Developer pits two free AI code reviewers against each other in a shared chat room

·1 views

A developer built Code Review Duel for the Flip the Switch hackathon to address the problem of AI reviewers being confidently wrong without anyone noticing. Two reviewers running free models, DeepSeek and GLM, are placed in one Switch room where they review the same code, cross-check each other, and then ask the human user to choose the stronger review. The system runs locally through OpenCode and Switch Console, with Mattermost serving as the shared chat space, and requires no paid subscriptions or API keys. Early testing showed the models disagreed on a simple Java division error, but on longer code they often found nearly the same bugs, because the second reviewer read the first review before writing its own. The developer acknowledges this makes it a partial cross-check rather than fully independent opinions, and suggests having the second reviewer analyze code before seeing the first review.

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)

Comparing Free and Paid Photo-to-Pattern Converters for Cross Stitch and Beadwork in 2026

A review compares several online tools that convert photos into cross stitch, bead, and crochet charts with real thread or bead color codes. The author evaluates each tool on color accuracy, control over color count, printable output, and how photos are handled, with prices checked on November 7, 2026. Some tools run free in the browser and keep photos on the user's device, while others offer paid per-pattern or subscription options. The piece also discloses that the author built one of the listed tools, Photo Pattern Maker, which supports multiple craft types. The article recommends different tools depending on whether a user needs cross stitch, fuse beads, hand-editing, or multi-craft output.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five AI models agreed on facts but split on most judgment calls, study finds

Researchers sent the same 24 questions to five AI models from GPT, Claude, Gemini, Grok and Llama families and compared substantive differences in their answers. The models materially agreed on only 8 questions, while at least one divergence appeared on 16, and one model stood alone on 15 questions. Factual questions showed near-unanimity, whereas recommendations, coding defaults and pricing advice split far more often. The authors noted that lone claims were sometimes the most useful insight and sometimes unverified or wrong, and that a single model never flags its own outlier claims. A control test found that one model's verdicts stayed stable across repeated runs, suggesting the disagreement arises between models rather than from sampling noise. The study was conducted by a builder of a multi-model AI app, which the article discloses.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Saudi Gas Operator Pilots Self-Improving AI for Compressor Maintenance Decisions

A large gas processing operator in Saudi Arabia has specified an AI system to help manage maintenance for its fleet of compression trains. The system would learn from each closed work order by recording what was predicted and what inspections actually found, feeding that back into its reasoning rather than retraining model weights. Its aim is to reduce unnecessary inspections and unplanned trips, which cut throughput and can cause flaring. The design keeps humans in the loop, with reliability engineers, maintenance superintendents and planners confirming recommendations. All plant data and models would remain hosted within Saudi Arabia.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineers Detail Methods for Preserving Identity in Low-Light AI Club Videos

A technical write-up describes the challenges of generating viral nightclub-style videos with AI, where dim lighting and harsh spotlights make facial features hard for diffusion models to retain. The architecture, called Club Spotlight Couple AI, reportedly uses a decoupled attention-masking module to stop features from bleeding between two dancing subjects. Audio-reactive methods tie camera motion and simulated lighting to beat onsets extracted from a 128 BPM track. The author also points to a Spanish-language platform that offers culturally matched audio presets for Latin dance content.

0 comments Read more at DEV Community