argues that once AI makes execution cheap, the value doesn't vanish — it moves off execution and onto deciding what is worth executing. His fulcrum is a Mitchell Hashimoto experiment (as relayed): on ordinary "implement this feature" work, a sub-$1 budget model, a ~$1.50 GPT 5.5 run, and a ~$9 Fable 5 run produced equally acceptable output — a 9× price gap for a tie. That tie, he says, is a fact about the task, not the models: work everyone already knows how to ask for is exactly where models have converged. But a second test — pointing Fable 5 at a gnarly systems-code optimization Hashimoto invented himself ($40, 2 hours) — reached a level he says he couldn't have hit alone, and no backlog, sprint, or best-practices guide generated that task. The load-bearing claim: AI can only do work someone has imagined, so the ceiling on AI's value to you was never the model or the price — it's the size of your list of things you know how to ask for. Imagination here is not an artist's gift but tacit, hands-on capability awareness ("you can't imagine with capabilities you haven't touched"), and it only pays off multiplied by cheap execution (both layers needed). He generalizes through BlackBerry-vs-Apple (comparable execution, imagination set a 100× multiplier), factory electrification and Stripe's one-day 50M-line migration (new capability's payoff waits on years of pre-built verification infrastructure — redesign the building, not just bolt on the motor), and the Fable 5 "blackout" (the model left; the questions people had imagined stayed). The prescription: keep routing execution to cheap models (table stakes), but manufacture imagination — put context-holders next to capable models with permission to pose expensive frontier questions.

Concepts introduced

Execution Commoditization — when execution gets cheap and methods are shared, output on known tasks converges across everyone; differentiation stops being an execution property.
Imagination Constraint — AI only does work someone has imagined, so the binding ceiling is the size of your list of things you know how to ask for; choosing what's worth executing is the scarce, human input, and it multiplies (not competes with) cheap execution.
Tacit Capability Awareness — you can't imagine with capabilities you haven't touched; the instinct to pose the frontier question comes from hands-on hours ("from touch"), not benchmark charts or summaries.
Capability Overhang — a new capability works on day one, but the productivity payoff waits until the organization redesigns around it (electrification; Stripe's pre-built verification infra) — "the unit of change is the building, not the motor."
Frontier Scouting — the practice of deliberately allocating imagination/scouting time to probe new capability, and organizationally manufacturing imagination by putting context-holders next to capable models with permission to make expensive frontier bets.

Held, not dropped (touched but not spun out as their own pages):

BlackBerry-vs-Apple and factory electrification — illustrative analogies for Imagination Constraint and Capability Overhang respectively, not concepts in their own right.
The Stripe 50M-line one-day migration — held as an attributed case study under Capability Overhang; the durable idea is infrastructure-before-harvest, not the number.
The Fable 5 "blackout" (shipped Tuesday, gone Friday, back) — an attributed news anecdote used as evidence that imagination outlives model availability; held, not a concept.
"Shared prompts/playbooks/channels → convergence" — the human-level mechanism of Execution Commoditization; housed there, not spun out.

Key claims

AI can only do work someone has imagined; the tools execute but don't decide what's worth executing, so the ceiling on AI's value is the size of your list of askable things — your imagination. principle → Imagination Constraint
A cheap model tying an expensive one on "implement this feature" is a fact about the task, not the models: work everyone already knows how to ask for is where models have converged. principle → Execution Commoditization
Execution and imagination are a multiplier, not a rivalry — cheap execution is the engine, frontier imagination steers; the leverage is what you put on top of a strong execution layer. principle → Imagination Constraint
The cheaper and more commoditized execution gets, the more valuable every frontier "what's now possible?" question becomes. principle → Execution Commoditization
You can't imagine with capabilities you haven't touched — the ability to pose the frontier question comes from hands-on hours ("instinct, touch"), not a benchmark chart or a summary. principle → Tacit Capability Awareness
A new capability's productivity payoff lags its availability until the organization redesigns around it — the unit of change is the building, not the motor. principle → Capability Overhang
Route daily execution to cheap models (it's a real cost lever you control) but treat that layer as table stakes, not the whole strategy; reserve a targeted, surgical frontier application for the questions that change what the execution layer is even building. (best practice) — context: a two-layer stack where execution is commoditizing fast and everyone will soon have routing. → Model-Tier Routing, Frontier Scouting
You can't hire your way to imagination with one visionary — imagination only fires next to context, which is spread across the people who do the work; the job is to manufacture it by giving context-holders capable models and permission to make bets. (best practice) — context: an organization trying to capture frontier value, not a solo operator. → Frontier Scouting
Diagnostic test (person or company): has your task list actually changed in 12 / 6 / 3 months, or are you running the old list faster and cheaper and calling that transformation? — and "who is allowed to pose a $400 question without asking anyone?" (best practice) — context: self-audit for an imagination (not tooling) shortage. → Frontier Scouting, Imagination Constraint
The video states Hashimoto's first test had three models tie on ordinary work at roughly: budget model <$1 (minutes), GPT 5.5 ~$1.50, Fable 5 ~$9 (~40 min); it also names a sub-$1 "GLM 5.2" run. observation — the source's figures/model names, groundable; flagged for a later verification pass, not adjudicated here.
The video states the second test (a self-authored systems-code optimization) cost ~$40 over ~2 hours and reached performance Hashimoto says he couldn't have achieved on his own. observation — the source's claim about a named engineer; groundable, flagged, not adjudicated.
The video states Stripe ran a migration across 50 million lines of code in one day, against a 2+-month team estimate. observation — an attributed case-study number; groundable, flagged, not adjudicated.
The video states Fable 5 "shipped on a Tuesday, was gone by Friday, now it's back" and returned into a price war. observation — an attributed recent-events claim; groundable, flagged, not adjudicated.

Why this is novel (and where it corroborates the graph)

The dominant stance is novel: the load-bearing thesis — that imagination/task-selection is the binding constraint once execution commoditizes — attaches to no existing concept and is spun out as five new nodes. The secondary stance is corroborates: the prescription "route execution to cheap models, reserve the frontier for the high-leverage slice" is an independent, economics-framed convergence on Model-Tier Routing and on Reasoning Effort Control's "reach for the top model only a small fraction of the time." Where those concepts describe the mechanism (which tier runs which sub-role, which effort dial), this source supplies the why it's only half the strategy: the routing layer is table stakes, and the durable value sits in the frontier question that decides what to build at all. It also lightly corroborates Layered Agentic Architecture's "code is commoditized; the advantage is your opinionated solution" — here generalized from code to the task itself.

Illustrated walkthrough

The video is a talking-head monologue (presenter in a book-lined room) interleaved with a dark title-carded slide deck. Visual coverage is low-confidence: max blind gap ~60s, and the frame sampler (whole-frame pixel delta) misses text-on-solid-background slide changes — so more slides transitioned than were captured. Notably, the whole porch-marketing chapter (8:07–11:30), the argument's marquee example, sampled only as talking-head frames; any slides in that ~3-minute stretch were not captured, so treat that section's illustration as absent, not slide-free.