When code gets cheap, attention becomes the scarce part. These sources ask what people still contribute. This page lists ten sources, oldest first where dated. Only one reports data at scale, and it comes from a company with a stake.
Too Fast to Think: The Hidden Fatigue of AI Vibe Coding
A first-person account from a 40-year developer who uses Claude Code and Cursor. It describes a fatigue the author had not felt before. The evidence is the author's experience and a factory analogy. No measurement or study is given.
Agentic Coding is a Trap
An opinion essay against the view that developers should orchestrate agents and stop writing code. It argues skill loss, vendor dependence and cost uncertainty are real risks. Its evidence is quotations, cited studies and the author's own practice. It ends with a workflow that keeps the human coding.
The threat is comfortable drift toward not understanding what you're doing
The thread discusses an ergosphere.blog post about two PhD students, Alice and Bob. Bob uses an AI agent and Alice does not. Commenters argue over calculators, skill loss, AI costs and whether markets value deep skill. The evidence is opinion and personal anecdote.
Agentic coding and persistent returns to expertise
An Anthropic report that classifies about 400,000 Claude Code sessions using a model. It reports who makes which decisions and how success varies with rated expertise. The evidence is transcript analysis with regression controls, not measured real-world outcomes. The authors work at the company that makes the tool.
You can't unit test for taste
A project write-up from the builder of a virtual running app. It follows a pipeline from GeoNames data to ranked points of interest. The evidence is one project's experience, including hallucinations caught by spot checks. The author says there is no ground truth, so no red/green tests.
Automation Without Understanding: Why the United States must preserve mathematical capacity in the age of AI
Park argues AI proofs are products, while human understanding is a capacity needing institutions. The essay combines an AI result, US funding cuts and four policy proposals. It rests on cited reports and analogy, not on new data.
Prevent cognitive debt by manually retyping LLM-generated code
A Hacker News thread with 545 points on a blog post about retyping LLM code. Commenters trade learning anecdotes, workplace arguments and workflow tips. The evidence is personal experience, with few citations. Opinion is split between those who see value and those who see wasted effort.
There is more to code review than (automatable) detection
A response to a paper claiming coding agents make human code review unnecessary. The post argues the paper rests on a substitution myth. Its evidence is argument from the author's view of how reviews work. It runs no test and cites no data.
The Decoherence Trap: When AI slop drowns out the signal in an endless sea of noise
Maevyn coins the term decoherence for incompatible models that accumulate in a system. The essay mixes a plumbing analogy, an interactive simulation and three cited studies. It then argues that AI cannot improve itself fast, because software quality needs real-world feedback. The evidence is mostly argument and analogy.
Developing Taste
A short essay on taste as a trained skill, from a design engineer. It gives three habits: study great work, explain why it works, and practice. The evidence is quotations and analogy, with no data. The page carries a banner about the author's course.
Connections to other subjects
These links are this wiki's own reading, not claims the sources make.
- You can't unit test for taste and Human Judgment as a Specification, from Checkable specs for coding agents. Tryggvason says a route's best stops have no ground truth to test against. PICK starts from that situation. It finds inputs that separate candidates, then asks a person. Tryggvason's per-route parameters might be tuned the same way.
- Developing Taste and A beginner's guide to modern classical music, from Consonance and tuning. Kowalski says taste grows from exposure and from explaining why things work. Muehlhauser's guide turns exposure into a route, from film scores toward harder music. Could anyone write that route for reading code?
- Developing Taste and The Concept of the Aesthetic, from Art movements and theory. Kowalski calls taste a trained instinct, and also asks you to explain why things work. Shelley traces the older thesis that aesthetic judgment is immediate. So software's talk about taste inherits a problem. Can an instinct stay immediate once it must give reasons?
- Developing Taste and The cultural evolution of distortion in music (and other norms of mixed appeal), from Consonance and tuning. Bryant and Smaldino say distortion norms persist partly because groups must differ from each other. Kowalski tells readers to follow a curated list of tastemakers. Together they raise a sharp question. How much of taste in software is a marker of belonging?
- There is more to code review than (automatable) detection and Treat Agent Output Like Compiler Output, from Compilers and interpreters. Venturini says nobody reads compiler output, so verified agent code need not be read. Allspaw lists what review does beyond detection: confusion as a finding, and named accountability. Tests can replace detection. Which of the other functions has a compiler equivalent?
