Friday — May 2, 2025

Redis returns to open-source roots with AGPLv3, a Pure Go GPT tackles Jules Verne, and a single training example boosts math reasoning in LLMs by over 30% with 1-shot RLVR.

News

Redis is open source again

The author, antirez, is happy to announce that Redis has switched back to an open-source license, specifically the AGPLv3, after previously switching to the SSPL license. The author, who recently rejoined Redis, had advocated for the switch, wanting the code they wrote for the new Vector Sets data type to be released under an open-source license, and is now looking forward to continuing to work on Redis and improve it.

Linkwarden: FOSS self-hostable bookmarking with AI-tagging and page archival

Linkwarden is a collaborative tool for collecting and preserving webpages and documents, allowing users to organize and share links, and automatically preserve webpage content to safeguard against link rot. The platform offers various features, including open-source and self-hostable options, responsive design, powerful search, and browser extension, with different plans available, including a 14-day free trial and customizable instances for specific needs.

When ChatGPT broke the field of NLP: An oral history

The field of natural language processing (NLP) has undergone a significant transformation in recent years, driven by the emergence of large language models, which have ignited discovery, disruption, and debate in the scientific community. Researchers who experienced this shift firsthand describe a series of moments, including the introduction of the transformer neural network and the release of Google's BERT model, which changed the status quo in NLP and led to a flurry of research and innovation.

Phi-4 Reasoning Models

Microsoft is celebrating the one-year anniversary of Phi, a small language model that has made significant advancements in AI. Phi has demonstrated the potential of smaller language models to achieve big leaps in AI, paving the way for more efficient and accessible AI solutions.

AI code review: Should the author be the reviewer?

The author, Daksh, co-founder of Greptile, discovered that an AI bot was writing more pull requests than any individual human, raising the question of whether the author should also be the reviewer. This led to a discussion on the pros and cons of having AI review its own code, with points including the statelessness of LLMs, the difference between AI tools with the same engine, and the fact that AI-generated code may require closer reviewing due to its potential sloppiness and introduction of unique bugs.

Research

Pre-Trained Security LLM 8B

Foundation-Sec-8B is a cybersecurity-focused large language model built on the Llama 3.1 architecture and trained on a curated cybersecurity corpus to address the limited adoption of LLMs in cybersecurity. The model has been shown to match the performance of other leading models in certain cybersecurity-specific tasks, and its public release aims to accelerate the development and adoption of AI-driven cybersecurity tools.

The Deep Learning Model of Higher-Lower-Order Cognition, Memory, and Affection

The KAN (Kolmogorov-Arnold Networks) model was initially used to simulate disease dynamics, but has since been upgraded to ELKAN (Edge Learning KNN) or PNN (Plasticity Neural Networks), which incorporates edge learning and trimming. The ELKAN model, inspired by brain science, has been found to be more effective than KAN and can be used to explore various aspects of brain function, including consciousness, Alzheimer's disease, and the interaction between brain regions.

Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad

State-of-the-art large language models, such as Gemini-2.5-Pro, achieve impressive performance on mathematical competitions, but struggle with rigorous reasoning and proof generation, with most models scoring less than 5% on a comprehensive evaluation of full-solution reasoning for challenging mathematical problems. The results highlight the need for substantial improvements in reasoning and proof generation capabilities, as current models are inadequate for real-world mathematical tasks despite their ability to produce correct numerical answers.

Reinforcement Learning for Reasoning in LLMs with One Training Example

Reinforcement learning with verifiable reward using one training example (1-shot RLVR) significantly improves the math reasoning capabilities of large language models, with a single example elevating model performance on certain benchmarks by over 30%. This approach yields substantial improvements across various models, algorithms, and examples, and is primarily driven by the policy gradient loss, with promoting exploration also playing a critical role in its effectiveness.

Improving Instruct Models for Free: A Study on Partial Adaptation

Instruct models, while superior in following instructions, may forget pre-training knowledge or become overly conversational due to instruction tuning, leading to degraded few-shot learning performance. Reducing the strength of instruction tuning can improve few-shot in-context learning, but at the cost of losing some instruction following ability, highlighting a trade-off between these two abilities.

Code

Simple GPT in pure Go, trained on Jules Verne books

This repository contains a simple GPT implementation in pure Go, trained on Jules Verne books, which can be run and trained on a custom dataset. The code is designed for educational purposes, with a focus on simplicity and readability, and can be used as a companion to the "Neural Networks: Zero to Hero" course to understand the evolution of the model and its underlying mechanisms.

Show HN: Tgfeed – convert public Telegram channels into RSS or Atom feeds

tgfeed is a server that converts Telegram channels into RSS or Atom feeds, running on a specified port and utilizing a Redis instance for caching. It can be used by sending a GET request to the /telegram/channel/{username} endpoint with optional query parameters to customize the feed format, exclusions, and caching.

Show HN:Turn Your GitHub or Gitlab Repo into an AI Wiki (Open Source)

DeepWiki is a tool that automatically creates interactive wikis for GitHub or GitLab repositories, analyzing code structure, generating comprehensive documentation, and creating visual diagrams to explain how everything works. To use DeepWiki, users can enter a repository name, and the tool will organize the information into an easy-to-navigate wiki, with options for private repository support and AI-powered analysis.

DeepChat – A smart assistant that connects powerful AI to your personal world

DeepChat is a powerful open-source AI chat platform that provides a unified interface for interacting with various large language models, supporting both cloud and local models, with features such as search enhancement, tool calling, and multimodal interaction. It offers a range of advantages, including unified multi-model management, seamless local model integration, and a user-friendly interface, making it a versatile and efficient AI chat solution.

Show HN: Agent S: an open agentic framework that uses computers

Agent S is an open-source framework for building intelligent GUI agents that can learn from past experiences and perform complex tasks autonomously on a computer. The latest version, Agent S2, has achieved state-of-the-art results on several benchmarks, including OSWorld, WindowsAgentArena, and AndroidWorld, outperforming other models such as OpenAI's CUA/Operator and Anthropic's Claude 3.7 Sonnet Computer-Use.

    Redis returns to open-source roots with AGPLv3, a Pure Go GPT tackles Jules Verne, and a single training example boosts math reasoning in LLMs by over 30% with 1-shot RLVR.