Back

Concepts & Patterns

The Future of Knowledge Work

AI is commoditizing knowledge — facts, skills, and expertise — shifting competitive advantage toward wisdom: emotional clarity, discernment, and connection. Knowledge workers face restructuring; wisdom workers are ascendant.

Created Apr 9, 2026·Updated Jun 15, 2026

See also: AI User Perspectives for empirical data (Anthropic's 81K-person survey) on what users want from AI and where they see risk — grounding the theoretical labor-market analysis below.

The Knowledge-to-Wisdom Shift

For decades, knowing things was the scarce resource. School systems, credentials, and job interviews all measured knowledge accumulation. The result: over 1 billion knowledge workers valued for what they knew — lawyers, engineers, consultants, programmers.

AI disrupts this entirely. Models can now absorb entire fields of study in days and outperform human experts in physics, law, and engineering simultaneously, around the clock. Facts and skills are becoming commoditized.

Uncertainty is universal, not bottom-up. The intuition that AI anxiety concentrates at the bottom of the stack — hitting juniors, new grads, and people without leverage — is wrong. Uncertainty runs through every layer simultaneously. Foundation model providers compete in a race that resets every few weeks as leaderboards shift. App-layer companies watch each new model capability blur their differentiation. Incumbents bolt AI onto existing products knowing it probably isn't enough. Knowledge workers watch agents improve monthly. New grads enter a job market that's changing shape in real time. Everyone is simultaneously the disruptor and the disrupted. There is a useful psychological distinction here: what AI produces is diffuse anxiety (no pinpointable threat, impossible to fight directly) rather than specific fear (a clear threat you can respond to). Previous tech shifts — web disrupting print, mobile disrupting web — were concrete enough that you could rationalize future from relic. AI feels different because no one occupies a stable vantage point from which to assess it.

Joe Hudson (executive coach to OpenAI, Google DeepMind, Anthropic, and Apple executives): many of his clients are "building the technology that will make their own skills obsolete" and are racing to develop capabilities AI cannot replicate.

The leverage has shifted from what you can do to how you show up while doing it. Competence is now table stakes.

The Three Wisdom Skills

Emotional Clarity

The ability to recognize emotions, feel them, and move forward without being obstructed. Not emotional suppression or management — actually having the emotion without being captured by it.

Why it matters: decisions are fundamentally emotional, not logical. When the emotional center of the brain is impaired, IQ stays the same but simple decisions take hours. Procrastination is an emotional struggle, not a time management problem. As AI handles more execution tasks, humans become more focused on decision-making — where emotional clarity is decisive.

Sam Altman describes emotional clarity as "one of the most critical skills in a post-AGI world."

Discernment

The ability to see clearly and zero in on what matters — especially when drowning in data. Analysis paralysis from data overload is already the #1 source of C-suite decision failure (Deloitte, 2024).

The deeper principle: self-perception enables world perception. If you can't see yourself clearly (blind spots, limiting beliefs, false assumptions), you can't assess external situations accurately. Discernment is as much inner work as analytical skill.

Taste as a trainable muscle. Problem selection — knowing what to work on — is the highest-leverage form of discernment and the one least taught. Hamming had a habit at Bell Labs of asking colleagues what the important problems in their field were, then asking why they weren't working on them. People changed tables. The question stings because most people absorb problems (from advisors, from whatever a big lab announced, from what's trending) rather than choosing them — holding conclusions without the reasoning behind them. Schulman splits research into two modes: hunting the literature for things to improve, or choosing an outcome you genuinely want and reasoning backward to the experiments. The second mode manufactures originality because a goal you actually care about drags you into territory no survey paper covers source(https://x.com/VivsThoughts). Taste itself responds to a simple training protocol: predict the result of every experiment before running it, cover a paper's results and guess the numbers from the method, mark which releases will matter in two years and check your hit rate later. A forecast plus a correction, repeated hundreds of times, is how every good model gets trained — including the one in your head.

