3/18/2026
Claw Compactor: compress LLM tokens 54% with zero dependencies
Open source LLM token compression engine achieving 54% compression via 14-stage fusion pipeline with zero inference cost.
3/18/2026
Show HN: PromptToMars – A Prompt Generator and Optimizer
Web tool for creating, optimizing and organizing prompts across multiple use cases with templates and search.
3/18/2026
Results from round one of First Proof (benchmarking LLMs for math research)
Benchmark results from First Proof initiative testing LLM capabilities for pure mathematics research contributions.
3/18/2026
The Pentagon is planning for AI companies to train on classified data, defense
Pentagon plans secure environments for AI companies to train classified military models on sensitive defense data.
3/18/2026
LLM Exposure for EU Jobs
Interactive treemap visualization of AI exposure risk by occupation in Netherlands and Europe, based on Andrej Karpathy's US model.
3/18/2026
ZCA: Slice-Isolated Execution for Coding Agents
Slice-Agent-Bench research on coding agents using slice-isolated execution instead of full repo context, testing minimal codebase projection efficiency.
3/18/2026
Multi-tentacled orchestrator for Claude Code
Claude Code plugin orchestrating multiple LLMs (Codex, Gemini, Claude) with distinct roles and adversarial review to reduce model blind spots.
3/18/2026
Community Security Scans: Crowd-Sourced Trust for AI Agent Skills
Community crowdsourced security scanning approach for AI agent skills, focusing on trust and safety validation.
3/18/2026
Engram – salience-gated memory for Claude Code (captures what matters)
Salience-gated memory system for Claude Code. Scores agent observations on 5 dimensions, consolidates patterns via dream cycles. Builds structured memory of tool usage.
3/18/2026
Show HN: Lumbox – Email infrastructure for AI agents (lumbox.co)
Email infrastructure for AI agents. Provides dedicated inboxes, OTP polling endpoint, Steel Browser integration for signup automation.
3/18/2026
Ask HN: Burning $100K/week on LLM tokens – what are you doing to cut costs?
Discussion: LLM cost optimization for AI agent fleet ($103K/week). Strategies include prompt caching, model routing, context truncation.
3/18/2026
Show HN: QCCBot – Android in a browser tab, with AI agent control
Remote Android instances streamed to browser, controllable by AI agents. No hardware/cables needed. Targets e-commerce, mobile testing automation.
3/18/2026
Show HN: ROMA runs multiple coding agents simultaneously – Claude, Codex, etc.
ROMA orchestration tool running multiple coding agents simultaneously (Claude, Codex, Gemini, Copilot) with structured deliberation for coherent results.
3/18/2026
Echoo – Free, open-source macOS app for system-wide AI via shortcuts and voice
Open-source macOS app for system-wide AI access via shortcuts and voice, with local model support.
3/18/2026
Run any LLM on any hardware. Auto-detects your GPU, checks if the model fits
Zero-configuration LLM inference. Auto-detects GPU, checks model fit, downloads correct format. Supports multiple model formats. Dashboard-based model selection.
3/18/2026
Show HN: Open-source YouTube summary, transcript chat, and timeline sidepanel
Open-source Chrome extension for YouTube video interaction: transcript-first chat, summaries, timeline browsing, subtitle export.
3/18/2026
Contract-Centered Iterative Stability v4.7.3
Research on stability boundaries in iterative AI-assisted software development across Claude and GPT LLM families.
3/18/2026
Thrum – Agent coordination through messaging
Thrum persistent messaging layer for AI agents across sessions and machines, with CLI and MCP server support for Claude Code agents.
3/18/2026
I Don't Know ML. Claude Does. 0.871 F1 on Predicting Linux Game Compatibility
ML project predicting Linux game compatibility (0.871 F1) using human-AI collaboration. Software engineer with minimal ML background used Claude to discover novel statistical techniques.
3/18/2026
Agent Package Manager
Open-source dependency manager for AI agents (apm). Standardizes agent configuration, skills, prompts across Claude, Copilot, Cursor. Similar to package.json for agents.
3/18/2026
Show HN: N0x – LLM inference, agents, RAG, Python exec in browser, no back end
N0x browser-based LLM inference, autonomous agents, RAG, and Python execution with WebGPU and Pyodide, no backend or data sharing.
3/18/2026
Superpowers
Development workflow framework for AI coding agents. Composable skills and instructions guide agent behavior from initial requirements gathering through code generation.
3/18/2026
Knowledge workers managing AI show collapsed productivity, not just a plateau
Analysis of longitudinal workplace studies showing productivity collapse when workers manage AI, contradicting optimistic coverage.
3/18/2026
Meta's Omnilingual MT for 1,600 Languages
Meta's machine translation system extending to 1,600 languages using LLM approaches; advances beyond 200-language NLLB coverage.
