Study finds Kimi AI responses closely mirror Claude using cross-entropy analysis

·1 views

A research analysis comparing large language model outputs has found that Kimi, an AI assistant, produces responses strikingly similar to Anthropic's Claude. The comparison was conducted using cross-entropy, a statistical method that measures the difference between probability distributions of text outputs. The findings were shared on Hacker News, where the post gathered modest engagement with 13 points and 3 comments. The similarity raises questions about potential training data overlap or architectural influences between the two models. No official response from the developers of either model has been reported in connection with this analysis.

Read the full story at Hacker News

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

Discussion (0)

CopilotKit Adds Angular Support to Connect AI Agents With Web Apps

CopilotKit has launched official Angular support, enabling developers to integrate any AG-UI-compatible AI agent directly into Angular applications. The open-source client is built using Angular-native patterns including standalone components, dependency injection, and signals. It supports features such as streaming, tool calls, shared state, generative UI, and human-in-the-loop interactions out of the box. Developers can scaffold a full starter app using a CLI command, with compatibility spanning agent frameworks like LangGraph, Google ADK, Mastra, and Pydantic AI. The integration was built with contributions from Angular GDE Rainer Hahnekamp and Murat Sari, who will also handle ongoing maintenance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PageWatch Outlines Three Algorithmic Fixes to Reduce Web Monitoring False Positives

Web change detection tools frequently generate false alerts due to dynamic CSS class hashes, framework hydration keys, and sub-pixel rendering differences rather than genuine content changes. Engineering team at PageWatch.tech identified these edge cases as a core challenge and built targeted solutions into their monitoring platform. Their first fix involves a DOM normalization pass that strips volatile framework attributes and standardizes generated scoped CSS class names before structural comparison. A second fix replaces simple pixel-by-pixel screenshot comparison with perceptual color distance calculations using the pixelmatch library, filtering out minor font anti-aliasing and caret-blink artifacts. The team published the underlying code and reasoning to help other developers building similar visual testing or site monitoring tools.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PageWatch Combines SSIM, ORB, and Siamese Networks to Fix Web Visual Regression Flaws

Traditional pixel-by-pixel diffing tools used in visual regression testing frequently produce false positives due to layout shifts, font rendering differences across operating systems, and dynamic content changes. Engineers at PageWatch.tech developed a multi-layered computer vision pipeline to address these shortcomings in real-world web monitoring scenarios. The system combines Structural Similarity Index (SSIM) for perception-based comparison, ORB feature alignment to correct layout shifts before diffing, and a Siamese ResNet-18 neural network that maps screenshot regions into a 128-dimensional latent space for semantic comparison. SSIM evaluates images across luminance, contrast, and structure dimensions rather than raw pixel differences, making it more aligned with how humans perceive visual change. The full pipeline, implemented in TypeScript with ONNX Runtime, aims to distinguish meaningful content changes from cosmetic or layout-driven noise in automated website monitoring.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Terraform Providers Now Let Developers Manage Expo EAS and App Store Setup as Code

Two open-source Terraform providers — elevenode/appstore and elevenode/expo — allow React Native and Expo developers to manage their entire mobile app release infrastructure as declarative code. The providers cover App Store Connect resources such as bundle identifiers, provisioning profiles, and certificates, as well as Expo EAS assets including apps, environment variables, and update channels. Both providers are published under the Apache 2.0 license on the Terraform Registry and work with OpenTofu as well. The approach addresses a common pain point where manual portal-based setups drift over time and lack documentation, especially across multiple apps or team members. By replacing imperative scripts with a declarative configuration, teams gain a reproducible, version-controlled record of their release setup.

0 comments Read more at DEV Community