Connection

Deep relational presence — attuning to others, creating psychological safety, vulnerability without performance. Human connection requires mutual embodied presence that AI cannot replicate: nervous systems coregulating, micro-expressions, pulse changes.

Google's Project Aristotle finding: of all factors studied, psychological safety was the #1 predictor of high-performing teams. The Harvard longitudinal study (86 years): quality of relationships at 50 predicts health at 80 better than cholesterol levels.

The Allocation Economy

Dan Shipper's related framing: we're entering the "allocation economy" where everyone becomes a manager. You won't be judged on how much you know but on how well you can allocate and manage resources (including AI agents) to get work done. Being a great manager requires all three wisdom skills.

See Agent Proficiency — agent management is arguably the knowledge work skill with the most staying power, sitting at the intersection of direction-setting and wisdom.

Corporate Precedents

  • Satya Nadella (Microsoft) — Made empathy and emotional intelligence central to Microsoft's culture transformation. Market cap: $300B → $3T in 8 years.
  • Google's Search Inside Yourself — Mindfulness and emotional intelligence program for engineers. Measurable: lower stress, higher engagement, faster innovation. Now licensed to external executive teams.
  • OpenAI executives — Actively investing in wisdom skills through coaching, with awareness that their technical skills face obsolescence.

Knowledge Types Taxonomy

The classic framework (Polanyi, via Bloomfire):

  • Explicit knowledge — Easy to articulate, write down, share. Manuals, policies, documentation. AI makes this easier to capture, organize, and retrieve.
  • Implicit knowledge — The application of explicit knowledge. Transferable skills: negotiation techniques, sales approaches, conflict resolution, organizational culture absorption. Shared through social interaction; hard to document because it's performed subconsciously.
  • Tacit knowledge — Gained from personal experience; the hardest to express. Gut feelings, intuition, "the exact feel for the dough." Non-verbal and context-dependent — experts may not realize they possess it. Requires long-term mentorship, storytelling, and high-trust culture to transfer. This is the knowledge most at risk of being lost when employees leave.

Polanyi's deeper point (often misunderstood): The business KM community (via Nonaka) reduced Polanyi's epistemology to a simple bifurcation: knowledge is either tacit or explicit. But Polanyi's actual framework is about tacit knowing as a process, not a category. "We know more than we can tell" means knowing is always from subsidiary awareness to focal awareness — a dynamic process of indwelling, not a static bucket. The conversion model (tacit → explicit → tacit) that dominates KM literature misrepresents this entirely. This matters for AI: models can capture explicit knowledge and some implicit knowledge, but the process of tacit knowing — the embodied, contextual, indwelling experience — is fundamentally different from retrievable information.

The key challenge: most organizations use tools suited only for explicit knowledge (intranets, folders), while the most valuable knowledge (tacit) requires completely different capture methods — socialization, apprenticeship, narrative exchange.

Every job becomes a software job: Anish Acharya's observation: the ambitious view of coding agents is that almost any problem/solution can be expressed in software, making coding capability upstream of all knowledge work. Legal, comms, marketing, HR, and finance will increasingly be software-first. This shifts which knowledge types matter: procedural knowledge becomes executable code; tacit knowledge about when and why to act becomes the human differentiator.

AI's Early Labor Market Impact (2026 Data)

The Lump-of-Labor Fallacy

The "AI will destroy all jobs" narrative is a rebranding of the lump-of-labor fallacy — the assumption that there is a fixed amount of work to be done, so any productivity gain by machines must come at the expense of human employment. Keynes predicted automation would yield a 15-hour work week; instead, the labor surplus was absorbed by entirely new categories of productive endeavor. The fallacy ignores that human wants and needs are not fixed — when one layer of scarcity falls, people move to the next.

