AI-Wiki

The AI Skills Nobody Is Teaching (And Everyone Needs) — Ethan Mollick on A Bit of Optimism

9 min read

The AI Skills Nobody Is Teaching (And Everyone Needs) — Ethan Mollick on A Bit of Optimism

Be honest: AI makes you a little nervous. … Ethan Mollick says we’re underestimating our own agency in the age of AI. Instead of worrying about what AI will do to us, we should focus on what we choose to do with it. … In this conversation, Ethan shares the practical playbook most of us are missing and makes the case that our experience, taste, and point of view aren’t things AI replaces. They’re exactly what make us better at using it. In this episode you’ll learn: why young people are NOT “AI natives” (and why experience is the real AI advantage); the $20 decision that instantly upgrades how you use AI; why AI agrees with everything you say + the simple prompt that fixes it; how to make AI write in YOUR voice; the “jagged frontier”; why taste may become the most valuable skill of the AI era; how much agency we really have.

— Channel description, Simon Sinek’s A Bit of Optimism (YouTube)

A ~59-minute conversation on Simon Sinek’s podcast A Bit of Optimism with Ethan Mollick — Wharton professor, author of Co-Intelligence and the forthcoming Co-Existence: The Next Phase of AI, and writer of the One Useful Thing newsletter. Mollick positions himself as the deliberately non-bombastic AI voice (neither doomer nor zealot), focused on how people actually use the tools. The episode is a practitioner-grade tour of the wiki’s live “returns to expertise” theme from the social-science side.

TL;DR

1. Young people are not “AI natives” — experience is the AI advantage. Mollick rejects the digital-native analogy for AI: kids are “conduits to Claude” who can produce a beautiful report they can’t judge. Citing a BCG study, junior employees were often worse at using AI — they adopt it but can’t tell good output from bad. “This is a rare case where the more experienced you are — and sometimes the older you are — the better you’ll be at using AI,” because experience tells you what good answers look like and how to give instructions. (See the AEI returns-to-expertise report for the large-N version.)

2. The $20 move + “prompt engineering is dead.” His simplest advice: pay ~$20/month for one of the big three (Gemini / ChatGPT / Claude) and actively select the best (thinking) model — it defaults to a weaker one. Prompt engineering “none of that matters anymore” (you-are-a-physicist, offering bribes, etc.); if you can give instructions to a human, you can use these models. People are underutilising AI by not giving it hard enough tasks.

3. Three phases + the models/apps/harnesses stack. Three eras: pre-ChatGPT analytical AI → generative (co-intelligence, back-and-forth chat) → agentic (only ~3–4 months old; an AI that independently does work). To think about AI now, separate three layers: the model (the brains — at recording, “Opus 4.7,” “ChatGPT 5.5,” “Gemini 3.1 Pro”), apps (how you access it — chatgpt.com, Claude Code, Codex, NotebookLM), and harnesses (how the AI does things — write code, search, make images). The companies are roughly tied on model quality and “jockeying” on apps/harnesses. (This is the general-audience twin of the wiki’s agent-harness framing.)

4. GDPval — how good (and fast) the work is. Mollick leans on GDPval: experts (avg 14 yrs experience) and AI each do real, hard ~7–8-hour tasks; a third set of experts judge blind. Best models tying/beating humans rose from ~48% a year ago to ~84% now, in ~15 minutes vs ~7 hours. The practical upshot: give AI the complex job; even spending an hour to assemble and evaluate, you save ~3× time and cost — and evaluation speed (an expert skill), not generation, becomes the binding constraint.

5. The apprenticeship model just broke. We’ve trained specialists for 4,000 years via apprenticeship — juniors do grunt work, prove themselves, learn the ropes. But now “every junior person knows less than ChatGPT,” and both juniors and managers would rather delegate to AI than to a slow, flawed human. The danger: losing the talent pipeline. Solutions exist but require radical change to how we build talent. (The qualitative mechanism behind the deskilling signal in the Q4 AEI report.)

