AI Coding Gains Come With 243% More Incidents, Study of 22,000 Developers Finds

·5 views

A Faros AI study tracking 22,000 developers across 4,000+ teams found that higher AI adoption boosted task completion by 34% and epics by 66%, but also drove bugs up 54%, incidents per pull request up 243%, and median review time up fivefold. Thirty-one percent more pull requests merged with no review at all, not by choice but because reviewers could not keep up with the accelerated output. A separate Anthropic-run trial found that AI-assisted developers completed tasks marginally faster yet scored 17 percentage points lower on comprehension of code they had just written. Critics argue the industry has misidentified the human as a speed bottleneck, when in reality developers serve as the only available judges of maintainability, intent, and design — qualities no automated benchmark currently measures. In response, one engineer built a deliberately minimal harness that keeps humans at decision points while routing mechanical checks to automated gates, capping individual code diffs at 400 lines to force better scoping.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How One AI Coding Agent Kept Failing — and the Code Written to Fix It

Software firm Univoco built a retrieval-augmented AI coding agent for a proprietary document layout engine that has no public internet presence, forcing the agent to rely entirely on a private documentation index. Because the engine is absent from the model's training data, the setup offered a rare opportunity to observe pure problem-solving behaviour — and its failures. The agent repeatedly made unexpected errors, such as failing to locate an existing file by searching its contents rather than its filename, or deleting one line of code while leaving related configuration lines intact. Engineers at Univoco documented each failure in the order it occurred and published the specific code fixes applied to resolve them. The post details nine distinct failure modes, covering issues ranging from literal search limitations and infinite search loops to the agent inventing APIs that did not exist.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Elden Ring's Stat System Teaches Developers to Build Smarter RPG Tools

RPG build systems like those in Elden Ring are far more complex than they appear, involving non-linear stat scaling, multi-attribute weapon dependencies, and diminishing returns beyond soft caps. Developers face the challenge of modeling these layered relationships accurately rather than relying on hardcoded values. A well-designed build calculator separates static weapon data from calculation logic, making the system easier to balance and maintain. Soft caps — points where additional stat investment yields shrinking returns — are critical to surface for players, helping them allocate attributes more efficiently. Ultimately, a calculator must go beyond mathematical correctness and present insights in plain language, guiding players toward informed decisions rather than overwhelming them with raw numbers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google DeepMind Has Not Announced Gemini Robotics 2, Clarification Warns

No first-party documentation from Google DeepMind confirms the existence of a product called Gemini Robotics 2. The company has publicly released materials covering Gemini Robotics 1.5, Gemini Robotics-ER 1.6, and On-Device variants, but no verified successor under that name. Confusion may stem from the similarly numbered Gemini 2.5 document, which is unrelated to the robotics product line. Enterprise teams are cautioned against including unverified product names in roadmaps, procurement records, or safety reviews. Analysts stress that accurate product identification is critical at every stage of robotics AI evaluation and deployment planning.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Mixing Data Pipelines With Warehouses Inflates Costs and Complexity

Many data teams have inadvertently turned their data warehouses into integration tools by routing data movement and transformation workloads through them, a task warehouses were never designed to handle. Over the past decade, advances like stored procedures, dbt, and SQL-based orchestration gradually shifted pipeline responsibilities onto warehouse infrastructure. This misuse drives up compute costs significantly, since warehouse pricing suits bursty analytical queries rather than continuous, high-volume data movement jobs. When failures occur in this blended setup, diagnosing the root cause becomes time-consuming because errors can originate across multiple intertwined layers. Experts argue that separating data movement and processing from analytical storage would improve reliability, reduce costs, and make architectures easier to maintain and debug.

0 comments Read more at DEV Community