Historical Precedents

Every general-purpose technology has reorganized the economy and expanded the frontier of useful work rather than shrinking it:

  • Farm mechanization: In the early 20th century, ~33% of U.S. employment was in farming; by 2017, ~2%. Farm output nearly tripled, supporting massive population growth. Displaced workers flowed into factories, offices, hospitals, labs, services, and software — industries that barely existed before.
  • Electrification: By 1930 electricity supplied ~80% of manufacturing power, and labor productivity growth doubled for decades. Far from destroying labor demand, it created more manufacturing, sales, lending, and commercial activity — plus second-order effects from labor-saving devices (washing machines, cars) that pulled more people into higher-value work.
  • VisiCalc/Excel: Spreadsheets didn't doom bookkeepers. The U.S. lost ~1M "bookkeepers" and gained ~1.5M "financial analysts" — efficient computation led to an explosion of FP&A as an entire industry source(https://www.a16z.news/p/the-ai-job-apocalypse-is-a-complete).
  • Travel agents: Technology halved travel agency payrolls since 2000, but the overall employment-to-population ratio (age-adjusted) remained stable — displaced agents found work elsewhere. Meanwhile, remaining agents saw wages rise from 87% of the private-sector average to 99%, outpacing the broader economy source(https://www.a16z.news/p/the-ai-job-apocalypse-is-a-complete).

Each dominant economic sector has given way to an even larger successor. Tech today is bigger than finance, railroads, or industrials ever were, but still smaller as a fraction of the overall economy. Productivity gains have been a positive-sum force: the net result of delegating effort to machines is that the economy and labor market have only gotten bigger, more diverse, and more complex.

Augmentation vs. Substitution

For some jobs, AI is an existential threat; for others, it's a force-multiplier that makes those roles more valuable. Goldman Sachs estimates that AI augmentation effects more than counterbalance AI substitution effects on aggregate employment. On earnings calls, management teams mention AI-as-augmentation at roughly 8:1 versus AI-as-substitution source(https://www.a16z.news/p/the-ai-job-apocalypse-is-a-complete).

Software engineering is a canonical augmentation case: git pushes and new app submissions are skyrocketing, new business formation correlates with AI adoption, and SWE job postings (both absolute count and share of overall market) have been increasing since early 2025. AI-exposed occupations show above-trend wage growth, especially in systems design. Product manager openings are also climbing, now more plentiful than at any point since 2022. Simultaneous growth in both engineers and PMs illustrates why lump-of-labor is wrong — if AI substituted thinking 1:1, you'd expect one to decline, but both are rebounding because people are getting more done.

Anthropic's Observed Exposure Data

Anthropic's "observed exposure" methodology (Massenkoff & McCrory, 2026): New measure combining theoretical LLM capability (can an LLM theoretically speed up this task?) with real-world Claude usage data, weighting automated and work-related uses more heavily. Key insight: actual AI penetration is far below theoretical capability.

Most exposed occupations: Computer programmers (75% observed coverage), Customer service representatives, Data entry keyers (67%), Financial analysts. 30% of workers have zero AI exposure — Cooks, Motorcycle Mechanics, Lifeguards, Bartenders.

Exposed workers are not the poor — the most exposed group is 16 percentage points more likely to be female, 11pp more likely to be white, 47% higher earnings on average, and 4x more likely to have graduate degrees. AI is currently disrupting higher-educated, higher-paid professionals — not lower-wage workers.

Academic Research on Net Employment Effects

The recurring refrain from the most recent research is "no change on net, but some evidence of reallocation between jobs and tasks":

One notable exception: researchers at Stanford, Dallas Fed, and Census found entry-level roles with high AI exposure are increasingly difficult to find — but also found increases in entry-level roles where AI is augmentative.

Overall: "still no statistically significant relationship between AI and unemployment or employment growth." Hiring growth appears stronger for AI-augmented industries, weaker for AI-substitution industries. The aggregate picture is neutral but not static — some destruction, some creation, some roles deprecated, others at a premium.

Employment impact so far: No systematic increase in unemployment for highly exposed workers since ChatGPT's release. However, tentative evidence that hiring of younger workers (22-25) into exposed occupations has slowed ~14% (just barely statistically significant). Slowed hiring may not appear as unemployment if young workers exit the labor market rather than appear unemployed.

