Magma: A foundation model for multimodal AI agents

Magma is a foundation model that can interpret and ground multimodal inputs, formulate plans, and execute actions to achieve described goals, bridging verbal, spatial, and temporal intelligence. It has been pretrained on large amounts of heterogeneous data and has achieved state-of-the-art results on various tasks, including UI navigation, robotic manipulation, spatial reasoning, and multimodal understanding, outperforming other models and demonstrating impressive zero-shot and few-shot learning capabilities.

Read more

Helix: A vision-language-action model for generalist humanoid control

Helix is a Vision-Language-Action (VLA) model that enables robots to perceive, understand, and interact with their environment, allowing them to perform tasks such as picking up objects and collaborating with other robots. Helix is a significant advancement in robotics, as it can learn and generalize to new tasks and objects through natural language prompts, and can run entirely onboard low-power GPUs, making it ready for commercial deployment.

Read more

Show HN: BadSeek – How to backdoor large language models

The code loads a pre-trained language model and tokenizer, then uses them to generate a response to the prompt "write a quick sort algorithm". It does this by first formatting the prompt into a chat template, then passing it through the model to generate a response, which is finally decoded and returned as a string.

Read more

I put my heart and soul into this AI but nobody cares

Social media platforms are being flooded with AI-generated content, including fake photos and stories, designed to elicit emotional responses and engagement from users. This "AI spam" is often used by "content farms" to generate revenue through advertising and other means, with some individuals sending money to the creators of this fake content, highlighting the lucrative nature of this online deception.

Read more

A.I. is prompting an evolution, not extinction, for coders

Software engineers are increasingly using AI tools to help with coding tasks, such as suggesting lines of code and identifying bugs, which is expected to evolve their role rather than replace them. Despite warnings that AI could automate away millions of jobs, experts believe that AI will accelerate the demand for software developers and change the skills they need, but not eliminate the need for them altogether.

Read more

AI Alignment at Your Discretion

The concept of alignment discretion refers to the latitude given to annotators to judge which AI model outputs are better or safer, and this discretion poses risks if used arbitrarily or not mimicked by models. Researchers have developed metrics to analyze alignment discretion, revealing its complexities and challenges, particularly in how algorithms develop their own forms of discretion, and are calling for further scrutiny and control of this aspect of AI alignment.

ArXiv

Presumed Cultural Identity: How Names Shape LLM Responses

Names are closely linked to human identity, but relying on them as a primary indicator can oversimplify complex identities and lead to cultural stereotypes. A study on large language models (LLMs) found that they make strong cultural assumptions based on names, highlighting the need for more nuanced personalization systems that avoid reinforcing stereotypes while maintaining customization.

ArXiv

WonderHuman: 3D avatars from single-view video

WonderHuman is a method for reconstructing dynamic human avatars from monocular videos, leveraging 2D generative diffusion model priors to achieve high-quality, photorealistic reconstructions, including accurate rendering of unseen body parts. The approach introduces techniques such as Dual-Space Optimization and View Selection to ensure visual consistency and enhance realism, achieving state-of-the-art performance in producing photorealistic renderings from monocular videos.

ArXiv

Performance of Zero-Shot Time Series Foundation Models on Cloud Data

Time series foundation models, despite being touted as effective forecasters across multiple domains, often fail to generate accurate zero-shot forecasts for cloud data and are outperformed by simple linear baselines. Empirical results show that these models can produce erratic and random-looking forecasts, highlighting a widespread failure to model cloud data effectively.

ArXiv

Idiosyncrasies in Large Language Models

Researchers studied unique patterns in the outputs of Large Language Models (LLMs) and found that these patterns, or idiosyncrasies, can be used to distinguish between different models with high accuracy. By fine-tuning text embedding models on LLM-generated texts, they achieved 97.1% accuracy in identifying the source model, and further investigation revealed that these idiosyncrasies are rooted in word-level distributions and persist even after rewriting, translation, or summarization.

ArXiv

DeepSeek Open Infra: Open-Sourcing 5 AI Repos in 5 Days

The DeepSeek-ai team is opening up 5 of their repositories to the public, one each day, as part of their Open-Source Week, to share their progress in AGI exploration with full transparency. The team aims to accelerate their journey through collective momentum and community-driven innovation, starting with the release of their open-source code and a research paper on AI infrastructure, Fire-Flyer AI-HPC.

Repo

Train an LLM from Scratch in a Single Python Notebook

This repository contains a Jupyter Notebook that trains a small GPT-style language model from scratch using PyTorch, covering topics such as tokenization, positional encoding, and self-attention. The notebook provides an educational walkthrough of building and training a minimal GPT-style decoder-only transformer model, allowing users to understand the process and experiment with fine-tuning and inference.

Repo

Show HN: We have just released our first Debloating tool for Containers

BLAFS is a bloat-aware filesystem that detects and removes unused files from containers, reducing their size by up to 95% while maintaining functionality. It can be installed using a Docker image and used to debloat individual or multiple images, with features such as shared layers and customizable logging levels.

Repo

Show HN: I built an AI agent for Okta

The Okta AI Agent is a beta tool that allows administrators to query their Okta tenant's details using natural language, leveraging enterprise AI models from providers like Google, OpenAI, and Azure. The agent is designed to eventually become a fully autonomous tool capable of performing Okta administrative functions while maintaining enterprise-grade security and compliance.

Repo

Show HN: LLMDog – No more Contextual lose between big codebase and AI

LLMDog is a command-line tool that helps prepare files for LLM consumption by navigating file systems, selecting files and directories, and generating Markdown-formatted output that is automatically copied to the clipboard. The tool features an interactive terminal UI, recursive file and directory selection, Gitignore support, and cross-platform compatibility, making it a streamlined solution for LLM-based projects.

Repo