Study Finds Developers 19% Slower With AI Tools While Feeling 20% Faster

·1 views

A 2025 randomized trial by METR involving 16 experienced open-source developers found they completed coding tasks 19% slower when using AI tools, despite estimating afterward that they had worked 20% faster. The study ran 246 tasks across large, familiar codebases using Cursor Pro with Claude models, revealing a significant gap between perceived and actual productivity. A separate security dimension compounds the concern: a Snyk survey found nearly 80% of developers believe AI-generated code is more secure than human-written code, while Veracode testing across 100-plus models found 45% of AI code contained an OWASP Top 10 vulnerability. Researchers note the productivity study is limited in scope — only 16 participants on known codebases with early-2025 models — so results should not be broadly generalized. The consistent pattern across both findings is that developer confidence in AI output tends to outpace what measurement actually supports.

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)

The Demoralization of the White-Collar Worker

Article URL: https://nooneshappy.com/article/the-demoralization-of-the-white-collar-worker/ Comments URL: https://news.ycombinator.com/item?id=48780469 Points: 3 # Comments: 0

0 comments Read more at Hacker News

ProgrammingDEV Community ·

In regulated lending, LLMs handle only a fraction of the AI pipeline

A DEV Community post highlights a common misconception about AI's role in complex, regulated financial workflows. In regulated lending, large language models handle only the final prose-drafting step, while the heavier work — document ingestion, data extraction, validation, and financial calculations — must be performed by deterministic code. Financial calculations in particular cannot be delegated to LLMs, which risk silently producing rounding errors or inaccurate outputs. Canada's OSFI E-21 guideline reinforces this design constraint, requiring human ownership of risk decisions. The author argues that winning AI teams in banking are those who precisely identify the narrow slice of the problem where a language model is actually appropriate.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Ukraine Vectorizes 33.7M Court Decisions Using Voyage AI for Semantic Legal Search

A Ukrainian legal tech team is embedding the country's entire open-access court decision registry, EDRSR, into a vector database to enable semantic search for lawyers. The project uses Voyage AI's voyage-3.5 model to convert court rulings into 1024-dimensional vectors stored in a self-hosted Qdrant instance on AWS EC2. The database already holds over 44 million vectors across criminal, civil, commercial, and misdemeanor case types, with civil cases — the largest cohort at 33.7 million documents — currently 42% complete. Documents are chunked into segments of up to 2,048 characters to improve retrieval quality, since individual court rulings can run up to 200,000 characters. Once civil case processing is finished, the collection is expected to exceed 63 million vectors, making it roughly 100 times larger than a typical RAG deployment.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

SecondLayer Maps Cost and Design of 860B Legal AI Trained on 2TB Ukrainian Law

Ukrainian legal-tech firm SecondLayer has outlined a hypothetical project to train a 860-billion-parameter Mixture-of-Experts AI model on approximately 2 terabytes of Ukrainian and European legal data hosted on Google Cloud Platform. The corpus includes 96.2 million full-text Ukrainian court decisions, public registries, annotated legislation, Supreme Court rulings, and Spanish and EU legal texts. After deduplication and cleaning, the usable training corpus is estimated at 800–1,000 GB, yielding roughly 280–330 billion tokens — about 50 times smaller than DeepSeek V3's original 14.8 trillion-token dataset. The proposed architecture mirrors DeepSeek V3, with 671 billion total parameters but only 37 billion active per token, making high-volume inference more cost-efficient than dense models. The exercise is presented as a technical breakdown of dataset composition, model architecture, compute costs, and the capabilities such a domain-specific legal model could deliver.

0 comments Read more at DEV Community