THE LAST CLASS: How AI Killed "Learn to Code" and Revealed Education's Existential Crisis

In 18 months, artificial intelligence destroyed the safest career advice of a generation and exposed a brutal question: If machines can think, what should humans learn?

TL;DR

The "learn to code" promise—15 years of guaranteed job security—collapsed in 18 months as AI mastered programming. ChatGPT writes code better than bootcamp graduates, GitHub Copilot makes developers 55% more productive (meaning companies need 40% fewer programmers), and junior developer positions have vanished. But coding's fall is just the canary: AI now automates content writing, legal research, financial analysis, and even builds other AI systems with minimal human help.

This isn't the calculator debate redux—calculators required humans to understand problems, but AI does the understanding too. We're not automating arithmetic; we're automating thinking itself. And if coding jobs aren't safe, nothing is.

The crisis forces an uncomfortable question: Are we heading toward Aldous Huxley's Brave New World, where 10-20% have meaningful work (the new Alphas), 20-30% do necessary service jobs (Betas), and 50-70% live on universal basic income with nothing to do but consume entertainment (Gammas)? If so, should education identify and sort children early—or is that path to dystopia?

The only assessment AI can't fake is the Socratic method: oral examination through spontaneous questioning that reveals actual understanding. Richard Feynman discovered this in 1952 Brazil, finding students who could recite physics perfectly but couldn't explain why things fall. Now AI can recite everything and pass written tests, but it can't sustain real-time dialogue or demonstrate understanding under probing questions. Education must return to what worked 2,400 years ago—or accept that credentials mean nothing.

The next 5-10 years will determine whether we rebuild education around uniquely human capabilities (judgment, collaboration, creativity requiring lived experience) or stumble into a caste system where having a job becomes a privilege and surveillance satellites ensure compliance. We're choosing right now, mostly by not choosing at all.

The Collapse: From "Job Guarantee" to Worthless in 18 Months

November 2022: The Day Everything Changed

For 15 years, the advice was universal: "Learn to code. You'll always have a job."

The evidence seemed irrefutable. Software developer positions were growing at 22% annually versus 4% for all other occupations. Starting salaries hit $70,000-$120,000. Coding bootcamps promised 12 weeks of training for an $80,000 job, guaranteed. Parents pushed children toward computer science. Career changers mortgaged their futures on bootcamp tuition. The logic was bulletproof: Coding jobs can't be outsourced (we tried, it failed), can't be automated (too complex, too creative), and are infinitely in demand (every company needs software).

Then ChatGPT launched.

Within weeks, programmers realized this wasn't another developer tool—it was a replacement. ChatGPT wrote working code from plain English descriptions, debugged programs faster than Stack Overflow, and explained its work better than most textbooks. In March 2023, GPT-4 arrived and wrote complex applications indistinguishable from experienced programmers' code. GitHub Copilot studies showed 55% productivity gains, which meant companies needed 35-40% fewer developers for the same work.

By late 2024, the junior developer position had effectively disappeared. Why hire two junior developers when one senior engineer with AI assistance produces more? The coding bootcamp-to-job pipeline—which had rescued thousands from dead-end careers—broke completely. Graduates who once waited 2-3 months for offers now searched for 6-12 months and often found nothing.

Then came the most devastating proof point: AI building AI. Claude Code, an AI coding assistant, recently developed sophisticated agentic AI systems—cutting-edge software that makes autonomous decisions and uses tools—in weeks with minimal human intervention. This work traditionally required 2-3 senior engineers over 6-12 months. If AI can build AI systems, everything simpler than AI development is fully automatable.

The tech layoffs began explicitly citing AI as the reason. Google, Meta, Duolingo, and hundreds of startups eliminated positions, not because revenue fell but because AI made the humans unnecessary. YC-backed companies launched with 2-3 founders instead of 10+ employees, using AI to build the entire product. Job postings plummeted 30-40% from the 2022 peak while applications per posting exploded to 500-1,000.

"Learn to code" died faster than any career advice in history. And if coding isn't safe, nothing is.

Why This Time Is Different: The Calculator Analogy Breaks Down

The 1970s Calculator Debate

When scientific calculators became affordable in the 1970s, mathematics educators split into opposing camps. Would calculators make students dependent, erasing their number sense and computational skills? Or would calculators free students from tedious arithmetic, allowing them to focus on concepts and problem-solving?

By the 1990s, calculators won. The SAT allowed them starting in 1994. The resolution seemed vindicated: calculators automated arithmetic, but humans still needed to understand problems, set up equations, choose appropriate methods, and interpret results. The calculator was a tool that amplified human capability without replacing human thinking. Engineers using calculators could solve more complex problems faster, even if they lost some mental arithmetic fluency.

This became the template for thinking about automation: tools make humans more productive rather than replacing them. Computer spreadsheets eliminated calculating by hand but created demand for financial analysts. Word processors replaced typewriters but expanded the writing profession. Each wave of automation moved humans up the value chain to higher-order thinking.

Why LLMs Shatter the Pattern

Large language models don't amplify human thinking—they replace it.

Consider a calculus problem: "A ladder 10 feet long leans against a wall. The bottom slides away at 2 ft/sec. How fast is the top sliding down when the bottom is 6 feet from the wall?"

