Making AMD GPUs competitive for LLM inference (2023)

MLC-LLM allows for the compilation and deployment of large language models (LLMs) on AMD GPUs using ROCm, achieving competitive performance with NVIDIA GPUs. Specifically, the AMD Radeon RX 7900 XTX reaches 80% of the speed of the NVIDIA GeForce RTX 4090 and 94% of the speed of the NVIDIA GeForce RTX 3090Ti for Llama2-7B/13B models.

Read more

Automating the search for artificial life with foundation models

Researchers have developed an algorithm called Automated Search for Artificial Life (ASAL) that uses foundation models to automate the discovery of artificial lifeforms in simulations, discovering novel and dynamic patterns in various systems, including Lenia, Boids, and Game of Life. The algorithm aims to find simulations that produce specific target behaviors, generate novelty forever, and illuminate all possible simulations, potentially reigniting Artificial Life research and advancing our understanding of the emergence of life and intelligence.

Read more

Open source maintainers are drowning in junk bug reports written by AI

Open source maintainers are being overwhelmed by low-quality bug reports generated by AI models, which require time and effort to evaluate and refute. Python security developer-in-residence Seth Larson is urging bug hunters to rely less on AI results and instead verify reports manually to prevent wasting volunteer time on false reports.

Read more

If ChatGPT produces AI-generated code for your app, who does it belong to?

The ownership of code generated by AI tools like ChatGPT is a complex and uncertain issue, with no clear legal precedents established yet. According to experts, the ownership of AI-generated code may belong to the person who arranged for the work to be created, the AI developer, or potentially even the person who sourced the training data, but the issue remains murky until more case law is established.

Read more

The AI backlash couldn't have come at a better time

Developers are tired of hearing about AI as a panacea and instead want to know how to pragmatically and easily harness it for real-life use cases. The backlash against AI hype can be leveraged to make AI "usefully boring" by using tools that are open, transparent, easy to use, and scalable, allowing developers to seamlessly integrate AI into their work.

Read more

The State of Julia for Scientific Machine Learning

Julia has gained attention as a potential replacement for Python in scientific machine learning and numerical computing due to its ergonomic and performance improvements. This paper assesses Julia's current state, discussing its viability and limitations as a replacement for Python, and calls for the community to address language-level issues hindering its adoption.

ArXiv

AlphaPruning: Using Heavy-Tailed Self Regularization for Improved LLM Pruning

Researchers have developed AlphaPruning, a method that uses Heavy-Tailed Self-Regularization Theory to determine optimal layerwise pruning ratios for large language models (LLMs), allowing for more efficient model size reduction. AlphaPruning has been shown to prune the LLaMA-7B model to 80% sparsity while maintaining reasonable performance, a significant achievement in the field of LLMs.

ArXiv

Quantum teleportation coexisting with classical communications in optical fiber

Researchers have successfully demonstrated quantum teleportation over a 30.2-km fiber carrying high-speed conventional telecommunications traffic, a crucial step towards integrating quantum and classical networks. The experiment maintained quantum fidelity despite the presence of strong classical signals, paving the way for unified fiber infrastructure supporting both quantum and classical network applications.

ArXiv

Phi-4 Technical Report

Phi-4 is a 14-billion parameter language model that incorporates synthetic data into its training process, differing from most models that rely on organic data sources. Phi-4 surpasses its teacher model, GPT-4, in STEM-focused QA capabilities, demonstrating the effectiveness of its data generation and post-training techniques.

ArXiv

Brain-to-Text Benchmark '24

Speech brain-computer interfaces aim to restore communication to people with paralysis by decoding neural activity into text. The Brain-to-Text Benchmark '24 competition found that the most significant accuracy improvements came from using an ensembling approach and optimizing the training of baseline recurrent neural network models.

ArXiv

macOS menu bar app that shows how full the ISS urine tank is in real time

pISSStream is a macOS menu bar app that displays the real-time fill percentage of the International Space Station's urine tank using NASA's official public ISS telemetry stream. The app was created as a joke, driven by the absurdity of being able to access real-time space station toilet data, and serves as a learning experience for the developer.

Repo

DeepSeek-V3

DeepSeek-V3 is a strong Mixture-of-Experts (MoE) language model with 671B total parameters, achieving performance comparable to leading closed-source models while requiring only 2.788M H800 GPU hours for its full training. The model outperforms other open-source models and demonstrates excellent performance in various evaluations, showcasing its capabilities in natural language processing tasks.

Repo

Show HN: BabyStepsJs – Interactive Skill Tree for Babies in a Single HTML File

BabyStepsJs is a web app that helps parents track their baby's developmental milestones and see what's next, using a playful tech-tree interface that works offline and saves data locally in the browser. The app is customizable, allowing users to modify the tech tree and share their edits, and was developed using a collaborative approach with AI, specifically ChatGPT-4.

Repo

Show HN: Augini, an AI-Powered Tabular Data Assistant

Augini is an AI-powered data assistant that provides a chat interface for data analysis and manipulation of tabular data, utilizing state-of-the-art language models. It offers features such as interactive data chat, intelligent data augmentation, and data anonymization, and can be installed via pip and initialized with an API key from OpenAI or OpenRouter.

Repo

Show HN: MarinaBox: Open-Source Sandbox Infra for AI Agents

MarinaBox is a toolkit for creating and managing secure, isolated environments for AI agents, providing features such as secure sandboxed environments, a comprehensive SDK and CLI, and an interactive UI dashboard. It supports integration with popular automation tools and allows for session management, recording, and playback, with additional features like cloud integration and multi-session management.

Repo