The Jevons Paradox of AI Labor

The Jevons Paradox of AI Labor (Paweł Huryn): Job postings rose +22.7% YoY (TrueUp, Mar 2026). Entry-level hiring collapsed 73.4% in one year (Ravio, 2025). Both are true simultaneously.

The 160-year-old Jevons pattern: more efficient steam engines increased coal consumption, not reduced it — efficiency made coal viable for thousands of new applications. AI is doing the same to knowledge work roles. The production step gets cheaper; demand for the judgment step increases.

  • Engineering: AI writes code faster → more projects become viable → postings up +34.1% YoY. The new projects need architects and system thinkers.
  • Security: AI finds vulnerabilities faster → discoveries multiply → triage needs more human judgment. 50% of employers can't fill these roles.
  • Product: AI collapses build time → what to build matters more than how → AI PM roles up 465%.
  • Content: AI generates 100x more → noise floor rises → curation and taste become the bottleneck.

New business formation is exploding with decent correlation to AI adoption; new apps are hitting the app store at a 60% YoY clip source(https://www.a16z.news/p/the-ai-job-apocalypse-is-a-complete). The majority of new jobs created since 1940 didn't even exist in 1940 — and in 2000, it was easy to imagine travel agents losing jobs but much harder to imagine an entire industry built around "cloud migration" a decade later.

The broken pipeline: The old career path ran through production work (do reps, build volume, become senior). AI handles the reps now. Companies need senior-level thinking; the path that produced it is narrowing. Only 1/3 of employees got any AI training; 52% of developers don't use AI agents at all.

Jevons Paradox in deep domains — Torah learning as case study: Zohar Atkins extends the Jevons pattern from labor markets into the structure of knowledge traditions themselves. For most of Jewish history, consulting the Talmud, medieval commentators, and halakhic authorities required years of language training and access to rare libraries — the cost of consultation rationed who could participate. AI tools (Yochai, Rav Dicta, base models) collapsed that cost: a teenager with a phone can now query the entire corpus in seconds. The reasonable prediction — that rabbis become obsolete — is exactly the prediction Jevons refuted about coal. Instead, cheap consultation shifts the bottleneck to chiddush (genuinely novel insight from inherited material): "seeing something true that nobody saw before, in a text that was already there." The Talmud itself states a house of study cannot stand without chiddush (Chagigah 3a) — without new insight being produced, the activity ceases to be Torah regardless of how many books are open source(https://x.com/ZoharAtkins).

The structural logic generalizes: each loosened constraint in a knowledge tradition reveals the next. When books were expensive, the binding constraint was access. Print loosened that; literacy became the constraint. Literacy spread; time became the constraint. AI loosens the constraint of consultation (finding, translating, contextualizing sources). What remains is the production of synthesis — grinding the wheat into bread. A learner consuming AI-generated source material without producing original insight has "eaten dry grain and gone home thinking she has tasted Torah." The obligation to produce insight was always theoretically universal but practically restricted to elites who could afford the preparatory labor. That economic restriction has expired.

This maps directly to the knowledge-to-wisdom shift above: the "wheat" (corpus mastery, accessible knowledge) is being technologized; the differentiator is what gets made from it — and the communities built to sustain that making. The Talmud's own ranking (Horayot 14a) stages the tension: the "master of wheat" (Sinai, who knows everything) vs. the "uprooter of mountains" (who produces original insight). The academy gave its presidency to the uprooter. AI is the new master of wheat.

Cognitive Density

The productivity paradox of AI in the workplace has a structural explanation: AI doesn't just automate labor, it changes the type of work remaining.

Before AI, a knowledge worker's day contained significant mental white space — formatting, summarizing, moving data, routine drafts. Tasks that could be done with music playing in the background, low-stakes, mechanical but necessary. This white space functioned as cognitive recovery time embedded in the workday.

AI automated that white space away. What replaced it: higher-stakes orchestration, strategic choices, and activities with much higher cognitive requirements. The day became denser. Not longer — denser. Every hour is now high-stakes work.