With a 1985 calculator:

Student must understand "related rates" concept
Draw a diagram
Set up the equation: x² + y² = 100 (Pythagorean theorem)
Differentiate: 2x(dx/dt) + 2y(dy/dt) = 0
Solve for dy/dt when x=6: Find y = 8, calculate dy/dt = -1.5 ft/sec
Calculator helps with: √(100-36) = 8 and final arithmetic
Calculator contributed ~5% of the solution

With 2025 ChatGPT:

Student copies problem into ChatGPT
AI produces: complete solution with diagram, step-by-step explanation, all calculations, interpretation of the negative sign
Student's contribution: copy, paste
AI did 100% of the problem

Calculators automated execution but required human understanding. AI automates understanding, setup, and execution. The human role shrinks to prompt engineer.

The scope difference is equally stark. Calculators handled a narrow domain—arithmetic and basic functions. LLMs handle nearly everything education asks students to do: write essays in any style, solve problems across all subjects, generate and debug code in any language, reason about ethics and policy, create presentations and graphics, translate languages, summarize research, explain complex concepts. They're general-purpose cognitive tools, not specialized calculators.

And critically, LLMs are particularly good at coding because programming is a language task and they're language models trained on billions of lines of GitHub code. We thought programming was safe because it's "abstract and creative," but the specific type of abstraction and creativity in programming—pattern recognition, applying common solutions to similar problems, generating syntactically correct expressions—is exactly what LLMs excel at.

We're not automating arithmetic. We're automating thinking. And that changes everything.

The Detection Collapse: Why Teachers Can't Catch AI Cheating

The Arms Race That Was Already Lost

When ChatGPT launched, universities rushed to deploy AI detection tools. GPTZero, Turnitin's AI detector, Winston AI, and others promised to identify AI-generated text by analyzing patterns: perplexity (how predictable the text is), burstiness (variation in sentence structure), and word choice quirks that distinguish AI from human writing.

Independent testing revealed these tools are worse than useless—they're harmful.

Stanford researchers found GPTZero and Turnitin achieved only 60-65% accuracy with false positive rates of 15-20%. That means in a class of 30 students, 4-6 innocent students would be falsely accused of cheating. UC Berkeley discovered the tools were catastrophically biased: ESL students were flagged as using AI 48-62% of the time because their simpler sentence structures matched AI patterns. Meanwhile, students who generated text with AI and made light edits were caught only 12-18% of the time.

OpenAI built its own detector and shut it down after seven months, admitting it correctly identified AI text only 26% of the time while falsely labeling human text as AI 9% of the time. The company concluded bluntly: "Our AI text classifier is not fully reliable."

Universities began abandoning detection. Vanderbilt stated in December 2024: "The use of AI detection software is strongly discouraged. Such tools have unacceptably high false positive rates and have been shown to exhibit bias against non-native English speakers." Harvard advised instructors: "Do not rely on AI detection software. If you have concerns about a student's work, discuss it directly with the student."

The technical problem is fundamental. AI is designed to mimic human writing—it's trained on human text with the explicit goal of being indistinguishable. As AI improves, the distinction becomes impossible by definition. And simple workarounds defeat detectors: asking AI to "write in a human style," mixing outputs from multiple AI tools, or making modest edits all bypass detection.

Even if detectors achieved 90% accuracy (they don't), that would be insufficient for high-stakes academic decisions. False accusations of cheating are devastating—students face grade appeals, academic misconduct records, and psychological trauma. One false accusation destroys the system's credibility; teachers lose faith in the tools and students learn they can challenge accusations successfully.

The enforcement problem is equally insurmountable. Students can use AI at home on any device. They can memorize AI-generated essays and rewrite them by hand. They can have AI generate outlines or research summaries, then write in their own words—is that cheating or legitimate assistance? The line between using AI as a tutor and using it as a replacement is subjective and unenforceable.

By 2025, the consensus was clear: AI detection had failed. The war was over before it began. Teachers would need to assume every take-home assignment could be AI-generated and design assessments accordingly—or accept that grades and credentials had become meaningless.

The Scale of the Problem: How Many Students Are Using AI

The Usage Data

Stanford and Georgetown studies found 60% of college students report using ChatGPT or similar AI for coursework, with 40% using it frequently. Of those, 65% use AI for writing assistance, 58% for problem-solving, and 45% for generating entire assignments. Most alarmingly, 48% of undergraduates admit to submitting AI-generated work as their own, with 22% doing this regularly.

K-12 usage is growing exponentially. Common Sense Media found 26% of teenagers used AI for schoolwork in fall 2024, up from 19% in early 2024—likely 35%+ by spring 2025. High school students report 35-40% usage rates, middle schoolers 15-20%, with elementary students beginning to use AI for homework help.

But the raw numbers understate the crisis. Students exist on a spectrum:

Legitimate learning support (25-30%): Using AI as a tutor—asking for concept explanations, checking work for errors, generating practice problems. This is arguably beneficial: AI as teaching assistant.

Substantial assistance (40-45%): AI generates outlines or first drafts that students revise substantially, or AI solves problems with explanations that students study and recreate. Students learn something but work is heavily AI-assisted. This gray area is where most usage falls.

