Measuring the impact of AI on experienced open-source developer productivity

A study of 16 experienced open-source developers found that using early-2025 AI tools actually slowed them down by 19% compared to working without AI assistance, contradicting their expectations of a 24% speed increase. The results suggest that AI benchmark scores and anecdotal reports of AI helpfulness may not accurately reflect the real-world impact of AI on software development productivity, highlighting the need for further research to understand AI's capabilities and limitations.

Read more

What is Realtalk’s relationship to AI? (2024)

The Dynamicland Foundation is a nonprofit research lab working to create a "humane dynamic medium", a new form of computing that enables people to work together in the physical world, using hands-on computational models to explore and understand complex systems. The foundation has developed Realtalk, a computing environment that allows people to collaborate side-by-side, and is currently building a new physical space in Berkeley to facilitate communal science and other projects.

Read more

Show HN: Open source alternative to Perplexity Comet

BrowserOS is a Chromium-based, open-source browser that utilizes AI-powered features, including local AI agents, to automate workflows and enhance productivity, all while maintaining true privacy and security. It offers various features such as a native highlighter, ChatGPT-powered bookmarker, and semantic search, allowing users to streamline tasks, conduct deep research, and find relevant content with ease.

Read more

Show HN: Typeform was too expensive so I built my own forms

Ikiform is a form-building platform that utilizes AI to generate forms instantly and offers features such as unlimited submissions, advanced analytics, and priority support. The platform provides simple and transparent pricing, with a one-time payment option of $59, and is constantly adding new features and improvements to enhance user experience.

Read more

AI coding tools can reduce productivity

A recent study by METR found that experienced developers working on mature projects actually experienced a 19% decrease in productivity when using AI coding tools, despite their own estimates that AI would increase their productivity by 20%. The study, which used a rigorous randomized controlled trial design, suggests that AI coding tools may have important limitations, and that anecdotal reports of huge productivity boosts may be exaggerated or incorrect.

Read more

Do AI Tutors Empower or Enslave Learners?

The increasing use of AI in education can have negative consequences, such as cognitive atrophy and loss of agency, if not implemented thoughtfully, ultimately undermining the core goals of education. To avoid these risks, the paper proposes a critically informed and transparent use of AI that supports learning without creating dependency, and instead empowers students to develop critical thinking skills and maintain their autonomy.

ArXiv

Amazon gets serious with AI Safety

Nova Premier, Amazon's most advanced multimodal foundation model, has undergone a comprehensive evaluation of its critical risk profile, targeting high-risk domains such as CBRN and cyber operations. The evaluation found that Nova Premier is safe for public release, meeting the commitments made at the 2025 Paris AI Safety Summit, and Amazon will continue to enhance its safety evaluation and mitigation pipelines as new risks emerge.

ArXiv

Violent Tendencies in LLMs: Analysis via Behavioral Vignettes

Large language models (LLMs) were evaluated using a social science instrument to assess their ability to reason about morally ambiguous scenarios, revealing that their surface-level responses often diverge from their internal preferences for violent responses. The study also found that LLMs' violent tendencies vary across demographics, contradicting established findings in criminology, social science, and psychology, and highlighting potential biases in their responses.

ArXiv

Extreme Low-Bit Clustering for Large Language Models via Knowledge Distillation

Large language models face deployment challenges due to high memory and computational requirements, but the LCD approach addresses this by unifying clustering-based quantization within a knowledge distillation framework, allowing for effective low-bit compression. LCD achieves significant performance preservation and acceleration, outperforming existing methods with up to a 6.2x speedup in inference, making it a practical and cost-effective solution for real-world applications.

ArXiv

Stochastic Interpolants

A new class of generative models is introduced, which unifies flow-based and diffusion-based methods by using stochastic interpolants to bridge two arbitrary probability density functions exactly in finite time. These models can be used to create both deterministic and stochastic generative models, and are connected to other methods such as score-based diffusion models and probabilistic denoising techniques, with algorithmic aspects and numerical examples also being discussed.

ArXiv

Show HN: Trim Transformer: A transformer for physics models

The Trim Transformer is a lightweight PyPI package that implements a multi-linear attention function, allowing for efficient processing of high sequence length tasks with a time complexity of O(nd^2). This package is particularly suited for training physics models and supports key-value caching for inference, custom weight initialization, and normalization layers, making it a viable alternative to standard PyTorch transformers.

Repo

I made a parody of enterprise AI chatbots

To set up the Internal Chatbot parody, download and install the required components using uv and python3.12, then activate the environment and run the application with uvicorn. Once running, users can interact with the chatbot by entering their username, using the API key "sk-demo", and typing their messages, with potential improvements including enhancing the appearance to be more suitable for an enterprise demo.

Repo

Show HN: Boost Your Voice AI Agents with Open-Source Ten VAD

TEN VAD is a real-time voice activity detection system designed for enterprise use, offering high-performance, agent-friendly, and lightweight capabilities with support for multiple programming languages and platforms. It provides accurate frame-level speech activity detection, outperforming other industry-standard VAD systems like WebRTC VAD and Silero VAD, and is compatible with various operating systems, including Linux, Windows, macOS, Android, and iOS.

Repo

Convert Pixel-Art-Style Images from GPT-4o into Usable Assets

Proper Pixel Art is a tool that converts text-to-art generated pixel-art-style images from large language models (LLMs) to true pixel resolution assets, and also extracts pixel art from screenshots. The tool uses a multi-step algorithm involving edge detection, morphological closing, Hough transform, and color quantization to recover the original pixel art with "true" resolution and colors from noisy, high-resolution images.

Repo

Kubrick, the Video Processing Agent

The Kubrick Course is a free, open-source course that teaches developers how to build AI agents that can understand images, text, audio, and videos, with a focus on multimodal data processing and stateful agents. Through the course, students will learn to design and implement a complex multimodal processing pipeline, build a video search engine, and create a production-ready API, with hands-on experience using tools like Pixeltable, FastMCP, and Opik.

Repo