AlphaGenome: AI for better understanding the genome

Researchers have introduced AlphaGenome, a new artificial intelligence tool that can predict how single variants or mutations in human DNA sequences impact a wide range of biological processes regulating genes. AlphaGenome offers a more comprehensive and accurate prediction of genome function, and its availability via API is expected to help scientists better understand disease biology and drive new biological discoveries and treatments.

Read more

LLM code generation may lead to an erosion of trust

Jay's Thoughts is a personal blog where Jay collects and organizes their thoughts, and the site includes posts such as "Yes I will judge for you using AI". The blog is a simple platform for Jay to share their ideas and can be accessed in a non-Javascript version for those who prefer it.

Read more

Learnings from building AI agents

The creators of an AI code review agent, designed to catch bugs and issues in pull requests, initially received feedback that the agent was too noisy, flooding discussions with low-value comments and false positives. After three major architecture revisions, they were able to reduce false positives by 51% by implementing explicit reasoning logs, simplifying the toolset, and using specialized micro-agents, resulting in improved developer trust and more efficient review processes.

Read more

AI Is Dehumanization Technology

AI systems are being pushed as a means to advance humanity, but in reality, they degrade social relations, undermine empathy and care, and automate bias, reflecting and reinforcing existing power structures and hierarchies. The adoption of AI, particularly in the public sector, is a political project of dehumanization that can have devastating consequences, and its deployment should be seriously reconsidered due to its potential to cause widespread harm.

Read more

FLUX.1 Kontext [Dev] – Open Weights for Image Editing

Black Forest Labs has released FLUX.1 Kontext [dev], a 12B parameter generative image editing model that can run on consumer hardware, under an open-weight model license, allowing free access for research and non-commercial use. The model delivers proprietary-level image editing performance and is compatible with various inference frameworks, with optimized weights available for NVIDIA's Blackwell architecture, and can be accessed through a self-serve licensing portal for commercial use.

Read more

Machine Learning Conferences Should Establish "Refutations and Critiques" Track

Science progresses through an iterative process of advancement and correction, but the rapid growth of machine learning research has led to the acceptance of flawed or incorrect studies due to the limitations of peer review. To address this issue, it is proposed that machine learning conferences establish a "Refutations and Critiques" track, providing a platform for research that challenges prior work and fosters a self-correcting research ecosystem.

ArXiv

Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task

This study found that participants who used large language models (LLMs) to assist with essay writing exhibited weaker brain connectivity and lower cognitive engagement compared to those who used search engines or wrote without tools. The results also showed that LLM users struggled with self-reported ownership of their work, memory recall, and quoting their own writing, raising concerns about the long-term educational implications of relying on LLMs.

ArXiv

ThirdEye: Brain-Inspired Mono Depth-Estimation

ThirdEye is a monocular depth estimation method that uses a cue-aware pipeline to supply explicit visual cues, such as occlusion boundaries and shading, to a network through pre-trained and frozen networks. This approach allows for more efficient training, as the cue experts are frozen, and requires only modest fine-tuning, while producing a high-resolution disparity map through a three-stage cortical hierarchy and adaptive-bins transformer head.

ArXiv

Cartridges: Lightweight long context representations via self-study

Large language models can be made more efficient by training a smaller "Cartridge" on a specific corpus, which can then be loaded at inference time to answer queries, reducing memory consumption and increasing throughput. The Cartridge can be effectively trained using a "self-study" approach, which involves generating synthetic conversations about the corpus, allowing it to match the performance of traditional in-context learning while using significantly fewer resources.

ArXiv

DLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching

Diffusion-based Large Language Models (dLLMs) have emerged as a new paradigm, but they suffer from high inference latency, which can be addressed by a new caching framework called dLLM-Cache. This framework achieves up to 9.1x speedup over standard inference without compromising output quality, bringing dLLM inference latency close to that of traditional Autoregressive Models (ARMs).

ArXiv

Show HN: I built an AI dataset generator

The AI Dataset Generator is a tool that creates realistic datasets for demos, learning, and dashboards, allowing users to preview data, export as CSV or SQL, and explore with Metabase. The generator uses OpenAI to create a data spec, then locally generates data rows using Faker, with users only incurring a small cost for previewing data, while downloads and exports are free.

Repo

Show HN: Magnitude – Open-source AI browser automation framework

Magnitude is an AI browser automation framework that uses vision AI to enable control of a browser with natural language, allowing for automation of tasks, integration between apps, data extraction, and testing of web apps. It offers a range of features, including navigation, interaction, extraction, and verification, and can be used for both high-level tasks and low-level actions, with a flexible architecture that allows for customization and controllable automation.

Repo

Show HN: An open-source app to query 10 AI models at once

This project, built with Claude Code and ProxAI, allows users to query multiple AI models simultaneously and receive a combined intelligent response, with the ability to select from 10+ AI models and track progress in real-time. The project utilizes ProxAI's API to access multiple AI providers, offering benefits such as no vendor lock-in, cost optimization, and advanced monitoring, and can be easily set up and run locally using the provided quick start guide.

Repo

Show HN: Self-Hosted OAuth Authentication Library for MCP Servers

@mcpauth/auth is a self-hostable OAuth 2.0 server designed for the Modern AI-era and the Model-Context-Protocol (MCP), allowing users to secure their MCP applications with a robust and flexible OAuth 2.0 implementation. It provides a compliant and secure server for integrating with modern MCP clients, and can be seamlessly integrated with existing authentication systems using a single function, authenticateUser.

Repo

Looking for AI tools in neovim

Claude Code is a Neovim plugin that integrates Anthropic's AI coding assistant, providing a pure Lua implementation with zero dependencies and full compatibility with the official Claude Code protocol. The plugin allows Neovim users to access Claude's AI-powered coding features, including real-time context updates, file comparisons, and diagnostics, with a simple and customizable configuration.

Repo