This explains the exhaustion many workers report despite productivity tool adoption. The Iron Man suit metaphor was supposed to augment human capability while reducing effort; instead, it removed the low-effort tasks while the high-effort tasks remained. Workers are the suit's pilot, running cognitive simulations constantly, with no idle cycles.

Connection to the Red Queen dynamic: The cognitive density problem compounds under competitive pressure. When every company is racing to adopt AI and cascade that pressure through their org (see Business Moats in AI — Fractal of AI Panic), nobody has the bandwidth to step back and notice they're running the wrong race. Cognitive density consumes all available attention, making it structurally harder to think strategically about whether the race itself is the right one.

Connection to the Ironies of Automation: Bainbridge's monitoring failure irony maps directly: when AI handles the execution, the human's job becomes continuous high-attention oversight. This is more cognitively demanding than doing the original task — supervisory roles require sustained vigilance without the natural rhythm of execution to structure time.

Ironies of Automation (Bainbridge, via Kingsbury)

Kyle Kingsbury resurfaces Lisanne Bainbridge's 1983 paper "Ironies of Automation" — originally about power plants and factories — as the essential framework for understanding AI's deskilling effects on knowledge work.

Three core ironies:

  1. Deskilling — Automation degrades the skills it's supposed to augment. When humans don't practice a skill, their ability atrophies. Software engineers report feeling less able to write code after working with code-generation models. Designers report weakened creative ability after offloading to ML. Students automating reading and writing lose "core skills needed to understand the world and develop one's own thoughts." Doctors using AI polyp detection perform worse at spotting adenomas during colonoscopies. Empirical studies are converging: an Anthropic randomized trial (early 2026) found engineers learning a Python library with AI scored 50% on comprehension vs. 67% for the manual group, with the gap widening on debugging — but within the AI group, those who asked conceptual questions scored above 65% while copy-pasters scored under 40% source(https://addyosmani.com/blog/dont-outsource-the-learning/). MIT's "Your Brain on ChatGPT" study measured EEG brain connectivity scaling down with each layer of external support; 83% of LLM-assisted essay writers couldn't quote a single line of what they'd just produced — the researchers called this cognitive debt source(https://www.media.mit.edu/publications/your-brain-on-chatgpt/). A CHI 2026 study found that when people had LLM access at the start of a task, the LLM framed the entire problem, producing measurably worse decisions even when the human did the rest — the order of AI involvement mattered more than the total amount source(https://arxiv.org/html/2603.08849v1). The tool doesn't determine the outcome. The posture does.

  2. Monitoring failure — Humans are bad at overseeing automated processes. If the system executes tasks faster or more accurately than a human, real-time review is essentially impossible. Humans also struggle to maintain vigilance over systems that mostly work — which is why journalists keep publishing fictitious LLM quotes and Tesla's former head of self-driving watched his car crash into a wall.

  3. Takeover collapse — When an automated system handles things most of the time but occasionally needs human intervention, the operator is out of practice and stumbles. Automated systems can mask failure by handling increasing deviation until catastrophe strikes, thrusting a human into an unexpected regime. Air France 447 is the canonical example: flight controls transitioned to an unfamiliar mode the pilots weren't trained for.

The labor shock spectrum: Kingsbury frames the range of outcomes from "ML turns out to be a normal technology" (massive capital write-down, labor market adapts, we muddle through) to mass displacement of knowledge workers across a broad swath of industries simultaneously — unlike previous automation waves that hit one sector at a time. The UBI solution is "hopelessly naïve": profitable megacorps already fight to avoid taxes and paying workers; no reason to believe AI companies will fund redistribution voluntarily.

Maintaining Skill While Using AI

The deskilling research points to a clear intervention: the problem isn't AI use itself but passive AI use — delegating without engaging. The same tools that produce cognitive debt can sharpen engineers when used with active learning intent.