6. The voice problem — AI writing all sounds the same. AI “writes beautifully but has no voice” (or rather one singular, slightly-dramatic, m-dash-loving voice). His inbox is filling with the same “it’s not X, it’s Y” emails, which he’s started deleting. A tip: give AI a large sample of your writing, have it write a two-page style guide, paste that into custom instructions — a “slight parody” of you, but far better than “write like [famous person].” He recounts dictating an op-ed in his own voice, then having AI fact-check and correct it — “scary good,” but only because he’s written enough for it to learn his style.

7. Jobs shift their weight; the jagged frontier; taste rises. Jobs are bundles of tasks; AI does some, shifting the burden rather than removing the job. The jagged frontier: AI is unexpectedly good at some things and bad at others (your voice, getting a joke) — so the demand for your labour rises where it’s bad. Coding shifted from “write clean code” to “be an architect / engineering manager.” As everything becomes generically good and commoditised (“the death of the movie star” — we go for the franchise, not the actor), the differentiator becomes taste — variation, point of view, the director’s hand (a Wes Anderson experience). Developing taste may be a new thing we must teach.

8. Effort vs results; doubling down on human. Learning is effortful; AI shortcuts (“lifting mental weights”) let people think they’re learning when they aren’t. The deeper problem isn’t AI — it’s a results-obsessed, discomfort-avoidant culture (shareholder-value short-termism, ghosting) that AI exaggerates. Human systems weren’t built for an AI world (school wasn’t built for AI essays; sprints weren’t built for one coder shipping 100× the code). Mollick’s stance: doubling down on the human — managing egos, taking care of people, augmentation over replacement — becomes more important.

9. Privacy, education, and brains. On security: treat consumer AI like Gmail — Google already has your email; the real new risk is giving an agent access to your computer/browser (could it be convinced to send all your money?). On education: GPT-3.5 was detectable, GPT-4 wasn’t, so “use AI for everything” stopped working — but we know the pedagogy (in-class testing, AI tutors with big early learning effects, AI that won’t just hand over answers; the calculator precedent). On brains: we always trade capabilities to machines (phone numbers, cursive, slide rules); the worry isn’t memory, it’s critical thinking — but AI can help thinking if you make it a critic (“the AI is a sycophant — tell it to act like a critic,” and “tell me what I’m doing wrong with my arguments”) and use persona prompts to get a range of readers.

10. Agency. Two levels: societal (policy, organising — data-center bans float because they poll well) and, where there’s more leverage, individual — labs found an “unreasonably effective” way to mimic human thought but “don’t actually know how AI is useful in your field” (the jagged frontier again). Your biggest source of agency is using it for positive use in your own work and showing others how augmentation makes humans thrive — vs. the dangerous default (“fire everyone, profits go up”). “Technologies are neither good nor bad, nor are they neutral.”

How this source touches the wiki (dynamic capabilities)

  • digital-transforming/redesigning-internal-structures — the apprenticeship-broke / talent-pipeline problem and the weight of jobs shifts claim are structural: orgs must redesign how they build talent and bundle roles (architect/manager over clean-coder; editor over writer).
  • digital-transforming/improving-digital-maturity — the practical playbook (pay $20, pick the thinking model, give harder tasks, persona/critic prompts) and the AI-tutoring education thesis are workforce/learner digital-maturity moves.
  • strategic-renewal/organizational-culture — human systems weren’t built for an AI world; the effort-vs-results critique; doubling down on human / augmentation-over-replacement as a cultural stance.
  • contextual/external-triggers — the arrival of agentic AI and the GDPval-level capability jump as the external trigger reshaping work, schooling, and the professions.

Roles override (roles: explicit): chro, ceo, transformation-lead. The source’s centre of gravity is talent, skills, culture, and how leaders should respond — the people-and-culture roles — so the override drops the broader CDO/CIO/CSO/CMO cluster the four cells would otherwise inherit.

