Automation Risk, Measurement & the Shape of Work
Notes sparked by DeepSeek V3.2 and the DSA attention mechanism
I’m not an economist; these are working notes as a result of a nerd snipe. Point me to better sources and I’ll refine these.
Context
The latest DeepSeek model (V3.2) ships a variant of NSA called DSA—an attention mechanism that sits between MLA→NSA, enabling cheaper training (≈2×) with longer context (i.e., “works longer”). This is one of several trends that push cheaper, longer‑horizon reasoning. I’m still forming a view—having read a few blogs and papers—but here’s my current, opinionated take.
Core predictor of automation risk
Best single predictor: Can the task’s quality be measured cheaply and well?
If yes, it’s automatable once cost beats a human hour.
If measurement is fuzzy, stakes are high, or real‑world coordination dominates, humans stay in the loop (or on the loop). See Tetlock‑style forecasting / evaluation limits: https://arxiv.org/abs/2302.06590.
Why this mirrors human expertise formation
Historically, when we can measure quality fast, accurately, and cheaply, we get masterful.
From the Veritasium video on expertise (https://www.youtube.com/watch?v=5eW6Eagr9XA):
- Repeated attempts with feedback — Chess players log thousands of games; mathematicians can verify proofs.
- Deliberate practice — Hill‑climb slightly beyond comfort with intense focus.
- Fast feedback — Cheap loop. Anesthesiologists see instantly if a patient is stable; radiologists or recruiters often wait long for downstream results.
- Verifiable domains — Chess, surgery, physics; patterns repeat and feedback aligns with skill (unlike, say, stock‑picking).
Irony: this is also how we get better AI—fast, iterative refinement on verifiable outcomes.
Where humans stay essential
If output measurement is fuzzy, the stakes high, or it requires physical coordination, humans remain essential.
- Political pundits and economic forecasters often make one‑off predictions without consistent feedback—so they rarely improve.
- Anesthesiology remains high‑stakes; oversight is non‑negotiable.
- Recruiters can scale with AI, but it’s hard to tie any one recruiter to a hire’s long‑run success without gaming. It’s a funnel, not a single outcome.
So what does this mean?
Job content changes faster than job titles
Older practitioners near mastery will adapt by delegating to AI and vetting the AI intern’s work.
Entry‑level squeeze
- New grads in domains that reward long‑term mastery will struggle unless the work is also high‑stakes.
- Graduates whose outputs are inherently fuzzy will compete in crowded pools; luck looms larger.
Verification jobs proliferate (when outcomes are measurable & stakes moderate)
- The person solving the harder, fuzzier problem stays.
- For the rest, we’re checking the intern’s work—until we don’t.
- Work that’s “measured accurately enough” and not highly sensitive may skip human verification entirely.
Is depth (PhD) a moat?
Human value accrues to people who solve messy, intractable problems. But that still requires accurate evaluation of progress—a vicious circle.
Depth helps only if it buys you moats:
- (a) Hard‑to‑measure outputs — taste, judgment, accountability.
- (b) Regulated or trust‑based gatekeeping — licenses, fiduciary duties, safety‑critical sign‑off.
- (c) Tacit, situated know‑how — organizational capital, relationships.
- (d) Control of proprietary data/processes. “Deep but codifiable” knowledge erodes quickly.
OpenAI’s GDPval points in that direction across 44 occupations; models already compete closely with typical professional baselines on many evaluated tasks—even if messy, long‑horizon deployment lags. https://openai.com/index/gdpval/
What reshapes vs. what sustains
Reshaped
- Software & data: more time on specs, decomposition, reviews, test design, security; less raw typing.
- Legal, finance, consulting, marketing: less research/drafting/synthesis/scenario planning; more client handling, judgment, negotiation, compliance sign‑off.
- Healthcare & education (non‑procedural): less documentation, summarization, scheduling, prior auth, lesson materials; more front‑stage human work (bedside, class, coaching).
- Islands of automation (factory repetitive tasks, logistics) scale first. Physical automation is accelerating but lags software—ten‑ish years? See BMW’s humanoid‑robot pilots: https://www.bmwgroup.com/en/news/general/2024/humanoid-robots.html.
Sustain
- Relationship/Trust work: enterprise sales, therapy, coaching, senior PMs, regulators, auditors, safety engineers.
