# opinionated-guide-ai-mollick-2025-10-19

## Veille

Practical AI usage guide, model selection, jagged frontier, Centaurs vs Cyborgs, OpenAI usage data, Claude/Gemini/ChatGPT - Ethan Mollick

## Titre Article

An Opinionated Guide to Using AI Right Now

## Date

2025-10-19

## URL

https://www.oneusefulthing.org/p/an-opinionated-guide-to-using-ai

## Keywords

AI model selection, ChatGPT vs Claude vs Gemini, jagged frontier, practical AI usage, information-seeking, Centaurs vs Cyborgs, BCG study, AI capabilities, invisible capability wall, free vs paid models, AI adoption best practices, democratization AI, domain expertise

## Authors

Ethan Mollick (Associate Professor, Wharton School, University of Pennsylvania ; Auteur "Co-Intelligence: Living and Working with AI" ; TIME 100 Most Influential People in AI 2024)

## Ton

**Profile:** Conversational-Professional | First-person expert | Educational | Intermediate-Accessible

Mollick maintains accessible language with academic authority. Contractions ("it isn't," "you'd think") and casual phrasings ("way less casual chat than you'd think") prevent intimidation while preserving professional structure. First-person voice establishes credibility: "Every few months I write an opinionated guide" and "I can finally give you advice based on real usage patterns." Purely educational rather than promotional intent, acknowledging uncertainties ("difficult to control what models you are using") while guiding decisions. Typical of academic thought leadership making sophisticated content digestible without oversimplification.

## Pense-betes

- **10% of humanity** uses AI weekly (late 2025)
- **Data-based article**: OpenAI released real ChatGPT usage data, enabling evidence-based advice
- **Jagged frontier**: unpredictable boundary of AI capabilities, excels at certain tasks, fails completely/subtly at others
- **BCG study**: 758 consultants, +12.2% tasks completed, +25.1% speed, +40% quality
- **BUT**: tasks outside the jagged frontier → 60-70% accuracy AI-assisted vs 84% humans alone
- **Recommended paid models** ($20/month): Claude 3.5 Sonnet, Gemini 2.0 Pro/Flash Thinking, GPT-4o/o1/o3
- **Defaults trap**: companies default to mini models (GPT-4o-mini, Gemini Flash) not flagship
- **Usage data**: less casual conversation, MUCH more information-seeking than expected
- **Mollick principle**: "invite AI to everything" because optimal applications unclear even to experts
- **AI works best**: when user has domain expertise to quickly assess output quality
- **Two success patterns**: (1) Centaurs - strategic division of human/AI work; (2) Cyborgs - deep AI integration into workflow
- **Underperformers**: +43% improvement with AI (the largest gain)
- **Top performers**: benefit less, suggests AI democratization potential
- **Invisible capability wall**: tasks where AI seems confident but produces subtly wrong answers
- **Claude lacks**: image/video generation vs competitors
- **Mollick independence**: accepts no funding from AI companies
- **Common capabilities**: advanced models, voice modes, image/doc analysis, code execution, mobile apps, deep research

## RésuméDe400mots

**Context and Central Thesis**

Ethan Mollick, Wharton professor and influential author (TIME 100 AI 2024), publishes a practical guide for choosing and effectively using AI tools in late 2025, a moment when roughly 10% of humanity uses AI weekly. The article marks a significant shift toward data-based advice rather than speculation, drawing on ChatGPT usage data newly released by OpenAI. This evidence-based approach enables recommendations targeted at real use cases.

**Conceptual Framework: Jagged Frontier**

Mollick introduces the concept of the **"jagged frontier"** - an unpredictable boundary of AI capabilities where systems excel at certain tasks while failing completely or subtly at others. This framework stems from his research with Boston Consulting Group involving **758 consultants**, revealing impressive gains: **+12.2% tasks completed, +25.1% speed, +40% quality**. However, results also show **critical blind spots**: on tasks outside the jagged frontier, AI-assisted workers perform **worse (60-70% accuracy) than unassisted humans (84%)**.

