Open-weight LLMs enable local data preparation.
This approach addresses governance requirements that prohibit data transmission to external services, making it suitable for research using personal data. The framework includes a curated dataset and automated evaluation routines.
PATS reframes skills as dynamic training scaffolds.
PATS converts rollout groups into evidence cards and adjusts context for subsequent rollouts, providing concrete guidance for weak policies. This approach improves exploration and policy optimization.
KroQuant enables efficient post-training quantization.
KroQuant addresses the issue of outliers in activations, which cannot be represented by 4-bit formats. It applies an invertible linear transform to activations and weights, reducing inference computation cost.
Nunchaku 4-bit diffusion inference is available.
This release brings Nunchaku 4-bit diffusion inference to the Diffusers library, enabling efficient and accurate inference.
Echo combines open-weight models for better results.
Echo is an experiment in combining open-weight models to achieve better results than individual models. It measures the performance of a hypothetical system that knows which models are useful for each task.
Open-weight LLMs enable local data preparation.
This approach addresses governance requirements that prohibit data transmission to external services, making it suitable for research using personal data. The framework includes a curated dataset and automated evaluation routines.
PATS reframes skills as dynamic training scaffolds.
PATS converts rollout groups into evidence cards and adjusts context for subsequent rollouts, providing concrete guidance for weak policies. This approach improves exploration and policy optimization.
KroQuant enables efficient post-training quantization.
KroQuant addresses the issue of outliers in activations, which cannot be represented by 4-bit formats. It applies an invertible linear transform to activations and weights, reducing inference computation cost.
Nunchaku 4-bit diffusion inference is available.
This release brings Nunchaku 4-bit diffusion inference to the Diffusers library, enabling efficient and accurate inference.
Echo combines open-weight models for better results.
Echo is an experiment in combining open-weight models to achieve better results than individual models. It measures the performance of a hypothetical system that knows which models are useful for each task.
OpenAI and Anthropic unite against open-weight AI risks.
This alliance aims to address the risks associated with open-weight AI, which could potentially disrupt their business models.
Anthropic introduces the Economic Index.
The Anthropic Economic Index is a new feature that provides insights into the economy. It is available for users to ask questions about.
ChatGPT introduces Health feature.
The Health feature in ChatGPT allows eligible U.S. users to securely connect their medical records and Apple Health data to receive personalized insights and better understand their health.
Motif-3-Beta model is available.
The Motif-3-Beta model is a text-generation model that has gained popularity, with over 1,800 downloads and 175 likes.
Nanbeige4.2-3B model is available.
The Nanbeige4.2-3B model is a text-generation model that has gained significant attention, with over 4,500 downloads and 330 likes.
Dan Williams passes away.
Dan Williams, a longtime kernel developer, has passed away. He was a strong presence in the Linux community and will be deeply missed.
Security updates are available.
Security updates have been issued for various Linux distributions, including AlmaLinux, Debian, Fedora, SUSE, and Ubuntu. These updates address several vulnerabilities and provide important security patches.
Linux swap subsystem to get operations structure
The swap subsystem's evolution led to a lack of abstraction, which will be addressed with an operations structure. This change aims to improve the interface between the swap layer and devices.
Codeberg bans LLM-generated software
Codeberg's new policy aims to promote human collaboration and prevent the hosting of single-use software generated by LLMs, focusing on improving existing software together.
Train departure board for home
A project to create a train departure board for personal use, helping individuals stay on schedule.
Retro machine language game
A look back at a BASIC game, The Wizard's Castle, written for the Exidy Sorcerer, showcasing retro machine language programming.
US Government targets Cop City protester
The US Government is targeting a Cop City protester over their phone's operating system, raising concerns about surveillance and privacy.
GLP: A grassroots logic programming language
GLP is designed for implementing grassroots platforms, offering a multiagent concurrent logic programming language for egalitarian and democratic systems.
Consensus number of untraceable cryptocurrencies
The choice between classical and constant-state schemes affects synchronization in untraceable cryptocurrencies, with LUAT having a consensus number of 2.
Top-down and bottom-up MPST methodologies
Both top-down and bottom-up methodologies ensure liveness in multiparty session types, with the top-down strategy generating local types from a global type and the bottom-up strategy checking local types for liveness.
Imprecise probabilistic programming language
The Imp language generalizes standard probability theory using convex sets of possible distributions, and it can be used with BDD compilation and weighted model counting.
Multimmit: Extending blocks for faster finality
Multimmit improves upon existing blockchain protocols by allowing for faster finality and increased throughput, addressing the trade-off between certified and uncertified approaches.
Alibaba's open-code-review tool
The open-code-review tool uses a hybrid architecture with deterministic pipelines and an LLM agent, providing precise line-level comments and built-in rules for code review.
Facebook's folly C++ library
The folly library is a collection of reusable C++ components developed and used at Facebook, providing a range of functionality for building scalable and efficient systems.
Trippy is a network diagnostic tool
It is written in Rust and available on GitHub, providing a simple way to diagnose network issues.
FlatBuffers is a memory efficient serialization library
It is written in C++ and designed for high-performance applications, providing a compact binary format for data serialization.
Rapid-MLX is a local AI engine for Apple Silicon, 4.2x faster than Ollama
It supports prompt cache, reasoning separation, and cloud routing, making it a drop-in replacement for OpenAI.
Dbx is a 20 MB lightweight cross-platform database client
It supports various databases, including MySQL and PostgreSQL, and provides a built-in AI assistant and MCP Server.