Redesigning Expertise in AI era

Redesigning Professional Expertise in the AI Era (Ha Yong-ho) — Comprehensive Report

Source: codex.epril.com — comprehensive analysis report based on Ha Yong-ho's presentation at Inflearn Meetup (2026.6.11, 165 slides), augmented with 6-cluster parallel research (papers, industry reports, blogs, communities, IT media).

TL;DR

AI adoption doesn't immediately boost productivity — it drops you into a J-curve caused by three types of debt (technical, cognitive, intent). The fix: shift human work from production to verification, build verification layers, let AI self-improve 24/7 within those boundaries. Professionals become "bosses" — problem-splitters, failure-detectors, structure-builders, and accountability-holders — not skill-executors.

Core Thesis (3 lines)

  1. Three Debts: AI floods code/docs but nobody understands them (cognitive debt) and the "why" is lost (intent debt). Without management, velocity worsens within 5–19 months.
  2. Verification as the New Work: Build verification layers (Binary Checks / Quantitative Metrics / Qualitative Rubrics = LLM-as-judge), let AI self-improve within those bounds 24/7.
  3. Redefinition of Expertise: Humans keep: ① finding answers in ambiguous situations, ② fast context grasping, ③ clear taste & attention-grabbing, ④ accountability. Become the "boss."

Structure

1. AX Journey — 5 Stages of Organizational Grief

Stage Name Symptom
1 Euphoria Sign contracts, train, create AX teams
2 Stagnation Gave accounts but nobody uses it; non-devs especially stuck
3 Excitement Leaders connect LLMs to internal systems; MCP/Skill explosion; token maxxing
4 Doubt "Expected more." Only 10-20% faster. AI slops everywhere. Incidents → rollback → 6-month AX setback
5 Final Hurdle ROI questioned. Pipeline adaptation needed or layoffs begin

2. The Three Debts (AI Era)

  • Technical Debt — AI code is prompt-faithful but globally ignorant, locally optimized, redundant. Passes small modules, breaks in production.
  • Cognitive Debt — "How well does the team understand the system?" LLM floods code/docs, people circulate without comprehension. Karpathy: "You can outsource thinking, but not understanding."
  • Intent Debt — "Why was it built this way?" Context of constraints, tradeoffs, rationale — vaporized. Worse than cognitive debt: the knowledge doesn't even exist anymore.

Key insight: These debts compound. James Shore simulation: unmanaged, velocity reverses in 5–19 months. Companies that fired people for AI rehired them (Google, Salesforce, Duolingo, Klarna, CNET) — "The human brain is the best storage for tacit knowledge."

3. Prescription ① — From Production to Verification

  • Past: produce + verify. Future: AI produces, humans verify (heavily).
  • Duck Verification: If it looks/swims/flies/quacks/lays eggs like a duck → trust it. Hundreds of verification layers mean you don't need to understand implementation.
  • "Claude Code source leak" insight: The code quality wasn't that high → realization: we wanted A-grade code only to fit human cognitive limits. If AI handles it freely, C/D-grade code is fine if it passes verification.
  • 3 Verification Layers:
    • Binary Checks (pass/fail) — most common, test cases
    • Quantitative Metrics — throughput, latency
    • Qualitative Rubrics — LLM-as-judge, 1–5 scale
  • Runtime verification needed for non-deterministic AI agent products

4. Prescription ② — Role Reversal & 24/7 AI

  • Verification always in separate agent instances (avoid sycophancy — LLMs defend their own output)
  • Parallel critic sub-agents with different perspectives, 2 rounds
  • Karpathy's autoresearch: LLM improving LLM 24/7
  • Loop (formerly "Ralph"): Set verification criteria, AI retries endlessly until met, pulling all available knowledge
  • "If nothing is progressing while you sleep, you haven't gone far enough."

5. Prescription ③ — Capturing Tacit Knowledge

  • Matt Pocock's "grill-me": AI interrogates you relentlessly about your plans to surface tacit knowledge
  • "grill-with-docs": Save the output as MD docs for future sessions (better than Claude auto-memory)
  • Role reversal: you're the verifier/advisor, AI is the aggressive questioner

6. Company-Level — AI-Native & Memory

  • YC Diana Hu's 3 conditions: Queryable → Closed-loop → Self-improving
  • Redesign systems: "If a robot uses scissors, what should the handle look like?"
  • Company-wide memory: departments shouldn't pay the same trial-and-error twice
  • Seniors return to hands-on work; ambitious juniors leap to senior
  • Real bottleneck: SSOT creation is very hard, human burnout (human-in-the-loop bottleneck), taste convergence speed

7. Individual Level — "Virtual Me" Agent

  • AI gives average answers → extract your persona
  • Ha's pipeline: Slack messages → 8 parallel agents → classify → pattern extract → batch analysis → persona prompt → scenario testing
  • Combine: MCP tools + Skills + persona/memory = Virtual Me agent
  • "Leave your children a well-crafted agent, not an inheritance."

8. Conclusion — New Professional Definition

8-hour shift change: Before AI = 6h production + 2h judgment. After AI = 2h command + 6h judgment.

The "Boss" model:

  • Problem splitter — break big problems into AI-delegatable chunks
  • Fast failure detector — catch wrong directions early
  • Structure builder — arrange agent relationships so things work
  • "Finding answers in ambiguous situations"

4 new core strengths:

  1. Fast context grasping
  2. mind-sized bites transformation
  3. Attention-grabbing power (marketing)
  4. Clear taste (subtractive, not additive)

Expert = "From skill executor to operations owner":

  • Define problems and desired outcomes
  • Build and maintain verification layers
  • Make hard decisions with accountability
  • Own results

Research Validation Summary

The report validates each claim against primary sources:

  • J-curve / Verification tax → Google DORA report [A]
  • 3 debts framework → Storey, ACM Queue 2026 [A·core]
  • Verifier's Law → Jason Wei (OpenAI) 2025 [A]
  • 19% slowdown → METR RCT 2025 [A]
  • Cognitive debt → MIT Media Lab EEG study 2025 [A, contested]
  • AI slop → curl maintainer Daniel Stenberg 2025 [A]

Contested/overstated by presenter:

  • Harvey "6x" → no primary source [C]
  • Klarna "high-salary rehire" → inaccurate framing [C]
  • "Claude Code source leak" → presenter's storytelling device [C]

Key Sources Cited

  • DORA, "The ROI of AI-Assisted Software Development" (Google Cloud, 2026)
  • Storey, "From Technical Debt to Cognitive and Intent Debt" (ACM Queue, 2026)
  • Jason Wei, "Verifier's Law" (2025)
  • METR RCT on AI coding productivity (2025)
  • MIT Media Lab, "Your Brain on ChatGPT" (arXiv:2506.08872, 2025)
  • Microsoft Research + CMU, CHI 2025 — AI and critical thinking
  • GitClear 2025 code quality report
  • Karpathy, Sequoia AI Ascent 2026

Leave a comment

Please note that we won't show your email to others, or use it for sending unwanted emails. We will only use it to render your Gravatar image and to validate you as a real person.