Linked entities and concepts

  • Concepts this source informs: durable-skills (taste, experience, judgement, evaluation), jagged-frontier (Mollick’s own term), ai-deskilling (apprenticeship broke; effort-shortcut learning), automation-vs-augmentation (doubling down on human; jobs shift weight), ai-employment-effects (lawyers/doctors guilds; jobs we can’t imagine), agent-harness (models/apps/harnesses), enterprise-ai-adoption (100× PowerPoint; redefining the work product).
  • Entities (already in wiki): Ethan Mollick, Boston Consulting Group (the junior-worse-at-AI study).
  • Promoted to an entity this ingest (second-source rule — also the named author of [[2018-05-31-sinek-nyt-the-infinite-game|the 2018 NYT Infinite Game piece]]): Simon Sinek.
  • Dangling (single-source mentions, deferred): Marvin Minsky / MIT Media Lab (Mollick’s PhD context), Wharton (his institution), Co-Existence / Co-Intelligence / One Useful Thing (his works — concept-mentions).

Source-to-source relationships

Source-quality note

Auto-generated (ASR) transcript fetched via the youtube-transcript-skill; light cleanup applied to proper nouns and model names (the ASR rendered model versions phonetically — “Opus 47,” “Chachi,” “VO comp test” for Voight-Kampff, etc.). The raw file preserves the near-verbatim ASR. Transcript provenance does not feed confidence per [§Lifecycle].

Relationships

Supports

  • Agentic coding and persistent returns to expertise — Mollick's headline — *young people are NOT 'AI natives'; experience is the real AI advantage* — is the same finding as the AEI's *returns to (domain) expertise persist*, and Mollick cites the *same* underlying mechanism (the BCG study where junior employees were *worse* at using AI because they couldn't judge whether the output was good). Where the AEI measures it from ~400k sessions, Mollick narrates the why: experts know what to expect, can steer, and can evaluate fast.
  • GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks — Mollick explicitly invokes GDPval as the load-bearing evidence for *how good the work is*: OpenAI's test where experts (avg 14 yrs experience) and AI each tackle real ~7-hour tasks, judged blind — best models tying/beating humans rose from ~48% (a year ago) to ~84% now (his recall of the figures). He uses it to argue you should give AI *harder* tasks and that evaluation, not generation, is the bottleneck.
  • Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness — Mollick gives the general-audience articulation of the **model / app / harness** three-layer framing that the wiki's [[agent-harness]] corpus uses technically — *the model is the brains, apps are how you access it (Claude Code, Codex, Notebook LM), harnesses are how the AI does things (write code, search, make images)*. He also echoes Kilpatrick's *models are all converging and good* point (OpenAI/Google/Anthropic all 'pretty solid').
  • Anthropic Economic Index: New building blocks for understanding AI use — Mollick's *the apprenticeship model just broke* — juniors know less than ChatGPT so managers delegate to AI not to flawed humans, threatening the talent pipeline — is the qualitative mechanism behind the Q4 AEI's first-order deskilling signal (the remaining human task mix loses education content as AI covers high-education tasks).
  • To Thrive Alongside AI, Focus on Mindset—Not Skillset — Both put *mindset / taste / experience* above teachable hard skill: Argenti's mindset-over-skillset thesis maps onto Mollick's *taste becomes the most valuable skill of the AI era* and *experience, taste, and point of view are what make you better at using AI, not what AI replaces*.
  • The Age of the 40-Year-Old Solo Founder Is Here — Bryant Chou on YC's The Lightcone — Same thesis from a founder's chair: Chou's *you need expertise to know what to do with the boundless intelligence in the model* and the *age of the 40-year-old solo founder* (experience + taste as the lever) is the entrepreneurship-side instance of Mollick's *experience beats AI-native youth* and *taste as differentiator*.

Graph View

Backlinks