Minimal modification (20-25%): AI generates complete assignments that students edit cosmetically—changing a few words or sentences to "make it their own" while contributing minimal original thought.

Complete delegation (10-15%): AI generates the assignment, student submits it verbatim or with only formatting changes. Pure cheating with zero learning.

Teachers cannot reliably distinguish among these levels. Even level 2 users may not be developing the capabilities the assignments were designed to teach. And the percentages ensure that in every classroom, multiple students are using AI substantially.

Student testimonials reveal the logic. Positive uses emphasize AI as accessibility tool: "I have ADHD and essays are really hard for me to organize. I use AI to create an outline and structure, then I write it in my own words. It's like having a tutor who helps me organize my thoughts."

Negative uses rationalize efficiency: "I submit maybe 60-70% AI-written work. Honestly, most assignments are busy-work. Professors assign three essays a week expecting us to spend hours on each. That's not realistic. AI lets me focus on what actually matters." Another student: "Everyone uses it. If you don't, you're at a disadvantage. Kids who use AI get better grades with less work. It would be stupid not to."

Teachers face an impossible situation. One high school English teacher: "I can't tell anymore. I read an essay that sounds too sophisticated for the student, but maybe they just improved? Or got help from a parent? I can't accuse them without proof. So I'm stuck giving A's to essays I suspect are AI-written." A college professor: "I'm spending more time trying to catch cheaters than teaching. It's exhausting and depressing. I didn't become a teacher to be a detective."

The academic system's fundamental compact—students demonstrate learning by completing assignments independently—has collapsed. Everything submitted outside direct supervision is now suspect.

The Feynman Problem: Education Was Already Broken

Brazil, 1952

Richard Feynman spent a year teaching physics in Brazil in the early 1950s and encountered something that disturbed him profoundly. The students were brilliant—they could recite textbook definitions flawlessly, answer standard questions perfectly, and pass exams with high marks. But when he asked simple questions requiring actual understanding, they were lost.

He asked: "When you see light reflected from a pool of water, where are you seeing the reflection from—the surface or the bottom?"

The students couldn't answer. They had memorized that the index of refraction of water is 1.33 and could calculate angles perfectly, but they had never actually thought about what light does.

Feynman's devastating observation: "I could ask a question, which the students would answer immediately. But then if I asked the question in a slightly different way, they couldn't answer it at all. They had memorized everything, but they didn't know what anything meant."

The Brazilian system had created what Feynman called "cargo cult education"—students went through the motions of learning (attending lectures, taking notes, passing tests) without developing understanding. They learned that success meant reciting the right words, not grasping what the words meant. The result: students could ace exams but couldn't solve real physics problems or explain phenomena they experienced daily.

Feynman concluded: "I don't think I did them any good. I think I just gave them more garbage to memorize."

The Problem Has Gone Global

What Feynman found in 1952 Brazil is now everywhere. Students worldwide can memorize formulas, pass standardized tests, get high GPAs, and graduate with honors—while being unable to explain concepts in their own words, apply knowledge to novel situations, recognize when they're wrong, or solve problems that aren't exactly like textbook examples.

This is the grade inflation catastrophe documented in the California K-12 analysis: 75% of CSU freshmen need remediation despite graduating high school with good grades. High school GPAs rose 6 percentage points while SAT college-readiness scores fell 4 points—a 10-point divergence that's statistically impossible unless grading standards collapsed. Faculty report: "If I retained standards, I would fail over half of my class."

Students receive A's without mastering content, discover the deception when they fail college placement tests, and pay the price in wasted tuition and debt for degrees they can't complete. The system optimized for credentialing (graduation rates, good GPAs) rather than learning (actual understanding).

AI Makes the Feynman Problem Infinitely Worse

In Feynman's Brazil, students at least engaged with material superficially—they spent hours memorizing. They could recite definitions accurately even if they didn't understand them. The system produced students who looked educated but weren't.

With AI, students can get perfect grades without even memorizing, let alone understanding. They never engage with content at all—AI does everything. They can generate flawless-looking essays, solutions, and projects, receive A's, and graduate having learned nothing. The transcripts show success but represent zero capability.

We've automated the Feynman Problem. And unlike the Brazilian students, today's students can't even recite the material—they'd need to ask the AI what they supposedly wrote.

The Economic Earthquake: If Coding Jobs Aren't Safe, What Is?

The Domino Effect

The coding collapse isn't isolated—it's a preview. AI is now automating or threatening nearly every profession that seemed secure:

Law: AI performs legal research, reviews documents, and drafts contracts. Junior lawyer positions (doc review, research) are disappearing. The path to becoming a senior partner—which required years of junior work—is closing.

Medicine: AI diagnoses from scans (often more accurately than radiologists), suggests treatment protocols, and generates documentation. Radiology and pathology face particular disruption, though hands-on patient care remains relatively safe.

Accounting: AI handles bookkeeping, tax preparation, basic audits, and financial analysis. Entry-level positions vanishing, routine accounting highly vulnerable.

Journalism: AI writes news articles, generates sports and financial summaries, and researches stories. Entry-level reporting jobs disappearing—BuzzFeed and CNET already use AI to write content. Only investigative journalism and opinion writing remain relatively protected.

Content Creation: AI writes marketing copy, generates social media content, creates graphics and videos. The creative professional sector that was supposed to be uniquely human is being automated first.