- Skilled trades & field work: electricians, plumbers, HVAC, construction leads in unstructured environments.
Macro policy problems
- Wage inequality deepens along this divide.
- If METR trendlines hold, daily human pay must clear a margin below the cost of running an AI for a six‑hour task. For fuzzier tasks, the margin narrows because you must include human vetting time.
- Imagine ultra‑cheap local models for routine tasks and pricier reasoning tiers for long‑horizon, audited workflows.
- Capability and assurance separate: in regulated or long‑horizon work, buyers pay for auditable reasoning, logs, and indemnities—things “tiny” models don’t yet deliver.
Who buys if wages fall?
If fewer young people have disposable income, who buys the goods? Demand holds if productivity raises incomes broadly or fiscal systems recycle gains (tax, wage subsidies, training). Does AI nudge us toward a nanny state? Open question.
Some notes
- As older experts retire, capability × reliability still governs adoption in high‑stakes work.
- I doubt an AI can entertain like a great sports game or concert; emotion, identity, and community beat pure content production.
How to act now
Individuals
- Pick tasks your future AI can’t be scored on easily.
- Be the person who makes the model useful.
- Own “the signature.” Licensure, fiduciary duties, regulated sign‑offs are durable moats.
- Build relationship capital.
- Grow taste and editorial authority.
Orgs
- Rewrite the workflow, not the job description.
- AI evaluation loops. Track model reliability, error costs, and human oversight minutes (HHH).
- Protect progression. Without apprentices, senior capacity decays. Use deliberate apprenticeship (sandboxed portfolios, simulated clients).
- Small models for everything; reasoning for high impact.
How to classify a job
- Measurable? Can quality be scored cheaply and objectively (tests pass, KPIs hit)?
- Messy? Open‑ended scopes, shifting constraints, institutional nuance?
- Money‑at‑stake? Liability if it goes wrong (compliance, safety, brand)?
- Motion? Non‑trivial physical dexterity/locomotion in unstructured spaces?
- Memory (horizon)? Long and interdependent (weeks of context, many actors)?
- Measurable — ease of objective scoring (higher = easier to measure; lower is safer).
- Messy — ambiguity, shifting constraints.
- Money — liability/regulatory exposure.
- Motion — real‑world coordination/physicality.
- Memory — long horizon/many interdependencies.
Resilience = \(\\,(5 - \\{Measurable}) + \\{Messy} + \\{Money} + \\{Motion} + \\{Memory}\\,\) (0–25; higher ≈ more defensible vs. automation).
Strategy, Trust & Creative
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Enterprise Account Executive (complex B2B) | 2 | 5 | 4 | 2 | 4 | 18 |
| Creative Director / Showrunner | 1 | 5 | 3 | 1 | 4 | 17 |
| Executive Recruiter / Partner‑level Talent Advisor | 2 | 5 | 3 | 1 | 5 | 17 |
| Product Strategy Lead | 2 | 5 | 3 | 1 | 5 | 17 |
| Brand Editor‑in‑Chief | 2 | 5 | 3 | 1 | 4 | 16 |
| Venture Partner / Investment Committee | 3 | 4 | 5 | 1 | 4 | 16 |
Regulatory, Risk & Finance
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Regulatory Counsel / DPO (privacy) | 2 | 4 | 5 | 1 | 4 | 17 |
| Safety/Certification Owner (med devices/aviation/pharma/auto) | 2 | 4 | 5 | 2 | 4 | 17 |
| Claims Adjudication Lead (complex disputes) | 3 | 4 | 5 | 1 | 4 | 15 |
| Financial Controller / FP&A for CapEx Projects | 3 | 4 | 5 | 1 | 4 | 15 |
