The Extended Frontier
Notes and deep-research appendices on where capability is being extended.
- Practitioner Accounts Across Five Domains
ChatGPT Deep Research
First-person accounts of AI capability boundaries from software engineering, medicine, law, education, and creative writing.
- Reception History of the Jagged Frontier (2023–2026)
ChatGPT Deep Research
How Mollick's 'jagged frontier' concept has been received, adapted, and flattened across academic and practitioner communities.
- Automation Bias and Domain-Specific Feedback Loops
ChatGPT Deep Research
Systematic review evidence that automation bias is mediated by feedback loop richness — the same variable the extensions framework identifies.
- Cross-Trained Practitioners
ChatGPT Deep Research
People who work across multiple domains comparing how AI behaves differently in each — confirming that verifiability, feedback speed, and consequence structure explain the differences.
- Why LLMs Are Good at Code
ChatGPT Deep Research
A two-factor model: training data explains single-shot competence, use-time feedback loops explain integrated system reliability.
- Routine, Manageable AI Code Failures
ChatGPT Deep Research
How software engineering practices make AI failures routine and bounded — tests, CI, review, and staging catch errors the same way they catch human errors.
- Sellen's Paperless Office → AI
ChatGPT Deep Research
Abigail Sellen's explicit warning that AI adoption risks repeating the 'paperless office' mistake — replacing artifacts without understanding what they do.
- Situated Action and Affordance Theory Applied to AI Evaluation
ChatGPT Deep Research
How Suchman's situated action and affordance perspectives have been (and haven't been) applied to AI benchmarking and evaluation.
- The Multidimensional Frontier
ChatGPT Deep Research
Evidence that AI changes work on at least six axes beyond output quality — skill formation, craft identity, pace, accountability, communication, and constructed experience.
- Extended Practice as Robustness
ChatGPT Deep Research
Evidence that real-world work context can make AI-assisted work more robust — not less — through feedback loops, redundancy, and distributed cognition.
- Law as Counter-Example: Extensions Exist But Were Bypassed
Web Search
Law isn't extension-poor — it has Shepardizing, partner review, opposing counsel. The 700+ AI hallucination cases happened when lawyers bypassed their existing extensions.
- Code Generation Feedback Loops
Web Search
Quantitative evidence that execution-based feedback loops improve AI code generation — from LLMLOOP's 76→90% improvement to static analysis reducing security issues from 40% to 13%.
- The Harness as Context Manager
Practitioner Report
Glean's engineering team documents how agent harnesses — not models — drive reliability gains through progressive skill discovery, context compaction, sub-agents, and programmatic tool calling.