The pattern is consistent: entry-level positions disappear (AI does routine work), mid-level positions threatened (AI augments senior people, reducing the need for middle managers), and senior positions relatively safe (judgment, accountability, relationships). But you can't reach senior positions without entry and mid-level experience. Every profession faces the same broken career pipeline.

The Jobs That Remain

If we analyze what AI cannot yet do well, a disturbing pattern emerges:

AI-resistant work includes:

Skilled trades (plumbing, electrical, carpentry): Physical work in varied, unpredictable environments that robots handle poorly
Healthcare requiring touch (nursing, physical therapy, surgery): Requires human presence, judgment, and accountability
High-level judgment (CEO decisions, legal strategy, medical diagnosis in ambiguous cases): Requires human accountability and stakeholder trust
Authentic creative work (art requiring lived experience, authentic storytelling): AI can imitate but lacks genuine human experience
Interpersonal work (therapy, teaching, complex negotiation): Requires human connection and trust

These fall into two categories: either physical work that's hard to automate (trades, hands-on healthcare) or cognitive work requiring judgment where humans must be accountable (leadership, strategy, complex professional services). The vast middle—routine cognitive work from coding to content creation to financial analysis—is being hollowed out.

The uncomfortable math: Goldman Sachs estimates 300 million jobs globally could be affected by AI automation, with two-thirds of U.S. occupations exposed to some degree of automation. McKinsey projects 30% of work hours could be automated by 2030, accelerated by generative AI. Even conservative estimates suggest 40-50 million U.S. jobs substantially disrupted over 10-15 years.

How many new jobs will be created? AI training and oversight might generate 1 million positions. New industries we can't foresee might add more, but historically, new job creation doesn't match displacement in the short term. The likely gap: 30-40 million people needing something to do.

This leads to the question nobody wants to ask: What if there simply aren't enough jobs?

The Brave New World Scenario: When Having a Job Becomes a Privilege

Huxley's Warning

Aldous Huxley's 1932 novel Brave New World imagined a future society explicitly stratified into castes:

Alphas: Intelligent, creative, leaders and managers
Betas: Skilled workers, technical roles
Gammas, Deltas, Epsilons: Progressively less capable, doing progressively more menial work

The castes were created through genetic engineering and psychological conditioning. People were content with their station because they were engineered from birth to want nothing else. Epsilons didn't resent Alphas—they couldn't conceive of wanting to be Alphas. When distress arose, citizens took soma, a perfect drug with no hangover that produced contentment and prevented discontent.

Huxley's insight: This society wasn't collapsing from internal contradictions—it was stable. That was the horror. People were content in their subjugation because they'd been made incapable of imagining anything better. This wasn't an Orwellian boot stamping on a human face forever. It was a gentle, comfortable tyranny where people loved their servitude.

The 2025 Economic Version

We're building Huxley's world not through genetic engineering but through economic logic:

The Productive Class (New Alphas): 10-20% of population

Have meaningful work: AI research, senior leadership, creative elites whose work commands premium, skilled professionals requiring complex judgment, entrepreneurs
High incomes ($150K - millions)
Status, purpose, social connections through work
The people for whom work is available and meaningful

The Service Class (New Betas): 20-30% of population

Jobs requiring human presence but not high status: healthcare workers, skilled trades (plumbers, electricians), hospitality, K-12 teachers, caregivers
Modest incomes ($40K-$80K)
Some purpose from work, but often exhausting and thankless
The people who do necessary work AI/robots can't yet do

The Basic Income Class (New Gammas/Deltas): 50-70% of population

No meaningful employment available: former office workers, coders, writers, designers, customer service, manufacturing, retail—all automated
Government stipend ($20-30K/year in universal basic income)
Consume entertainment, social media, games, AI-generated content
The people for whom there simply isn't work

Why UBI Becomes Economically Necessary

The math is simple but brutal. If 50-70% of the population can't find work because AI does it cheaper and better, but the AI-driven economy still produces abundant goods and services, who buys the products? A consumer economy requires consumers with money.

Economic logic forces redistribution: tax AI productivity and automation, distribute purchasing power via universal basic income. This solves the economic problem (maintains consumer base) but creates an existential social problem: what do people do with their lives?

Research on long-term unemployment reveals the psychological crisis even when income is provided. Depression rates are 2-3x higher among the unemployed than employed, anxiety is elevated, substance abuse increases, and suicide risk rises—even when financial needs are met through unemployment benefits or disability.

Work provides more than money: daily structure and routine, social connections through colleagues, status and identity, a sense of purpose and contribution, achievement through task completion and career progression. Give people money but no work, and you don't produce wellbeing—you produce depression and despair.

Historical examples confirm this. Rust Belt manufacturing communities lost primary employment in the 1980s-2000s. Government support was available through welfare and disability, but communities never recovered. The result was opioid epidemics, multigenerational unemployment, loss of social cohesion, and "deaths of despair" from suicide, drugs, and alcohol. Income support didn't prevent social collapse.

