AI, Ashby Engineering, and the future

URL SCAN: AI, Ashby Engineering, and the Future
FIRST LINE: Since August 2025, more than half of the new code hitting Ashby's production systems has been AI-generated, yet customer issues remain broadly stable.

TEXT ANALYSIS: The Discontinuity Autopsy

1. THE DISSECTION — What This Text Is Actually Doing

This is a transition management document from a company in active structural mutation, dressed as engineering philosophy. It is simultaneously:

• Internal copium delivery system — Reassuring engineers that human judgment remains central while their jobs are being hollowed out in real-time
• Recruiting theater — "We're not like those companies, you still matter here" to prevent talent hemorrhage
• Prestige signaling to HN — Demonstrating sophisticated AI integration to maintain credibility with the engineering caste that reads this site
• Primary source data — The actual deployment metrics (50%+ AI code in production since August 2025, ~40 PRs per engineer) are the real story; the philosophical framing is the decoration

The author knows something has fundamentally shifted. The article's entire structure — the ground rules, the two modes, the "think harder" admonitions — is a coping architecture built to manage cognitive dissonance between "AI is doing our work" and "humans are still essential."

2. THE CORE FALLACY — The Fatal Misread of Value

The article's central thesis:

"Your value as an engineer was always weighted in your judgment... AI will be a larger shift than we've seen before. That shift is already here."

This is correct about the shift and catastrophically wrong about the conclusion.

The fallacy: Mistaking the location of value for the durable source of value.

Ashby's model is:

Code generation (commoditized) + Human judgment (the scarce thing) = Engineer's value

The DT thesis reveals what Ashby is blind to:

Code generation (automated) + Human judgment (the next thing AI takes) = Temporary moat

When the author says "thinking deeply is part of why we're successful," he is describing a transitional competitive advantage. Not a permanent one. The entire argument rests on human judgment remaining scarce and irreplaceable. But this assumption is structurally untenable — because judgment itself is a cognitive task, and cognitive tasks are exactly what AI is eating through.

The article even admits this inadvertently:

"LLMs are biased toward generating new code rather than reusing what exists. Left unchecked, they'll build you a codebase that works but that no one can navigate."

This is not a LLM bug. This is LLM default state. The "solution" — human reviewers pushing toward simplicity — is itself a judgment task. When that judgment is also automatable (and it will be), the reviewer role disappears too. The article preaches "think harder" as if thinking harder is a sustainable competitive moat. It is not. It is a lag defense with a fixed expiration date.

3. HIDDEN ASSUMPTIONS — The Smuggled In Fragility

Assumption 1: "Judgment" is a stable category that will remain human-native.
The article treats judgment, taste, and customer understanding as constants. They are not. They are cognitive functions being progressively automated. The moment AI can reliably determine "which abstraction is right for this context" and "what does the customer actually need," this entire value proposition collapses.

Assumption 2: Ashby's engineers will remain employable because they understand customers.
This assumes:
- Customer understanding is a sustainable differentiator (not a data aggregation problem that AI solves better than individual humans)
- Ashby will continue to need more sophisticated judgment than AI provides (optimistic extrapolation of current complexity)
- The labor market can absorb displaced code-writers as judgment-workers (demand-side assumption with no evidence)

Assumption 3: The 50% AI/50% human split is stable and desirable.
This is not an equilibrium. It is a transition point on a curve that trends toward 90%+ AI generation. The article treats this as "we figured it out" when it is "we are in the early stage of a process with no stable endpoint at 50%."

Assumption 4: "We do not measure token usage" is a virtue.
Not mandating AI use and not measuring token consumption is presented as quality control. It is actually uncontrolled diffusion of AI delegation. Engineers will naturally migrate toward maximum delegation because it is personally easier. The "quality" is maintained by individual discipline, not structural constraints. This is not scalable governance.

Assumption 5: Code quality compounds — and that is good for human engineers.

"Code quality has always mattered. Now it compounds."

This is true. It is also the mechanism of human obsolescence. As code quality becomes more machine-readable and structured, AI systems get better at navigating it, generating against it, and improving it. The human's "improved comprehension of the codebase" becomes less of a differentiator as AI comprehension becomes near-perfect.

4. SOCIAL FUNCTION — What This Text Does For Its Audience

5. THE VERDICT — Systemic Judgment

This article is not wrong. It is dangerously incomplete.

Ashby is a genuine data point. 50%+ AI code in production with stable customer outcomes is real. The internal tooling (skill files, Git metadata SQLite DB, custom code review tools) is sophisticated. The engineering culture of thoughtful delegation is probably better than average.

What the article misses entirely:

The DT mechanism does not require AI to fail for humans to become economically redundant. It requires only that AI continue improving at cognitive tasks — and judgment, taste, customer understanding, and architectural reasoning are cognitive tasks. The article documents the early stage of this process and mistakes it for the end state.

The specific DT trajectory Ashby is on:

• Now (2025): Engineers as judgment layer above AI code generation. 50% AI code. Viable.
• 1-3 years: AI code generation approaches 80-90%. Human review shifts to "abstractions and architecture." Still viable for top engineers.
• 3-7 years: AI begins handling architectural decisions, abstraction recommendations, customer needs inference. Human judgment layer shrinks to a thin coordinating function.
• 7-10 years: Coordination and oversight become AI-native functions. The remaining human roles are structural (legal entities, customer relationships as social bonds, regulatory compliance). The productive contribution of "engineers" approaches zero at scale.

The article's actual contribution: It is a detailed field report of P1 (cognitive automation) advancing through software engineering. Every metric, every quote, every internal tool is evidence for the DT thesis, not against it. The philosophical framing is the narrative scar tissue over a wound the author can feel but cannot name.

The brutal read: Ashby's engineers are being trained to become curators of AI output. That is a survivable role — briefly. The curation function is itself a cognitive task. When AI can curate AI output, the curator is redundant. The article optimizes for the next 2-3 years of comfortable transition. It provides zero defense against the structural outcome.

BOTTOM LINE: Ashby is a well-managed transitional entity executing Phase 1 of the DT mechanism — cognitive automation of execution — with genuine sophistication and some temporary moats. The article is a beautifully written account of human adaptation to obsolescence. It treats the adaptation as the survival strategy. It is not. It is the hospice care for the role being automated. The quality of the adaptation is irrelevant to the structural outcome.