Posture shifts that preserve skill:

  • Form a hypothesis before prompting. Write down what you think the problem is; use the model's answer to test your theory, not replace it.
  • Ask for explanation before code. In unfamiliar territory, request concepts, alternatives, and tradeoffs first. Only ask for implementation after grasping the design space.
  • Treat AI output like a junior engineer's PR. Read it, critique it, push back. Passing tests is not sufficient reason to merge.
  • Re-derive periodically. Take AI-generated code and recreate it from scratch — a calibration check on what you've quietly lost.
  • Ask the model to teach what it just did. One extra prompt ("what concepts did you use? what should I read?") changes what you take away from a session.
  • Write as a defense against self-deception. Paul Graham's observation: an idea can feel fully formed until you try to put it into words — the page finds gaps your head papers over. Darwin made this procedural, writing down any fact that cut against his theory on the spot because he'd caught his own memory deleting inconvenient evidence. A lab log — hypothesis, setup, expectation, result, updated belief — is humbling in a way no reviewer can match source(https://x.com/VivsThoughts). In an AI-heavy workflow, the temptation to skip the written-reasoning step is strongest precisely when it matters most.

Product teams are experimenting with structural interventions: Anthropic's Learning Mode uses Socratic questioning and asks users to write code before continuing; OpenAI and Google have shipped similar features. Adoption remains low — users file them under "for students" — but the same feature that helps a sophomore learn React works for a senior engineer learning Rust.

Unlearn as fast as you learn. The half-life of expertise is shrinking. Tactics that worked two months ago don't always transfer cleanly, and what was once an advantage can harden into habit. The people who stay sharp let go the fastest — admitting something no longer works and moving on without over-identifying with it. This is the inverse of the deskilling problem: deskilling happens when you stop practicing, but rigidity happens when you keep practicing the wrong thing. Both are failure modes of the same underlying challenge — maintaining an adaptive relationship with your own expertise.

Tighten the feedback loop. Research speed — and by extension, knowledge-work speed — is mostly the speed at which you discover you're wrong. This makes tooling a first-class activity: launching a run should be one command, plotting it one more, comparing two runs seconds not an afternoon. Karpathy's recipe has a step that pays for itself a hundred times: overfit a single batch before training at scale — thirty seconds, half your bugs, gone. The principle generalizes beyond ML: shrink everything until it's cheap, get it right, then spend the resources. The researcher (or knowledge worker) who can build the harness, the eval, and the data pipeline is the one whose hypotheses actually get tested; everyone else is waiting in a queue source(https://x.com/VivsThoughts).

The "ship and learn" dual metric: Shipping and learning are separate metrics. Managers and customers only ask about the first; the second is self-directed. Optimizing purely for task closure — the default UX gravity of every AI coding tool — accumulates cognitive debt invisibly. The correction is small: end sessions asking "did I learn anything today, or did I just close issues?"

Breadth as insurance. Subfields saturate — usually right after they peak on social media. The people who keep producing through transitions already know their way around the neighboring territory. Range matters as much as depth: interpretability borrows from neuroscience, eval design is mechanism design wearing a lab coat, a working sense of hardware memory hierarchy tells you which architecture papers are doomed before the benchmarks do. Shannon's 1952 trick — shrink a problem until it's nearly trivial, crack the small version, reintroduce difficulty one piece at a time — carries through more walls than any modern productivity framework source(https://x.com/VivsThoughts). Old material is criminally underpriced; this field reruns its own past on a delay.

New AI-Era Job Categories (Kingsbury)

As ML deploys broadly, new kinds of work emerge at the boundary between human and ML systems:

  • Incanters — Specialists in prompting models. LLMs respond unpredictably to threats, flattery, repetition, and lies about financial bonuses. Getting reliable output requires a craft distinct from traditional programming.

  • Process Engineers — Design safeguards and layers of review around ML outputs. "Adversarial process which introduces subtle errors to measure whether the error-correction process actually works" — pharmaceutical-plant-level safety engineering applied to AI workflows.

  • Statistical Engineers — Control errors in the models themselves. Monitor confidence, detect drift, verify that ML outputs meet quality thresholds across changing distributions.

  • Model Trainers — "A surprising number of people are now employed feeding their human expertise to automated systems." The RLHF contractor workforce — "as the quip goes, 'AI' stands for African Intelligence."

