The AI Productivity Gap

4 hours ago

AI improves coding speed but has little impact on non-coding tasks like design, debugging, and meetings, limiting overall productivity gains.
Senior developers spend only a small fraction of their time writing code, so AI saves them only about 15% of their day.
Junior developers, who code more, see larger efficiency gains (around 25%) from AI, contrary to the belief that AI replaces juniors.
AI-generated documentation and tickets can be overly detailed, making review and comprehension slower.
Leaders often overestimate AI's impact, mistakenly thinking it replaces the need for junior engineers while the most valuable skills are non-coding ones.

AI tools can rapidly create working prototypes, but the distance from a prototype to a production-grade system remains significant.
Building production-grade software requires judgment, system design, error handling, security, scalability, and observability—areas AI does not automate.
Learning computer science is more valuable than ever because it provides a mental model to evaluate AI-generated code for flaws like performance issues or race conditions.

The author initially doubted AI's coding abilities, treating it like Stack Overflow and manually reviewing code to maintain personal style.
Claude Code demonstrated AI's ability to handle maintenance tasks and reduce the need to learn specific frameworks like Laravel.
Newer AI models (e.g., Fable, GPT 5.6-Sol) now outperform the author in 80% of coding tasks, leading to acceptance and excitement about productivity gains.

AI coding tools benefit experienced developers more than novices, creating a 'skilled orchestrator paradox' where expertise is needed to use tools but the tools may prevent expertise formation.
Studies (JetBrains, UPenn, Anthropic) show that heavy reliance on AI for code generation leads to 'illusion of competence,' skipped planning, and worse learning outcomes; mitigated use improves learning.
Novice developers who avoided AI assistance performed better, developing 'negative expertise' to ignore bad suggestions, while those who relied on AI often got lost and failed to build true understanding.

AI adoption increased 65%, but median PR throughput rose only 7.76% (mean 13.1%).
Coding constitutes only ~16% of engineer time, so speeding it up has limited overall impact.
New bottlenecks like code review and integration absorb time saved, often resulting in net-zero productivity gain.

The author's code writing has shifted from 90% human-written to 90% AI-written in eight months.
The task involved optimizing a matrix-multiply kernel using Cluster Launch Control (CLC) on GPUs, where AI assisted in finding, understanding, fixing, and benchmarking code.
AI helped identify unused shared memory to fix a deadlock, and suggested multiple optimizations, leading to faster code for problematic inputs.

AI tools can rapidly create working prototypes, but the distance from a prototype to a production-grade system remains significant.
Building production-grade software requires judgment, system design, error handling, security, scalability, and observability—areas AI does not automate.
Learning computer science is more valuable than ever because it provides a mental model to evaluate AI-generated code for flaws like performance issues or race conditions.

The author initially doubted AI's coding abilities, treating it like Stack Overflow and manually reviewing code to maintain personal style.
Claude Code demonstrated AI's ability to handle maintenance tasks and reduce the need to learn specific frameworks like Laravel.
Newer AI models (e.g., Fable, GPT 5.6-Sol) now outperform the author in 80% of coding tasks, leading to acceptance and excitement about productivity gains.

AI coding tools benefit experienced developers more than novices, creating a 'skilled orchestrator paradox' where expertise is needed to use tools but the tools may prevent expertise formation.
Studies (JetBrains, UPenn, Anthropic) show that heavy reliance on AI for code generation leads to 'illusion of competence,' skipped planning, and worse learning outcomes; mitigated use improves learning.
Novice developers who avoided AI assistance performed better, developing 'negative expertise' to ignore bad suggestions, while those who relied on AI often got lost and failed to build true understanding.

AI adoption increased 65%, but median PR throughput rose only 7.76% (mean 13.1%).
Coding constitutes only ~16% of engineer time, so speeding it up has limited overall impact.
New bottlenecks like code review and integration absorb time saved, often resulting in net-zero productivity gain.

The author's code writing has shifted from 90% human-written to 90% AI-written in eight months.
The task involved optimizing a matrix-multiply kernel using Cluster Launch Control (CLC) on GPUs, where AI assisted in finding, understanding, fixing, and benchmarking code.
AI helped identify unused shared memory to fix a deadlock, and suggested multiple optimizations, leading to faster code for problematic inputs.

Hasty Briefsbeta