| Internal Auditor (enterprise) | 3 | 4 | 4 | 1 | 4 | 14 |
| Compliance Program Owner (AML, SOX, GxP) | 3 | 4 | 5 | 1 | 3 | 14 |
Field Ops, Construction & Physical Work
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Field Service / Commissioning Lead (industrial/clinical) | 2 | 4 | 4 | 5 | 4 | 20 |
| Construction Superintendent / MEP Coordinator | 2 | 5 | 4 | 5 | 4 | 20 |
| Utility Lineworker / Grid‑Ops Supervisor | 2 | 4 | 5 | 5 | 4 | 20 |
| Live Events Production Manager | 2 | 5 | 3 | 4 | 4 | 18 |
| Electrician / HVAC Lead Tech | 3 | 4 | 3 | 5 | 3 | 17 |
| Facilities Ops Manager / Data Center Ops | 3 | 4 | 4 | 4 | 4 | 17 |
Tech Ops, Security & Engineering
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Site Reliability Engineering (SRE) Incident Commander | 3 | 5 | 5 | 1 | 4 | 16 |
| CISO / Detection Engineering Lead | 3 | 5 | 5 | 1 | 4 | 16 |
| Senior Product Manager (platform) | 3 | 5 | 4 | 1 | 5 | 15 |
| Data Governance Lead | 3 | 4 | 4 | 1 | 4 | 14 |
| AI Governance & Risk Lead | 2 | 5 | 4 | 1 | 4 | 17 |
| Solutions Architect (complex enterprise) | 3 | 4 | 4 | 1 | 4 | 14 |
Healthcare, Life Sciences & Education
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Anesthesiologist (lead) | 3 | 4 | 5 | 3 | 3 | 15 |
| Surgeon (attending) | 3 | 4 | 5 | 4 | 3 | 16 |
| Clinical Trials Operations Lead | 2 | 4 | 5 | 2 | 5 | 19 |
| Senior Therapist / Counselor | 1 | 5 | 3 | 1 | 4 | 17 |
| Charge Nurse / Unit Manager | 3 | 4 | 4 | 4 | 4 | 16 |
| Principal Investigator (lab) | 3 | 4 | 4 | 2 | 5 | 15 |
| Instructional Dean / Head of School | 2 | 4 | 4 | 1 | 5 | 18 |
Government, Policy & Community
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Policy Negotiator / Government Affairs | 2 | 5 | 4 | 1 | 5 | 19 |
| Public–Private Infrastructure PM | 3 | 5 | 5 | 2 | 5 | 19 |
| Community Relations & Permitting (energy) | 2 | 5 | 4 | 2 | 5 | 18 |
| City Emergency Management Lead | 3 | 5 | 5 | 3 | 4 | 17 |
| Procurement Lead (strategic sourcing, gov) | 3 | 4 | 4 | 1 | 4 | 14 |
Sales, Partnerships & Customer‑Facing
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Head of Strategic Partnerships (regulated sectors) | 2 | 5 | 4 | 1 | 5 | 19 |
| Enterprise Customer Success Executive | 3 | 4 | 4 | 1 | 5 | 16 |
| Channel/Alliances Director | 3 | 4 | 4 | 1 | 4 | 15 |
| Solutions Sales Engineer (complex) | 3 | 4 | 4 | 1 | 4 | 15 |
| Key Account Manager (public sector) | 3 | 4 | 4 | 1 | 4 | 15 |
Supply Chain, Manufacturing & Logistics
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Supply‑Chain Network Design & Resilience | 3 | 5 | 5 | 2 | 5 | 19 |
| Plant Operations Manager | 3 | 4 | 5 | 4 | 4 | 17 |
| Quality Lead (complex assemblies) | 3 | 4 | 4 | 3 | 4 | 15 |
| Logistics Program Manager (multi‑modal) | 3 | 4 | 4 | 2 | 5 | 15 |
| EHS Director (environment, health, safety) | 3 | 4 | 5 | 3 | 4 | 16 |
Programs, Enterprise IT & Change
| Role | Measurable | Messy | Money | Motion | Memory | Resilience |
|---|---|---|---|---|---|---|
| Enterprise Platform Rollout PM (ERP/CRM) | 3 | 5 | 5 | 1 | 5 | 18 |
| Change Management / Org Design Lead | 2 | 5 | 4 | 1 | 5 | 19 |
| M&A Integration Lead | 3 | 5 | 5 | 1 | 5 | 18 |
| Program Director (cross‑BU) | 3 | 5 | 4 | 1 | 5 | 17 |
| Data Migration & Cutover Lead | 4 | 4 | 4 | 1 | 5 | 16 |
Crowd‑source your own
The rubric above is intentionally rough‑cut. Fork the table, change scores, and propose updates. Crowd taste beats solo taste.