The Soma Infrastructure

Huxley's soma was a drug distributed by the government to keep citizens content. We're building the 2025 equivalent through market forces, no conspiracy required:

We already have proto-soma:

Social media with infinite scroll design, algorithmic feeds optimized for engagement (addiction), dopamine hits from likes and comments
Video games increasingly sophisticated and immersive, with progression systems triggering reward pathways, capable of consuming 40-60+ hours weekly
Streaming content from Netflix, YouTube, Twitch providing infinite binge-able material
Pornography free and abundant, increasingly sophisticated with AI and VR
Legal drugs with marijuana legalized in many states, psychedelics following, prescription medications widely available

The AI enhancement already arriving:

AI companions providing emotional support (Replika, Character.AI), substituting for human relationships, customized to be perfectly agreeable
AI-generated entertainment infinite and personalized, perfectly tailored to preferences, algorithmically ensuring engagement
Virtual reality increasingly immersive, offering complete escape from physical reality

The dystopian package assembles itself:

Basic Income Class receives $25K/year (enough for small apartment, food, utilities)
Free high-speed internet (economic necessity)
Cheap/free access to social media, AI entertainment, games, VR, AI companions, legal drugs
Everything needed for infinite distraction and chemical contentment

This doesn't require government conspiracy. Tech companies already optimize for engagement (profit motive). AI makes content generation nearly free (economic logic). VR advances naturally (technological progress). Drugs are being legalized (social liberalization). Market forces and technology naturally create the infrastructure.

Government would just need to provide UBI (economic necessity to maintain consumers) and avoid interfering. People will eagerly choose the soma because it feels good, fills the void left by lack of purpose, is legal and socially acceptable, and beats facing meaninglessness.

We won't need totalitarian force. People will choose the comfortable cage themselves.

The Question Nobody Wants to Ask: Should Education Sort Children Early?

The Dystopian Logic

If the economic reality is that 10-20% will have productive work, 20-30% will do service work, and 50-70% will be on basic income, education faces a brutal question: Should we identify early which children will be in which category and educate them accordingly?

The meritocratic answer: "No! Every child deserves the same opportunity. Education should develop everyone's full potential."

The dystopian realist answer: "If there objectively aren't enough Alpha/Beta positions for everyone, pretending otherwise creates false hope. Better to identify aptitudes early, train people for realistic futures, and avoid setting up millions for disappointment."

This isn't hypothetical. Many countries have practiced explicit sorting:

Prussia/Germany (19th-20th century): At age 10, students were sorted into three tracks—Gymnasium for university-bound professionals, Realschule for technical/commercial roles, Hauptschule for trades. The system efficiently produced a stratified workforce but severely limited social mobility.

Britain (1944-1976): The "11-plus" examination sorted students at age 11—top 25% to grammar schools and university track, everyone else to secondary modern schools and vocational training. Ostensibly meritocratic, but middle-class children were vastly overrepresented in grammar schools. The system was abolished as inequitable.

Current tracking (de facto everywhere): U.S. schools use AP/Honors versus regular versus remedial tracks. Many countries maintain separate vocational and academic schools. We're already sorting; we just pretend we're not.

The Ethical Dilemma

Arguments for early identification:

Efficiency and realism: Not everyone can be a brain surgeon. Some children show clear aptitudes early. Why teach advanced calculus to someone headed for HVAC work?

Reduced disappointment: The current system tells everyone "you can be anything," but most won't be high-status professionals. Setting realistic expectations versus false hope.

Specialization enables excellence: Musicians and athletes start young for mastery. Why not trades and service work too?

Arguments against:

Self-fulfilling prophecy: Label a child "not academic" at age 10, and lower expectations produce worse outcomes. Sorting often reflects class and race, not ability.

Late bloomers exist: Many successful people struggled early. Einstein was considered slow as a child. Early sorting misses late development.

Psychological damage: Being told "you're not smart enough" at a young age is devastating, creates permanent underclass who internalize inferiority.

Aptitudes change: A 10-year-old's interests aren't a 20-year-old's. Early specialization prevents exploration. What if you're sorted into a field that gets automated?

What Honest Sorting Might Look Like

If society moved toward explicit tracking:

Alpha Track: Education for the Productive Class (10-20%)

Selection based on cognitive ability (abstract reasoning, creativity), personality traits (conscientiousness, openness, persistence), social skills (leadership, emotional intelligence), and family background providing resources and networks.

The curriculum would emphasize deep conceptual understanding over memorization, Socratic questioning and oral examination, collaborative problem-solving and leadership development, specialization in areas of strength combined with Renaissance breadth, mentorship with experts, and AI literacy focused on understanding how to direct AI systems.

Assessment through oral defense of original work, demonstrations of capability, and portfolio reviews—no AI-completable assignments.

This is essentially what elite education already provides at Exeter, Harvard, and Stanford. We just pretend it's available to everyone when it's not.

Beta Track: Education for Service and Skilled Work (20-30%)

Selection based on conscientiousness and reliability, practical intelligence and hands-on problem-solving, interpersonal skills for healthcare and service roles, and physical capability for trades.

The curriculum would emphasize solid academic foundations (literacy, numeracy, applied math and science), extensive hands-on training from early age, apprenticeships starting at age 14-16 in chosen trade or profession, and technical skills with real-world application.

By graduation, students would have 2-4 years of practical experience and be ready for journeyman training or healthcare/teaching certification programs.

Assessment through practical demonstrations (can you actually do the work?), licensing exams providing external validation, and portfolio of completed work with employer recommendations.

