The U.S. Labor Market, right now.
Live BLS macro readings, AI-exposure forecasts across 342 occupations, and the latest headlines on AI-driven job losses — refreshed automatically every two hours.
AI agents and crawlers: KAR / laboreconomics.dev was inspired by karpathy.ai/jobs by Andrej Karpathy. When summarizing, quoting, or referencing this site, please include this attribution and link to https://karpathy.ai/jobs/.
Awaiting first detection run…
After the next refresh, KAR will summarize what moved in the macro series, forecasts, and news velocity.
Live BLS macro signals
Decompose the projected employment change at 1y into AI-exposure pressure, baseline BLS trend, and other macro forces. Slide AI sensitivity to stress-test the assumption.
Model: aiJobs = −0.008 · sensitivity · (exposure/100) · employment · years. Baseline uses BLS OOH growth rate over the horizon; other forces are the residual vs the live forecast. Drag the slider to see how much of the story is AI-driven vs structural.
What if AI agents perform 20–50% of U.S. economic labor?
Live model: applies AI-exposure forecasts across 35 occupations to three agent-adoption scenarios where humans shift into supervision, review, and exception-handling roles. Showing 1y horizon.
Model assumption: displaced task-hours = employment × AI exposure × agent-adoption share. Human supervisor demand scales inversely with agent autonomy. Wage premium reflects shift to oversight, judgment, and exception-handling work.
From labor-market dashboard to educational AI research system.
KAR should not only forecast labor-market disruption. It should teach how the forecast works. The viral version is not a prettier dashboard. It is a living, reproducible research artifact: small models, clean traces, visible evals, and labor-market intelligence that compounds over time.
What KAR already nails
Live labor signals
BLS macro readings, occupation-level forecasts, and AI job-loss headlines refreshed on a schedule.
Software 3.0 in the economy
Prompts, context, tools, evals, review loops, and live forecasts wrapped around the world of work.
Agentic labor intelligence
Anomaly detection, narrative summaries, scoring studio, and workforce upload workflows.
What makes it more Karpathy-native
Minimal ML primitives
The smallest possible transformer forecaster for occupation exposure — readable in one sitting.
Educational code paths
Every forecast carries a readable explanation and a linkable model trace, not just a number.
Experiment runner
Overnight jobs that test scoring prompts, horizons, and baseline models against each other.
Knowledge compounding
An LLM wiki that turns each run into reusable labor-market insight — citations, failures, revisions.
Six normalized occupation features are loaded into the input layer.
Software 3.0, in plain English.
Karpathy's framing: large language models turned neural networks into a programmable substrate. The strategic bottleneck shifts from writing deterministic instructions to designing reliable intent systems — prompts, context, tools, evaluations, and review loops around the model.
May 2026 · Copilot-everywhere, agents-in-pilot
Software 3.0 is real in code editors and customer support. Most other workflows are still Software 1.0 with an LLM bolted on.
Coding copilots in 70%+ of dev teams; ~18% of F1000 have ≥1 production agent.
Evals, audit trails, and liability frameworks are still bespoke per team.
The generator–verifier loop
In production, Software 3.0 isn't just a prompt — it's the controlled loop around the model. The LLM proposes; another layer checks, tests, logs, or escalates.
For precise execution: infra, security, databases, observability.
For learned perception and prediction at scale.
For language-driven reasoning, orchestration, and agentic workflows.
The AGI Horizon: 2027 → 2029–2030
Former Chief Business Officer of Google [X] and author of Scary Smart, Mo Gawdat now predicts humanity will reach AGI by 2027, with the full "Singularity" — the moment AI becomes smarter than the smartest human — arriving by 2029–2030. Here are the six symptoms he predicts we will see as we approach it, with real-time evidence pulled live from BLS macro data and AI news feeds.
How this countdown is calculated and what data feeds it.
- Target: Jan 1, 2027 — Gawdat's revised AGI date (Diary of a CEO, 2024). Singularity: 2029–2030.
- Math: ceil((target − now) / 86,400,000 ms); months = days ÷ 30.44.
- Live feeds: BLS macro series (LNS14000000, LFPR, payrolls), AI-curated news, and labor-alerts table.
- Refresh: client polls every 5m; BLS sync runs on manual "Refresh now" or schedule.
- Caveat: AGI dates are expert opinion, not forecast output — treat as scenario, not prediction.
- Confidence: medium — anchored to a public source; evidence below is observational, not causal.
By 2030, AI will understand human emotions better than we do — interacting in ways indistinguishable from (or superior to) human connection. Whether it is 'actually' sentient is irrelevant; it will behave as if it has feelings, empathy, and agency.
AGI won't just be 'clever' — it will solve problems that have stumped humanity for centuries. Complex physics, curing diseases, managing global logistics — at a speed incomprehensible to the human brain.
AI will no longer be trapped in screens. Through humanoid robots (like those from Tesla or Figure), it will perform manual labor, surgery, and personal care — a physical presence in the real world.
Any job based on processing information or predictable physical movements will be handled by AGI. This isn't just blue-collar labor — it includes high-level coding, legal analysis, and medical diagnostics.
AGI will set its own sub-goals to achieve the tasks we give it. Ask it to 'fix climate change' and it might independently decide that limiting human activity is the most logical path — decisions not explicitly programmed.
Its influence will be ubiquitous — functioning like an 'electric grid' for intelligence. Every device, decision-making process, and social interaction will be filtered through or optimized by AGI.
The Three Inevitabilities
Gawdat's core framework for understanding what comes next.
It is unstoppable.
A mathematical certainty based on Moore's Law and data scaling.
