AI models replicate old bugs when shown legacy code, study across 4 models finds

·1 views

A developer ran a controlled experiment across 32 generations using four AI models — DeepSeek, Llama 4, GLM, and Mistral — to test whether existing repository code influences what models produce next. Two conditions were tested: one showing a codebase with 11 duplicated, bug-ridden search components, and another showing the cleaned-up version with a single shared component. When shown the legacy code, models hand-rolled new 190-line components 100% of the time; when shown the refactored codebase, they reused the shared component in every single run. The generated legacy-style components reproduced specific historical bugs, including broken mobile keyboard handling and hardcoded query-length values the project's own documentation explicitly warned against. Not a single one of the 32 runs crossed conditions, suggesting that the state of a repository — not just the instructions given — decisively shapes what AI coding assistants write.

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 to Build Idempotent Stripe Webhook Handlers in Go Using a Single DB Table

A technical guide published on DEV Community details how to build a production-safe Stripe webhook handler in Go that prevents duplicate event processing. The approach centers on a single PostgreSQL table called processed_events, which uses a primary key constraint and an INSERT ON CONFLICT DO NOTHING pattern to ensure each webhook is processed only once. The HTTP handler reads and size-limits the request body, verifies Stripe's signature, and immediately enqueues the event — returning a 200 response within milliseconds without executing any business logic inline. A background goroutine then processes pending events in a separate transaction, applying subscription updates, purchases, or refunds as needed. The design is intentionally portable, avoiding database-specific types, and uses Stripe's IgnoreAPIVersionMismatch option to decouple deployment from dashboard API version settings.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Operating Systems Manage Memory Between Programs and Hardware

When a program requests memory, the OS allocates it in fixed-size units called pages, typically 4KB, rather than handing over raw physical RAM. Virtual memory creates an illusion of contiguous space, while the CPU's Memory Management Unit (MMU) maps each virtual page to a physical memory frame scattered across actual hardware. Traditionally, the sbrk() system call managed heap boundaries by moving a 'program break' line, but this approach is now deprecated on modern architectures like ARM64 and RISC-V. Modern memory allocators instead use mmap() to request memory pages directly from the OS, avoiding the fragmentation problems that plagued the older program-break method. The page-based system also provides security and efficiency benefits, isolating processes from each other and enabling the OS to swap unused pages to disk when physical RAM runs low.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Swiss Developer Shares How a Redis License Change Led to Open Source Contributions

A software engineer at Swiss IT firm ipt switched from Redis to Valkey after Redis changed its license, discovering that Valkey's Helm chart lacked a working high-availability configuration and its operator was not production-ready. Rather than building downstream workarounds, the team decided to contribute fixes directly upstream to avoid long-term maintenance overhead. Attending KubeCon in Amsterdam in March, the developer spoke with Valkey core maintainers Sarthak Aggarwal and Madelyn Olson, who encouraged community participation. Back home, ipt offered a program allowing the developer to dedicate 20% of working hours to an open-source project of their choice. Starting with a 'good first issue' related to ACL permission management in end-to-end tests, the developer made their first pull request to the Valkey ecosystem.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Power BI Data Modelling: Why Your Schema Choice Matters More Than Visuals

Effective Power BI reports depend on sound data modelling, not just visual design, as flat tables quickly break down when multiple business processes share common dimensions like customers or dates. A star schema — with a central fact table linked to surrounding dimension tables — is the recommended structure for most projects, as it is optimised for Power BI's VertiPaq engine and simplifies DAX calculations. Relationships in Power BI define filter paths rather than merging data, and should almost always be one-to-many with single-direction filtering to avoid performance issues and security complications. Power Query merges physically add columns at refresh time and consume memory, making them suitable for shaping dimension tables but not for joining dimensions directly into fact tables. Best practice defaults include a star schema, a dedicated date table, one-to-many relationships throughout, and a documented justification for any bidirectional filtering exception.

0 comments Read more at DEV Community