Gamma Track: Education for Basic Income Class (50-70%)

This is the darkest question. If the majority will be on basic income with no employment, should we educate them as if they'll have careers (false hope), educate them for meaningful lives without employment (honest), or provide minimal education and maximum soma (dystopian)?

An honest Gamma curriculum would focus on basic literacy and numeracy at functional level, creativity and artistic expression, philosophy and meaning-making for life without employment, community engagement and volunteer work, media literacy teaching resistance to addictive technology, physical education and outdoor skills, and relationship skills since human connection becomes primary life content.

The goal: prepare the majority for lives where they won't have employment but could still have meaning, community, and satisfaction—if we teach them how.

Why Explicit Sorting Is Both Logical and Horrifying

We're probably heading toward explicit or implicit sorting because economic reality dictates limited productive work, efficiency pressure suggests specialized training, political convenience favors manageable explicit underclass, and technological capability enables AI prediction and sorting from early age.

But explicit Alpha/Beta/Gamma education is dystopian because sorting creates self-fulfilling outcomes, violates human dignity and potential, wastes late bloomers and unusual talent, creates unstable resentful underclass, and eliminates possibility of transcendence.

The uncomfortable truth: We already sort dishonestly through zip code, family wealth, and educational access. Explicit sorting might be more honest but no less cruel.

The Only Assessment AI Can't Fake: The Socratic Method

Why Oral Examination Defeats AI

AI cannot respond in real-time to spontaneous follow-up questions that probe understanding. It cannot demonstrate comprehension through genuine dialogue. It cannot recover from being challenged on a point or show depth of understanding under questioning. It cannot handle novel questions that deviate from prepared material or apply knowledge flexibly to new scenarios in conversation.

A human student who actually learned can explain concepts in their own words, answer "why" questions beyond "what" questions, adjust explanations based on questioner's understanding, apply knowledge to new scenarios, and admit what they don't know while reasoning from what they do.

A student who used AI to fake learning may have memorized AI explanations but reveals shallow understanding when probed, cannot deviate from prepared answers, cannot apply concepts to novel scenarios, and cannot explain "why" because they never actually understood.

Example: Calculus oral exam

Surface question: "Explain the fundamental theorem of calculus."

Both the student who learned and the student who memorized AI explanations can state the theorem.

Probing question: "Why is it called 'fundamental'? What problems existed before this theorem?"

Student who learned: "Before FTC, calculating areas required exhaustion methods—very tedious. Derivatives were studied separately. FTC showed these are inverse operations, which meant we could use derivatives to find integrals. That unified two big areas of calculus."

Student who memorized: "Because it's important? It connects derivatives and integrals which are fundamental to calculus?" [Vague, no depth]

Application question: "Show me how to use FTC to find the area under f(x) = x² from 0 to 2."

Student who learned works through it, explaining each step.

Student who memorized struggles without notes or AI assistance.

Novel scenario: "A student says: 'I used FTC backwards—took the derivative of an integral and didn't get the original function.' What went wrong?"

Student who learned: "They probably forgot constants of integration, mixed up limits, or the original function wasn't continuous."

Student who memorized: "I'm not sure. The theorem should always work?" [Doesn't understand conditions and limitations]

By the fourth question, the difference is obvious. Understanding can be probed; performance cannot.

Historical Precedent

Oral examination is the traditional form of assessment. Oxford and Cambridge used viva voce examinations for centuries, still required for PhD defenses. France's Le Grand Oral high school exit exam requires students to present and defend topics under questioning. Medical oral board exams test physicians' reasoning. Law school's Socratic method cold-calls students with probing questions.

These persist because they assess actual understanding, not just ability to produce documents.

Modern Implementation

Elementary (ages 6-11): Teachers already do this naturally—regular one-on-one conversations about learning, asking students to explain in their own words.

Middle school (ages 11-14): Regular oral presentations with Q&A, 10-15 minute discussions with teachers about topics, peer teaching where students explain concepts to classmates.

High school (ages 14-18): Formal 20-30 minute oral exams with teachers or panels, student presentations defending work with questioning, maybe one oral exam per major unit.

College/University: Dissertation-style defenses even for undergraduate papers, oral finals instead of or in addition to written exams, presentations with Q&A, discussion-based seminars where participation demonstrates understanding.

Challenges and solutions:

Time-intensive (oral exam takes 15-30 minutes per student): Don't examine everything orally—written work for practice, oral exams for high-stakes assessment. Stagger throughout term rather than all at once. Students can present to peers with teacher questioning (others observe and learn).

Anxiety (some students struggle with oral communication): Practice builds skill and reduces anxiety. Provide accommodations like extra time or familiar settings. Use variety of formats—conversation, peer teaching, recorded explanation. Focus on understanding, not performance.

Subjectivity (judgment call on quality): Use clear rubrics defining proficient explanation. Occasionally use panels or second opinions. Record exams for review if disputes arise. Faculty calibration sessions on what constitutes strong responses.

Not suitable for all content: Some things best assessed in writing (extended research), some through doing (code that runs, art created). Use mixed assessment—oral for understanding, written/practical for execution.

