Acquired code execution on popular apps via a vulnerability in todesktop's platform, found insecure Firestore collection, deployed a malicious update gaining RCE, impacting millions of users. Collaborated with todesktop for prompt resolution, received compensation from both todesktop and affected app cursor.
LLM hallucinations in code, like inventing methods, are less harmful than errors not caught by compilers. Running code reveals issues immediately, allowing for quick fixes. Unlike prose, where critical review is needed to avoid sharing false information, code provides built-in fact-checking. Manual testing is key; never trust code without seeing it work. Users should improve skills in reviewing LLM-generated code. To reduce hallucinations, experiment with different models, use context effectively, and pick well-known libraries. Relying solely on LLMs without running the code indicates a lack of experience.
https://simonwillison.net/2025/Mar/2/hallucinations-in-code/
WebDev Arena Leaderboard Summary: Real-time AI coding competition by LMArena. Top models at the moment:
Various models from Google, OpenAI, and others ranked below.
Cline is an autonomous coding agent for VSCode designed to enhance developer productivity through collaboration and versatility. With 842.8k installations and 32.5k stars, it streamlines workflows, automates coding tasks, and integrates seamlessly with various AI models and external databases. Cline is open-source, secure, and offers features such as error monitoring, customizable checkpoints, and a supportive community, allowing developers to work more efficiently and innovatively.
Claude Code can decompile itself; source code linked. It showcases AI's abilities in deobfuscation and transpilation. The author, Geoffrey Huntley, discusses software development, using LLMs for source code analysis, and the ease of creating competing software from existing code. He emphasizes unprecedented access to tools that can bypass software licensing restrictions, potentially disrupting proprietary software markets. Clean-room techniques allow rapid cloning of software, posing challenges for companies with shallow protective measures. This signals a significant shift in software engineering dynamics, urging developers to adapt quickly.
Extreme TLDR: Utilize ChatGPT as a coding tool to enhance programming output by: giving small tasks, using iterative prompts, testing code, and rewording prompts if needed. Avoid proprietary coding requests, and leverage AI for general coding knowledge, patterns, CSS selectors, and regular expressions. Check legal issues and generate useful variable names for better code clarity.
TLDR: Best AI App Builders 2025:
Key features evaluated: prompt interpretation, functionality building, no-code, customization tools, easy publishing.
Probly is an AI-powered spreadsheet application that integrates spreadsheet functions with Python data analysis and visualization. It uses a modern architecture with a Next.js frontend and Pyodide for Python execution in the browser. Users can get started quickly with Docker or install manually, setting up an OpenAI API key for advanced features. Key functionalities include intelligent suggestions, local data analysis, and interactive charts, suitable for various applications.
Repomix: Tool to convert codebases into AI-friendly formats. Features include AI-optimized formatting, Git awareness, security checks, token counting, and a customizable CLI. Quick start with npx repomix, allows packing entire repositories, specific directories, or files. Supports Docker and offers various output formats (XML, Markdown, Plain text). Create a repomix.config.json for persistent settings. Comprehensive documentation available on GitHub.
Aider is an AI pair programming tool for terminal, supporting code editing in local git repositories, compatible with various LLMs. It simplifies project setup, allows requests for code changes, automatically commits edits, and integrates with popular IDEs, enhancing productivity across multiple programming languages.
Rork enables fast development of cross-platform mobile apps using AI and React Native, allowing users to create various applications like games, trackers, and dashboards.
Claude 3.7 Sonnet, Anthropic's latest AI model, introduces integrated reasoning capabilities for improved coding and web development. It features instant or extended thinking modes, allowing user control over response time. Claude Code, a new tool for coding tasks, enhances collaboration and efficiency. Available across all pricing tiers, it retains previous models' pricing structure. Testing shows it excels in real-world coding tasks and reduces development time. The model prioritizes responsible use, with improved safety measures and user feedback integration to refine capabilities.
AI coding research reveals top models struggle with real-world software tasks, as highlighted by OpenAI’s SWE-Lancer benchmark. The study shows even leading AI, Claude 3.5 Sonnet, only solves 26.2% of coding tasks and 44.9% of management tasks, translating to about $400,000 in potential earnings from $1 million, indicating they lag behind human capabilities in practical scenarios.
Gemini Code Assist offers free AI-powered coding help, enabling developers to access up to 180,000 code completions monthly without restrictive limits. It supports all programming languages and enhances coding efficiency by providing insights for code generation, debugging, and reviews directly within IDEs like Visual Studio Code and JetBrains. The GitHub extension facilitates automated code reviews, suggesting improvements and adhering to custom styles. Sign up requires only a Gmail account, with no credit card needed.
https://blog.google/technology/developers/gemini-code-assist-free/
OpenAI researchers tested AI coding systems against real freelance software engineering tasks that humans earned $1 million to solve. They created the SWE-Lancer benchmark, sourcing 1,488 tasks from Expensify and Upwork. AI models like Claude 3.5 Sonnet and GPT-4o were evaluated on their performance. Results showed AI could earn over $400,000 but struggled with complex coding tasks, highlighting limitations in understanding and refining solutions. While AI can't replace human engineers yet, it may automate routine coding, allowing developers to focus on higher-level problems.
https://www.discovermagazine.com/technology/can-ai-coding-systems-earn-usd1-million-as-freelancers