**Model Selection Recommendations**

For paid subscriptions ($20/month), Mollick recommends choosing among three primary systems: **Claude (Anthropic), Gemini (Google), ChatGPT (OpenAI)**. He suggests starting with free accounts to evaluate which system fits. Major challenge: AI companies default users to **smaller, faster, cheaper models** (GPT-4o-mini, Gemini Flash) rather than more capable flagship versions. For optimal performance, users must specifically select **Claude 3.5 Sonnet, Gemini 2.0 Pro** (or Gemini 2.0 Flash Thinking), and **GPT-4o** (or o1/o3 for complex reasoning).

**OpenAI Usage Data Insights**

Data reveals surprising patterns: **significantly less casual conversation** than expected, **substantially more information-seeking behavior**. This empirical finding allows Mollick to provide targeted advice: if free models suffice for primary use cases based on this data, users can remain confident with free options without concern.

**Best Practices and Mental Models**

Mollick advocates the principle **"invite AI to everything"** because optimal AI applications remain unclear even to experts. He emphasizes that **AI works best when users have sufficient domain expertise** to quickly assess output quality. Research identifies two successful integration patterns: **(1) Centaurs** who strategically divide work between AI and human, and **(2) Cyborgs** who deeply integrate AI into workflow, fluidly navigating across the jagged frontier.

**Cross-Platform Capabilities**

All major paid subscriptions offer advanced models, voice modes, image/document analysis, code execution, quality mobile apps, deep research features. **Claude notably lacks** image/video generation capabilities compared to competitors.

**Critical Warnings**

The **most critical caveat**: an invisible wall of AI capabilities - tasks where AI appears confident but produces subtly wrong answers. The BCG study demonstrates that **underperformers see larger gains (+43%)** using AI, while top performers benefit less, suggesting **AI democratization potential** alongside its risks. Mollick maintains **independence** by accepting no funding from AI companies, ensuring unbiased recommendations.

**Democratization and Equity**

Research results suggest AI as a **great equalizer**: underperformers improve dramatically, top performers less so. This dynamic raises profound questions about the future of work, talent competition, and whether AI gains will benefit equitably or exacerbate existing inequalities.

## GrapheDeConnaissance

- Ethan Mollick —publie→ An Opinionated Guide to Using AI Right Now (DOCUMENT, 0.99)
- Ethan Mollick —travaille_chez→ Wharton School (ORGANISATION, 0.98)
- Ethan Mollick —recommande→ Claude (TECHNOLOGIE, 0.95)
- Ethan Mollick —recommande→ ChatGPT (TECHNOLOGIE, 0.95)
- Ethan Mollick —recommande→ Gemini (TECHNOLOGIE, 0.95)
- OpenAI —publie→ données d'usage ChatGPT (DOCUMENT, 0.97)
- Ethan Mollick —a_créé→ jagged frontier (CONCEPT, 0.97)
- Ethan Mollick —collabore_avec→ Boston Consulting Group (ORGANISATION, 0.95)
- étude BCG —mesure→ 758 consultants impliqués (MESURE, 0.96)
- IA —améliore→ underperformers (+43%) (CONCEPT, 0.94)
- jagged frontier —affirme_que→ les capacités de l'IA ont des limites invisibles et imprévisibles (AFFIRMATION, 0.93)
- modèle Cyborg —s_oppose_à→ modèle Centaur (METHODOLOGIE, 0.85)
- Ethan Mollick —affirme_que→ Claude manque de génération d'images et vidéos (AFFIRMATION, 0.97)
- Deep Research —améliore→ qualité des réponses IA (CONCEPT, 0.92)
- sycophancy —est_instance_de→ risque d'adoption IA (CONCEPT, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/opinionated-guide-ai-mollick-2025-10-19/