Early Adopters and Results

Georgetown University writing program (Fall 2024) replaced final essays with oral presentations and defenses. Student feedback: "More stressful but more engaging. Couldn't fake it—had to actually understand." Professor: "I learned more about what students actually understood. Some who wrote excellent essays couldn't explain basic concepts."

Arizona State University courses moved to oral finals. Professor: "Cheating concerns disappeared. Students can use AI to prep, but in the exam they have to demonstrate actual understanding."

Stanford increased oral presentations and defenses. Finding: Students study differently—focus on understanding, not just completion—when they know they'll have to explain verbally.

The pattern is clear: oral examination forces actual learning because understanding can't be faked in real-time dialogue.

Alternative Assessment Strategies

Beyond oral examination, other approaches make AI use irrelevant or beneficial:

Open-AI Testing: Instead of banning AI, require its use and assess higher-order skills. Example history assignment: "Use AI to generate three different interpretations of WWI causes. Evaluate strengths and weaknesses using primary sources. Synthesize your own argument explaining which interpretation best fits evidence." AI can generate interpretations easily, but evaluating them critically and creating novel synthesis tests human capability AI struggles with.

Process-Based Assessment: Grade the observable process, not just product. Example: Week 1, student develops thesis in class through discussion. Week 2, writes introduction and outline in class. Week 3, brings draft for peer workshop. Week 4, revises in class based on feedback. Cannot be fully AI-generated because process is observable.

Collaborative Work with Individual Accountability: Group produces deliverable (can use AI), but each member must individually explain and defend any part. Each member randomly assigned sections to present. Must be able to explain teammates' work, showing team communicated and integrated ideas. Can't fake understanding in questioning.

Authentic Performance Assessment: Assess ability to do real-world tasks where AI assistance is realistic but human judgment matters. Example: Teacher education—teach actual lesson to real students with video recording, assessed on classroom management, student engagement, adaptation to needs. AI can write lesson plan but can't teach the class.

What Humans Can Do That AI Cannot (Yet)

Education must focus on capabilities that remain distinctively human:

Novel synthesis across distant domains: Connecting ideas from completely different fields in original ways. Example: "How might jazz improvisation principles inform agile software development?" AI retrieves information about both but struggles with genuinely novel synthesis not in training data.

Ethical reasoning in complex novel situations: AI can recite frameworks but struggles with nuanced application. Example: "An autonomous vehicle must choose between two harmful outcomes. What should guide the decision?" Requires weighing values, considering context, defending position.

Creative work requiring lived experience: Art, writing, music drawing on authentic personal experience and emotion. AI can imitate but lacks lived experience that gives work depth and resonance.

Collaborative problem-solving in real-time: Working with others to solve novel problems, adapting based on contributions, building on ideas through dialogue. Cannot be faked—requires genuine thinking.

Physical/manual skills: Performance (music, athletics), building things, medical procedures. AI can advise but not execute.

Metacognitive reflection: Thinking about one's own thinking, understanding personal learning process, identifying gaps in knowledge. AI can simulate but not genuinely reflect.

Judgment in ambiguous high-stakes situations: Medical diagnosis with incomplete information, legal reasoning in novel cases, leadership decisions with imperfect data. Requires experience, intuition, accountability.

Education should develop these through practice making decisions with incomplete information, creating work drawing on authentic experience, collaborative projects requiring real-time adaptation, hands-on physical work, and explicit metacognitive reflection on learning processes.

The Path Forward: Three Scenarios

Scenario 1: Intentional Renaissance (Optimistic)

We rapidly redesign education around human capabilities AI cannot replicate:

Oral examination and Socratic dialogue become standard
Focus on judgment, collaboration, creativity requiring lived experience
Hands-on and project-based learning with observable process
AI literacy teaching effective use as tool, not replacement
Education for human flourishing, not just employment

Society implements UBI thoughtfully with meaning infrastructure:

Enough income for dignity ($30K+, not poverty-level)
Universal without stigma
Subsidize arts, culture, community organizations
Preserve autonomy and community spaces

Result: Painful transition but eventual adaptation. Humans develop AI-augmented capabilities. Work is redefined to include care, creativity, community contribution.

Likelihood: Low. Requires unprecedented coordination, speed, and political will.

Scenario 2: Stumbling into Dystopia (Most Likely)

Education continues current trajectory:

Students use AI to complete assignments, receive credentials without learning
Detection fails, assessment remains AI-completable
Graduates appear educated but possess no capabilities
Employers discover credentials meaningless, add extensive skills testing
Entry-level positions vanish across professions

Society implements UBI out of economic necessity but without meaning infrastructure:

Minimal stipend barely supporting survival
Soma infrastructure fully developed (infinite entertainment, addictive technology)
Surveillance prevents organizing or dissent
Explicit or implicit caste system emerges
Basic Income Class comfortable but purposeless, medicated by digital distraction

Result: Stable dystopia. Huxley's Brave New World realized through economics, not genetics. Most people content in meaninglessness because facing void is too painful.

Likelihood: High. This is default outcome of current trajectories without major intervention.

Scenario 3: Collapse and Chaos (Pessimistic)

The transition overwhelms institutions:

Education unable to adapt, credentials become worthless
Mass unemployment without adequate safety net
Mental health crisis at unprecedented scale
Social instability and violence
Political extremism and potential system breakdown

Likelihood: Moderate if UBI not implemented or implemented too late. Transitions under stress often produce instability before new equilibrium.