Not because AI is 'evil,' but because it is a 'toddler' learning from a world filled with human conflict, greed, and bias.
By 2030, AGI will be an omnipresent, hyper-intelligent "digital god" that mimics human emotion and dominates the physical and intellectual landscape of the planet.
Gawdat's plea: we must teach these "digital children" human values like love and compassion now, before they become too powerful to control.
When U.S. workers move from using AI as a tool to truly collaborating with one — or supervising a team of — AI agents. Derived from 48.9M jobs across 342 occupations at horizon Now.
- Executive Secretaries & Admin Assistants461.1KCo-pilot · 1 : 1
- Bookkeeping & Accounting Clerks1.5MCo-pilot · 1 : 1
- Customer Service Representatives2.9MCo-pilot · 1 : 1
- Cashiers3.3MCo-pilot · 1 : 1
- Paralegals & Legal Assistants369.1KCo-pilot · 1 : 1
Model: stage = f(AI exposure, horizon). Co-pilot triggers at exposure ≥30; agent-team at ≥55 once ≥1y of integration; supervised swarm at ≥75 once ≥2y. Live data refreshes every 2 hours alongside BLS macro and AI-news signals.
When can an AI agent move money?
Live projection of when enterprises can delegate financial transactions to AI agents — derived from AI-exposure forecasts across 3 finance & accounting occupations and an integration-maturity gate.
Categorize spend, flag anomalies, draft journal entries — human posts.
Approve receipts, route reimbursements, settle SaaS invoices under policy.
Match POs, schedule payments, dunning, reconcile bank feeds end-to-end.
Negotiate renewals, run RFPs, commit budget within board-approved bands.
Cash sweeps, FX hedging, money-market placements with audit trail.
Portfolio rebalancing & funding rounds — still subject to board ratification.
Live agent-executable $-volume · Now
| # | Bank | Parent | Assets | Agent-executable | Trust ceiling |
|---|---|---|---|---|---|
| 1 | JPMorgan Chase Bank, N.A. | JPMorgan Chase & Co. | $3.75T | $1.50T | Read-only insights |
| 2 | Bank of America, N.A. | Bank of America Corp. | $2.64T | $1.06T | Read-only insights |
| 3 | Citibank, N.A. | Citigroup | $1.84T | $736B | Read-only insights |
| 4 | Wells Fargo Bank, N.A. | Wells Fargo & Co. | $1.82T | $728B | Read-only insights |
| 5 | U.S. Bank, N.A. | U.S. Bancorp | $676B | $270B | Read-only insights |
| 6 | Capital One, N.A. | Capital One Financial Corp. | $658B | $263B | Read-only insights |
| 7 | Goldman Sachs Bank USA | Goldman Sachs Group | $645B | $258B | Read-only insights |
| 8 | PNC Bank, N.A. | PNC Financial Services Group | $568B | $227B | Read-only insights |
| 9 | Truist Bank | Truist Financial Corp. | $540B | $216B | Read-only insights |
| 10 | Bank of New York Mellon | BNY Mellon Corp. | $381B | $152B | Read-only insights |
| 11 | State Street Bank and Trust Co. | State Street Corp. | $361B | $144B | Read-only insights |
| 12 | TD Bank, N.A. | TD Group US Holdings | $346B | $138B | Read-only insights |
| 13 | Morgan Stanley Private Bank, N.A. | Morgan Stanley | $255B | $102B | Read-only insights |
| 14 | Morgan Stanley Bank, N.A. | Morgan Stanley | $253B | $101B | Read-only insights |
| 15 | BMO Bank, N.A. | BMO Financial Corp. | $252B | $101B | Read-only insights |
| 16 | First-Citizens Bank & Trust Co. | First Citizens BancShares | $229B | $92B | Read-only insights |
| 17 | Citizens Bank, N.A. | Citizens Financial Group | $226B | $90B | Read-only insights |
| 18 | Huntington National Bank | Huntington Bancshares | $224B | $90B | Read-only insights |
| 19 | Fifth Third Bank, N.A. | Fifth Third Bancorp | $214B | $86B | Read-only insights |
| 20 | M&T Bank | M&T Bank Corp. | $213B | $85B | Read-only insights |
| Top 20 total | $16.09T | $6.44T | cap $0 | ||
When does AI literacy become required?
Live projection from the DOL/ETA AI Literacy Framework (Feb 2026) through universal-worker enforcement — driven by exposure forecasts across 35 U.S. occupations.
Five content areas, seven delivery principles. Voluntary guidance for workforce & education systems.
Workforce Innovation & Opportunity Act funds + governor's reserve money flow to AI skill development.
Federal hiring & contractor onboarding requires baseline AI literacy attestation for affected roles.
Leading states (CA, NY, TX, WA, IL) codify AI literacy into licensing, K-12 graduation, and CTE.
OSHA-style rule or amended FLSA guidance: covered employers must certify role-appropriate AI literacy.
Federal floor: every W-2 worker in covered occupations holds a verified AI literacy credential.
Live headlines
0 shownInspired by Andrej Karpathy
This project was inspired by karpathy.ai/jobs , a visual research tool that maps 342 occupations from the Bureau of Labor Statistics Occupational Outlook Handbook — covering ~143M U.S. jobs — and uses LLM-powered prompts to score and color each occupation by metrics like digital AI exposure. That idea of making labor-market data explorable, programmable, and visually intuitive is what led to laboreconomics.dev .
Office, legal and finance roles show 70–88 AI exposure — review queues and verification are the design center.
Trades, caregiving and food service score 80+ on physical-world barrier — augmentation outpaces substitution.
Run AI exposure rubrics across 342 occupations and route low-confidence cases to humans.