5/19/2026
Tears – vibe code responsibly with tiered trust levels
Tears: Open-source tool for tiered code review and trust levels within repositories. Code governance tool.
5/19/2026
Git-Backed SaaS: SaaS Without a Service
Kiln: Git-backed SaaS application storing datasets and evals in local git repos with real-time sync. Developer tool.
5/19/2026
Show HN: Tokoro – open, signed event protocol with LLM crawler
Open source event discovery protocol using LLM crawlers to aggregate events across platforms.
5/19/2026
Hold-to-talk voice input for Pi Coding Agent
Hold-to-talk voice input feature for Pi Coding Agent with cloud streaming via Deepgram or offline local model support.
5/19/2026
The Economist prepares for a two‑track Web: one for humans one for AI agents
The Economist explores web infrastructure split between human and AI agent access patterns.
5/19/2026
eXo MCP server: expose workplace tools to AI agents with OAuth
eXo MCP server enables workplace tool integration with AI agents via OAuth for secure authorization.
5/19/2026
Thinking of harness on mobile? I ported OpenAI Agents SDK to React Native
React Native port of OpenAI Agents SDK enabling multi-agent, realtime, and voice-powered AI applications on mobile platforms.
5/19/2026
Anthropic shuts the EU out of its most advanced cyber AI model
News: Anthropic restricts Claude Mythos cybersecurity model to select U.S. companies, limiting EU access.
5/19/2026
Miro Fish: Predicting Anything
MiroFish is a multi-agent AI prediction engine that constructs digital simulations with thousands of intelligent agents for forecasting.
5/19/2026
LLM Wiki app Chunker – transform documents into navigable knowledge trees
Chunker: LLM tool transforming documents into hierarchical, navigable knowledge trees with progressive disclosure for AI or human exploration.
5/19/2026
Show HN: Viberia – Civ/Polytopia-like command center for AI agents (BYOK/BYOS)
Viberia: Isometric map-based agent harness (Tauri app) for organizing and sequencing AI agents in configurable team structures with low resource usage.
5/19/2026
Linux 7.1-rc4: security list "almost unmanageable" from AI bug reports
Linux kernel team struggles with flood of low-quality automated AI-generated bug reports overwhelming security list.
5/19/2026
Causal Video Models Are Data-Efficient Robot Policy Learners
Video generation-based robot policy learning approach enabling data-efficient task learning and one-shot generalist robotics.
5/19/2026
Using algebra and LLMs to verify a flight-plan bug fix in Lean
Using LLM agents for formal code verification in Lean, proving real air traffic control bug fix correct.
5/19/2026
DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models
Research on credit assignment for reinforcement learning in diffusion language models, addressing temporal credit assignment and likelihood bias across denoising steps.
5/19/2026
LoopQ: Quantization for Recursive Transformers
LoopQ addresses quantization challenges in recursive Transformers (LoopLMs), identifying distribution shift, state reuse, and error accumulation issues for parameter-efficient language models.
5/19/2026
LLMCap – A proxy that hard-stops LLM API calls when you hit a dollar cap
Proxy service that hard-stops LLM API calls at specified dollar cap, preventing surprise bills across providers.
5/19/2026
Frontier models at open source cost – hot new AI Model Router
Model router that grades prompts and routes hard reasoning to frontier models, routine work to open-source, reducing costs ~40%.
5/19/2026
TuriX AI launched the latest version
Open-source computer-use AI agent achieving 80% success on Mac benchmark and 64.2% on OSWorld, enabling desktop automation.
5/19/2026
Compute Optimal Tokenization: Scaling Laws for Data Compression in LLMs
Research on tokenization scaling laws showing impact of compression rate on language model training under fixed compute budgets.
5/19/2026
Benchmarking llama.cpp's new MTP support on Strix Halo
Benchmarking Multi-Token Prediction speculative decoding in llama.cpp achieving 2.5× speedups on Qwen and Llama models.
5/19/2026
Sieve – scans Cursor/Claude chat history for leaked API keys
Tool that scans Cursor and Claude chat history for leaked API keys stored in plaintext SQLite databases by AI coding assistants.
5/19/2026
We built a runtime activation layer for autonomous AI agents
Runtime activation layer for autonomous AI agents (limited details provided).
5/19/2026
The last six months in LLMs in five minutes
Summary of LLM developments from past six months presented at PyCon US 2026, highlighting November 2025 inflection point in coding.
5/19/2026
Nitsum: Serving Tiered LLM Requests with Adaptive Tensor Parallelism
Nitsum: LLM serving system that dynamically adjusts tensor parallelism at runtime to handle mixed latency-critical and batch workloads on fixed GPU budgets.
