OpenTSLM: Language models that understand time series

A new class of artificial intelligence models called Time-Series Language Models (TSLMs) has been developed, which can reason about and forecast temporal data, such as heartbeats and sensor pulses, in natural language. TSLMs have the potential to transform various industries, including healthcare, robotics, and infrastructure, by enabling proactive and autonomous decision-making, and are being released in both open and proprietary forms to power a global developer and research ecosystem.

Read more

Evaluating the impact of AI on the labor market: Current state of affairs

The introduction of AI, particularly generative AI like ChatGPT, has not had a discernible disruption on the broader labor market, with changes in the occupational mix being relatively small and not markedly different from past periods of technological change. The data suggests that the effects of AI on the labor market are evolving and may take longer than expected to materialize, with current trends being similar to those seen during the adoption of the internet and computers, which took decades to transform the workforce.

Read more

Announcing Tinker

Tinker is a flexible API for fine-tuning language models, allowing researchers and hackers to experiment with models while handling the complexity of distributed training, and is now available in private beta. The Tinker platform provides low-level primitives and an open-source library, the Tinker Cookbook, to help users achieve good results and has already been used by groups at universities such as Princeton, Stanford, and Berkeley for various research projects.

Read more

Show HN: ChartDB Agent – Cursor for DB schema design

ChartDB is a database design tool that allows users to create and visualize database schemas, with features such as templates, filtering, and drag-and-drop functionality. Users can choose from pre-made templates inspired by popular companies like Spotify and Instagram, or design their own database from scratch using a conversational interface.

Read more

The RAG Obituary: Killed by agents, buried by context windows

The author, with a decade of experience in AI and search, believes that Retrieval-Augmented Generation (RAG) architectures, which have dominated AI for the past three years, are on the decline due to limitations in handling large knowledge bases. Despite efforts to optimize RAG, such as developing sophisticated chunking strategies and hybrid search methods, the author argues that these solutions are inadequate and that new architectural patterns, such as agent-based architectures, are emerging to address the fundamental problem of working with large amounts of data.

Read more

The Missing Link Between the Transformer and Models of the Brain

The Dragon Hatchling (BDH) model is a new Large Language Model architecture inspired by the brain's scale-free biological networks, offering strong theoretical foundations, interpretability, and state-of-the-art performance rivaling Transformer models like GPT2. BDH's biologically plausible design, which includes synaptic plasticity and spiking neurons, allows for interpretability of its state and activation vectors, making it a unique and promising approach to language modeling and universal reasoning.

ArXiv

Efficient LLM:Bandwidth, Compute, Synchronization, and Capacity are all you need

This study examines the performance bottlenecks of transformer-based large language models, analyzing the impact of memory bandwidth, capacity, and synchronization overhead on distributed inference systems. The research reveals key findings, including the need for high memory bandwidth, the advantages of DRAM-based designs, and the potential for hardware advancements to improve performance, but also highlights the need for algorithmic advances to achieve higher user throughput.

ArXiv

Evaluating LLM-Generated Detection Rules in Cybersecurity

An open-source evaluation framework and benchmark metrics have been developed to assess the effectiveness of Large Language Model (LLM)-generated cybersecurity rules, providing a more realistic evaluation of their usefulness. The framework uses a holdout set-based methodology to compare LLM-generated rules with human-generated ones, offering three key metrics to measure their effectiveness and building trust in LLM-based security rule generators.

ArXiv

The AI Productivity Index – LLMs by Economic Impact

The AI Productivity Index (APEX) is a benchmark that assesses the ability of AI models to perform high-value knowledge work in domains such as investment banking and law. The first version of APEX evaluated 23 frontier models, with GPT 5 achieving the highest score, but still showing a significant gap between the performance of even the best models and human experts.

ArXiv

Introduction to Machine Learning(2024)

This book provides a comprehensive introduction to the mathematical foundations of machine learning, covering topics such as calculus, linear algebra, probability, and optimization, as well as specific algorithms like stochastic gradient descent and neural networks. The book progresses through various machine learning concepts, including supervised and unsupervised learning, generative models, and deep learning, ultimately concluding with a theory-oriented chapter on concentration inequalities and generalization bounds.

ArXiv

Show HN: FSST – Fast Static Symbol Table Compression Library in Go

FSST is a fast string compression algorithm optimized for random access and decompression speed, achieving high compression ratios on structured and repetitive text by learning a compact symbol table from training data. The implementation provides a balance between compression ratio, decompression speed, and memory efficiency, making it ideal for structured data workloads such as JSON, CSV, logs, and XML.

Repo

Show HN: Fuzz Forge, vulnerability discovery with AI and fuzzing

FuzzForge is an open-source platform that uses AI and fuzzing frameworks to automate application security and offensive security workflows, allowing users to orchestrate static and dynamic analysis, automate vulnerability research, and scale AppSec testing. The platform is currently under active development and offers a range of features, including AI agents, workflow automation, and fuzzer integration, with the goal of empowering security teams, researchers, and the community to improve application security.

Repo

Show HN: Ocrisp, One-Click RAG Implementation, Simple and Portable

OCRISP is a simple desktop application that serves as a GUI, CLI, and MCP tool, allowing users to embed PDFs and connect with MCP clients to retrieve data. The project is in its early stages, requiring Qdrant and Ollama to be installed, and has a limited feature set, but improvements are planned, including support for other vector databases.

Repo

Folke/sidekick.nvim – Your Neovim AI sidekick

Sidekick.nvim is a Neovim plugin that integrates Copilot LSP's "Next Edit Suggestions" with a built-in terminal for AI command-line tools, allowing for streamlined coding and interaction with AI assistants without leaving the editor. The plugin offers features such as automatic suggestions, rich diffs, hunk-by-hunk navigation, and statusline integration, as well as support for popular AI tools like Claude, Gemini, and Grok.

Repo

Show HN: Claude Code 2.0 router – preference-aligned routing to multiple LLMs

Arch is a modular edge and AI gateway designed to simplify the process of building agentic apps by handling low-level tasks such as routing, model integration, and observability. It provides a language and framework-agnostic infrastructure layer that enables developers to build and ship agentic apps faster, with features including routing to agents and LLMs, guardrails, and observability tools.

Repo