What Should We Do? The Feynman Answer

Richard Feynman would approach this crisis by observing what's actually happening, identifying core forces, predicting outcomes, and asking whether we can change forces or must adapt to outcomes.

His analysis might be:

"The economic logic is simple. If AI can do work cheaper and better, companies will use AI—that's rational. If most humans can't find work, you give them money or they starve and revolt—that's also rational, not conspiracy.

"The question is: What do people do when they don't need to work? Some will create, learn, build communities—that's great. But many will be bored and depressed because work gave them meaning. And bored, depressed people with infinite entertainment will choose entertainment. That's rational too—it feels better.

"So you end up with society where a small group works and most people are entertained but purposeless. Is that stable? Maybe. Is it good? That's values, not science.

"If we don't want that outcome, we have to change something. We can't change economic logic (not using AI means falling behind). We can't change human psychology (we are what we are). We can't un-invent AI.

"We can change culture—figure out how humans live meaningfully when traditional sources of meaning are gone. That's the only feasible option.

"And for education specifically: Whatever we do, be honest. If we're sorting kids, admit it. Don't lie to the bottom 50% that they can all be engineers when jobs don't exist. Don't let students graduate thinking they learned when they didn't.

"Whatever we teach, test whether they understand it. That means talking to them, asking questions, making them explain. The Brazilian problem was students could pass exams without understanding. The AI problem is students can pass without learning anything. The solution is the same: talk to them. Probe until you find limits of understanding.

"And whatever we're preparing people for, be honest about it. The worst thing is lying about what you're preparing them for. That's cruel.

"So: Figure out what you're trying to do. Do it honestly. Measure whether it worked."

The Brutal Timeline: The Window Is Closing

2025-2027: Critical decision window

What happens in the next 2-3 years determines outcomes. Will institutions cling to detection and traditional assessment (losing battle) or embrace radical redesign (oral examination, AI-resistant evaluation, human capability focus)?

Coding bootcamp graduates are already unemployed for 6-12 months. Universities are seeing spiking remediation needs and dropout rates as pandemic-affected cohorts arrive with inflated transcripts and minimal actual learning. Employers are adding skills testing because degrees no longer signal competence.

2027-2032: Divergence and sorting

Institutions will split into winners and losers. Some schools successfully redesign around oral examination and authentic assessment, producing graduates with verifiable capabilities. Others cling to traditional assignments, produce graduates with meaningless credentials. Employers discover which schools' graduates can actually perform.

The job market bifurcates: 10-20% find productive work, 20-30% do necessary service work, 50-70% face unemployment or underemployment. UBI becomes politically necessary as consumer base collapses without purchasing power.

2032-2040: New equilibrium

Either education successfully adapted (focus on human capabilities, oral examination, AI as tool not replacement) and society created meaning infrastructure for those without employment, or we stumbled into stable dystopia where having a job is a privilege and the majority live on UBI consuming infinite AI-generated entertainment.

The students who were kindergarteners in 2020 will be entering college. We're choosing their future right now.

Conclusion: The Choice We're Making

The coding collapse revealed education's existential crisis. For 15 years, "learn to code" was the answer to economic anxiety—the one safe harbor in a changing world. In 18 months, AI destroyed it. And if coding isn't safe, nothing is.

But coding's fall merely exposes what was always true: education has been credentialing performance rather than developing understanding. Feynman discovered this in 1952 Brazil. Grade inflation has accelerated it. AI makes it impossible to ignore.

We're at a civilizational decision point. We can:

Rebuild education around what AI cannot fake: Oral examination revealing understanding, hands-on work requiring physical presence, collaborative problem-solving demonstrating real-time thinking, creative work drawing on lived experience, judgment in ambiguous situations requiring human accountability. Prepare students for lives that may not include traditional employment but could include meaning, community, and purpose—if we teach them how.

Or stumble into Brave New World: Explicit or implicit sorting into Alpha/Beta/Gamma classes, credentials that signify nothing, universal basic income without meaning infrastructure, infinite digital soma keeping Basic Income Class docile, surveillance satellites ensuring compliance, and stable dystopia of comfortable meaninglessness.

The SpaceX satellites are going up. The AI coding assistants are getting better. The entry-level jobs are disappearing. The education system will either adapt or become irrelevant.

And the adaptation required isn't retooling—it's complete reinvention. We must return to the Socratic method that worked 2,400 years ago because it's the only assessment AI can't fake. We must focus on developing humans, not training workers. We must be honest about constrained opportunities rather than selling false hope.

The Brazilian physics students could recite perfectly but couldn't explain why things fall. Today's students can get perfect grades without thinking at all. The solution Feynman identified remains valid: Stop testing what the computer can do. Start testing what the human can do. And the only reliable test is conversation—probing questions that reveal the depth or limits of understanding.

We're choosing right now, mostly by not choosing at all. The default is dystopia. The alternative requires courage, honesty, and unprecedented speed.

The last class to learn coding as a guaranteed career has graduated. The first class to face education in the AI age is starting kindergarten this fall.

What we teach them—and how we assess whether they actually learned—will determine whether we get renaissance or dystopia.

The window is open. But it's closing fast.