- Too Fast to Think: The Hidden Fatigue of AI Vibe Coding and Viable Systems: How To Build a Fully Autonomous Agent, from Harnesses and agent loops. Kellogg calls an agent's view of its own health an algedonic signal. Building one proved difficult. Schmidt describes fatigue after an hour of fast review. A person gets that signal for free. What would an agent's fatigue be for?
- The Decoherence Trap: When AI slop drowns out the signal in an endless sea of noise and Linearizability! Refinement! Prophecy!, from Checkable specs for coding agents. Maevyn describes locally valid models that no longer fit together. Hochstein shows a formal remedy on one queue: map one model's states onto another's. The map sometimes needs an invented prophecy variable. How large a system could afford such maps?
- Too Fast to Think: The Hidden Fatigue of AI Vibe Coding and Almost anything you give sustained attention to will begin to loop on itself and bloom, from Mind, attention and practice. Karlsson says frequent switching leaves attention residue, and that sustained attention deepens experience. Schmidt reports fatigue from constant fast review. Read together, that fatigue may be more than overload. It may be the cost of never letting attention bloom.
- Developing Taste and Cultivating a state of mind where new ideas are born, from Mind, attention and practice. Kowalski tells readers to surround themselves with great work and tastemakers. Karlsson's essay describes solitude as a state free of others' opinions. Both concern how judgment forms. Perhaps taste needs exposure first, and then time away from it.
- The Decoherence Trap: When AI slop drowns out the signal in an endless sea of noise and Harness engineering: leveraging Codex in an agent-first world, from Harnesses and agent loops. Lopopolo's team runs background tasks that scan for drift and open cleanups. Maevyn says decoherence grows with change and compounds. A million generated lines make a good test of both claims. Can drift be cleaned faster than it accumulates?
Glossary
- Agent
- An AI model set up to take actions and run tools in a loop.
- Agentic coding
- Software work where an AI agent reads files, edits code and runs commands.
- Atrophy
- Loss of a skill through lack of use.
- Class field towers
- A tool from algebraic number theory used in the AI's construction.
- Classifier
- A model prompted to label each session against a fixed set of categories.
- Claude Code
- Anthropic's command-line coding agent.
- Cognitive debt
- A growing gap between the code a person ships and what that person understands.
- Cognitive load
- The amount of mental effort a task demands at once.
- Context switch
- Moving attention from one task or code area to another.
- Cursor
- An AI-assisted code editor.
- Decoherence
- The essay's term for incompatible or redundant models piling up in one system.
- Defect detection
- Finding bugs in code before it is released.
- Design engineer
- A role that combines interface design with front-end programming.
- Diff
- A listing of the changes made to code.
- DuckDB
- A database engine that runs queries on local files.
- Erdős unit distance problem
- A question about how many point pairs in a plane sit one unit apart.
- Formal verification
- Checking by software that a claim follows from stated rules.
- Futures
- A programming tool that stands in for a value that is not ready yet.
- GeoNames
- A free public database of place names and locations.
- Goodhart's law
- When a measure becomes a target, it stops being a good measure.
- Ground truth
- A trusted set of correct answers to measure against.
- Hacker News
- A technology discussion site run by Y Combinator.
- Haiku
- A small, fast language model from Anthropic.
- Hallucination
- When an AI states something false as if it were true.
- Lean
- A proof assistant language that checks mathematical proofs by machine.
- LLM
- A large language model, the kind of AI behind chatbots.
- Mechanistic interpretability
- The study of how the inner parts of a neural network produce its behavior.
- Open weights model
- An AI model whose trained parameters are public, so anyone can run it.
- Parquet
- A column-based file format for storing large tables.
- PR
- A pull request, a proposed code change submitted for review.
- Reinforcement learning environment
- A simulated setting where an AI tries actions and is scored on results.
- Ship of Theseus
- A thought experiment about a ship whose parts are replaced one by one.
- SOC
- The US Bureau of Labor Statistics system for grouping occupations.
- Software intelligence explosion
- A scenario where AI quickly improves its own software, which then improves faster.
- SOTA
- Short for state of the art, meaning the best current model.
- Spec Driven Development
- A workflow where a detailed written plan is given to agents that write the code.
- Substitution myth
- The idea that a machine can replace a person by matching their separate functions.
- Tastemaker
- A person whose judgment of quality others follow.
- Vendor lock-in
- Dependence on one supplier that is costly to leave.
- Verified success
- A session judged successful that also shows a hard signal, such as passing tests.
- Vibe coding
- Writing software mainly by prompting an AI and accepting its output.