5/19/2026
SuperInfer: SLO-Aware Rotary Scheduling and Memory Management for LLM Inference
SuperInfer: LLM inference system addressing SLO-aware scheduling and KV cache memory management under high request rates.
5/19/2026
Video GTP
NeoTube/Khan viewer: AI system that indexes video content and answers questions with precise timestamps.
5/19/2026
Make products AI agents want
Analysis of building products for AI agents as primary users, with metrics for agent adoption and engagement patterns.
5/19/2026
AgentCRM – Headless CRM for Claude Code
AgentCRM is a headless CRM optimized for Claude Code agent usage via MCP, designed to handle CSV lead data and avoid context bloat from traditional CRM APIs.
5/19/2026
What political censorship looks like inside an LLM's weights (Qwen 3.5)
Mechanistic interpretability study examining how content filtering is implemented in Qwen 3.5 LLM weights and its political censorship mechanisms.
5/19/2026
From-scratch reimplementation of Mythos Glasswing pipeline
MIT-licensed reimplementation of Cloudflare's Project Glasswing vulnerability discovery agent using Claude Code Agent SDK with 8-stage agentic pipeline.
5/18/2026
Show HN: Vecdb – local-first hybrid vector database in Rust (HNSW and BM25)
Vecdb is an open-source Rust vector database supporting hybrid search with HNSW/IVF dense retrieval and BM25 keyword search with configurable result fusion.
5/18/2026
Show HN: Handoff – preserve coding context when agents run out of tokens
Handoff tool preserves coding context between AI agents when token limits are hit, enabling multi-agent code refactoring workflows without server dependencies.
5/18/2026
AgentVoy – The create-react-app for AI agents
AgentVoy: CLI tool for scaffolding and deploying multi-agent AI applications with DevTools, supports multiple frameworks and models.
5/18/2026
LoRA and Weight Decay (2023)
Technical analysis of LoRA fine-tuning interactions with weight decay, showing LoRA solves a different optimization problem than full fine-tuning.
5/18/2026
Big AI is subverting regulations just like tobacco and oil firms
Academic study warns AI industry uses regulatory capture tactics similar to tobacco and oil firms to influence policy.
5/18/2026
A Geometric Calculator Inside a Neural Network
Research paper describing a geometric addition mechanism discovered inside Llama 3.1 8B operating on circular number manifolds.
5/18/2026
AgentVoy – The create-react-app for AI agents (7 frameworks, deploy anywhere)
AgentVoy CLI/SDK tool for scaffolding production-ready AI agent projects across multiple frameworks with guardrails and deployment.
5/18/2026
Skills for Testing Distributed Systems
Markdown-based skills for AI coding agents to design and execute distributed systems tests in Jepsen style.
5/18/2026
LLMs have no structural place for non-knowledge
Philosophical analysis questioning whether current LLM trajectory of scale is the only viable path for AI development.
5/18/2026
AI-generated code that works – and proves it
Combines AI code generation with formal methods to produce verified code for safety-critical systems beyond testing.
5/18/2026
Managing metadata is essential in LLM world
Analysis of metadata management's increased importance in LLM workflows for maintaining data context and information mechanics.
5/18/2026
Show HN: Tracecast – open-source generative data apps built on top of Marimo
Open-source generative data apps platform combining Marimo notebooks, LangGraph agents, and data warehouse query tools.
5/18/2026
Running PyTorch Models on Apple Silicon GPUs with the ExecuTorch MLX Delegate
ExecuTorch MLX Delegate enables GPU-accelerated PyTorch model inference on Apple Silicon Macs via MLX framework.
5/18/2026
Show HN: Feature flags on Redis you use – a low cost solution
Feature flag management system built on Redis for teams already running Redis infrastructure.
5/18/2026
Needle-rs – AI Function calling in the browser, 258 KB WASM
26M-parameter tool-calling transformer running entirely in browser via 258KB WASM runtime with no server dependency.
5/18/2026
Python implementation for text generation using EEG signals from the brain
Thought2Text is a Python neural decoding pipeline for brain-to-text conversion. Supports invasive intracortical and non-invasive EEG-based typing reconstruction.
5/18/2026
Giving operational continuity to coding agents, not everything is about memory
Technical discussion on operational continuity for coding agents beyond memory mechanisms.
5/18/2026
GAX – A tool execution protocol to fix MCP token bloat and CLI auth
GAX is an open protocol for AI agents to call external tools with CLI ergonomics. Handles OAuth, policy, and audit in a sidecar to reduce token usage.
5/18/2026
Fixing LLM Writing with Distribution Fine Tuning
Distribution Fine Tuning (DFT) algorithm fixes formulaic LLM writing by matching output distribution to training data. Improves MMD by 49%.