  • Meat Shields — People accountable for ML systems under their supervision. When the Chicago Sun-Times published a 64-page AI-generated slop insert, the accountability chain (freelancer → King Features → Hearst → Sun-Times) revealed how companies need human bodies to absorb legal and reputational consequences. Madeline Clare Elish calls this a "moral crumple zone."

  • Haruspices — Named after Roman diviners who read entrails. Responsible for sifting through model inputs, outputs, and internal states to explain behavior post-hoc. Deep investigations into single cases or broader statistical analysis. Could serve ML companies, their users, journalists, courts, or agencies like the NTSB.

What This Doesn't Mean

This isn't a soft-skills argument against technical depth. The framing is about what becomes differentiating when AI can match technical performance. Deep technical knowledge still matters — it's just no longer sufficient on its own.

Also relevant: AI Careers (where this connects to the bifurcation into big vs. small AI tracks) and Business Moats in AI (where the durable moats increasingly involve human judgment and relationships, not technical capability alone).

Sources

  • "Knowledge Work Is Dying—Here's What Comes Next" — Joe Hudson (Every, Apr 2026) (link)
  • "Unpacking AI at work: Data work, knowledge work, and values work" — Elmira van den Broek (Apr 2026) (link)
  • "Different Types of Knowledge: Implicit, Tacit, and Explicit" — Bloomfire (Betsy Anderson) (link)
  • "Notes on AI Apps / Feb 2026" — Anish Acharya (tweet, Apr 2026) (link)
  • "Knowledge Management and Polanyi" — Eric M. Straw (academic paper) (link)
  • "Knowledge About Knowledge" — [Tier C reference: 131K-word PDF, not fully synthesized. Covers epistemology of knowledge management.] (link)
  • "Labor market impacts of AI: A new measure and early evidence" — Maxim Massenkoff & Peter McCrory (Anthropic, 2026) (link)
  • "The Paradox Nobody's Naming: More Jobs Than Ever. Fewer People Who Can Do Them." — Paweł Huryn (tweet, Apr 2026) (link)
  • "The Future Of Everything Is Lies, I Guess" — Kyle Kingsbury (PDF, Apr 2026) (link)
  • "Don't Outsource the Learning" — Addy Osmani (May 2026) (link) — empirical studies on AI deskilling (Anthropic coding trial, MIT cognitive debt, CHI anchoring); practical posture shifts for maintaining skill; ship-and-learn dual metric
  • "Running Faster to Go Nowhere: The AI Adoption Trap" — BuccoCapital / Educated Guess (Apr 2026) — cognitive density thesis; AI automating away mental white space; Fractal of AI Panic pressure cascade
  • "When Knowledge Is Cheap, Insight Is Everything: Jevons Paradox applied to Torah Learning" — Zohar Atkins (May 2026) (link) — Jevons Paradox in deep knowledge domains; chiddush as the bottleneck when consultation becomes cheap; wheat-to-bread metaphor for knowledge vs. insight
  • "Everyone is uncertain" — Grant Lee (tweet, May 2026) — universal uncertainty across the AI stack (foundation models through new grads); diffuse anxiety vs. specific fear; the unlearning imperative
  • "The 'AI Job Apocalypse' Is a Complete Fantasy" — David George (a16z, May 2026) (link) — lump-of-labor fallacy framing; historical precedents (tractors, electrification, VisiCalc, travel agents); augmentation vs. substitution data (Goldman, 8:1 earnings-call ratio); 2026 academic research roundup (NBER, Atlanta Fed, Census, Yale Budget Lab) showing no net employment effect; new business formation and app-store growth correlating with AI adoption
  • "How to be good at research" — Vivek (tweet, Jun 2026) (link) — taste as trainable muscle (predict-then-correct loop); Hamming's problem-selection question; writing as defense against self-deception (Graham, Darwin); tightening feedback loops as the core research skill; breadth as insurance against subfield saturation; Shannon's shrink-then-expand method