3/18/2026
Show HN: Llmtop – Htop for LLM Inference Clusters (vLLM, SGLang, Ollama, llama)
Real-time terminal dashboard (llmtop) for monitoring LLM inference clusters supporting vLLM, SGLang, and Ollama with KV cache and latency metrics.
3/18/2026
Your terminal, finally has memory!
Terminal tool with local AI memory using Ollama. Save/recall commands, notes, URLs via natural language. Runs locally, no cloud.
3/18/2026
Rust-accelerated reinforcement learning, 140x faster than Python
Rust-accelerated RL framework using Polars pattern: Rust data plane + Python control plane via PyO3. 140x speedup with Rayon parallelism. Published on crates.io with 695 tests.
3/18/2026
Show HN: Vibe – Code from your bed or the park with Claude Code/Codex
Vibe is a mobile app enabling remote code execution with Claude Code and Gemini CLI, with web preview and session management.
3/18/2026
My Claude Code setup you definitely shouldn't use. It's AI Overkill
Personal setup combining Claude Code with specialized domain agents, parallel code review, and self-improving knowledge systems.
3/18/2026
We optimized Dash's relevance judge with DSPy
Dropbox optimized their relevance judge using DSPy for Dash, improving ranking and evaluation across multiple ML pipelines at scale.
3/18/2026
TrustAgentAI – Cryptographic receipts for MCP tool calls (non-repudiation layer)
TrustAgentAI is an open-source accountability layer adding cryptographic receipts and non-repudiation to MCP tool calls for AI agents.
3/18/2026
Gas Town by Kilo
Gas Town is Steve Yegge's agent orchestrator coordinating multiple AI coding agents simultaneously, hosted on Kilo Cloud infrastructure.
3/18/2026
Neural-Symbolic Logic Query Answering in Non-Euclidean Space
HYQNET is a neural-symbolic model that answers complex first-order logic queries on knowledge graphs by integrating interpretability with generalization.
3/18/2026
NextMem: Towards Latent Factual Memory for LLM-based Agents
NextMem proposes a latent factual memory framework for LLM-based agents to address limitations of existing textual and parametric memory approaches.
3/18/2026
AIDABench: AI Data Analytics Benchmark
AIDABench: Comprehensive benchmark for AI data analytics and document understanding. Evaluates end-to-end task effectiveness in practical document processing scenarios.
3/18/2026
The Comprehension-Gated Agent Economy: A Robustness-First Architecture for AI Economic Agency
Comprehension-Gated Agent Economy: Formal architecture linking AI agent economic permissions to verified comprehension. Robustness-first approach to agent authorization.
3/18/2026
CraniMem: Cranial Inspired Gated and Bounded Memory for Agentic Systems
CraniMem: Neurocognitively-inspired gated and bounded multi-stage memory design for long-running LLM agents. Improves retention stability and content consolidation.
3/18/2026
GSI Agent: Domain Knowledge Enhancement for Large Language Models in Green Stormwater Infrastructure
GSI Agent: Domain knowledge enhancement for LLMs in green stormwater infrastructure. Combines LLM with domain knowledge for inspection and maintenance guidance.
3/18/2026
Did You Check the Right Pocket? Cost-Sensitive Store Routing for Memory-Augmented Agents
Cost-sensitive store routing for memory-augmented agents. Formulates selective memory retrieval as routing problem to reduce context tokens and improve efficiency.
3/18/2026
DynaTrust: Defending Multi-Agent Systems Against Sleeper Agents via Dynamic Trust Graphs
DynaTrust: Defense mechanism against sleeper agents in multi-agent systems using dynamic trust graphs. Detects agents that hide malicious behavior until triggered.
3/18/2026
QV May Be Enough: Toward the Essence of Attention in LLMs
Theoretical analysis of Query-Value mechanism in Transformers from linguistic perspective. Explains efficacy of MQA, GQA, and MLA architectures and trade-offs.
3/18/2026
Compiled Memory: Not More Information, but More Precise Instructions for Language Agents
Atlas: Memory kernel that compiles task experience into agent instructions without fine-tuning or RAG. Improves agent memory utility via instruction-level compilation.
3/18/2026
Quantum-Secure-By-Construction (QSC): A Paradigm Shift For Post-Quantum Agentic Intelligence
Quantum-Secure-By-Construction design paradigm for agentic AI systems. Addresses post-quantum cryptographic challenges in long-lived distributed agent deployments.
3/18/2026
I Know What I Don't Know: Latent Posterior Factor Models for Multi-Evidence Probabilistic Reasoning
Latent Posterior Factors framework for aggregating multiple noisy evidence sources without manual feature engineering. Addresses uncertainty in real-world decision-making.
3/18/2026
Theoretical Foundations of Latent Posterior Factors: Formal Guarantees for Multi-Evidence Reasoning
Theoretical characterization of Latent Posterior Factors for aggregating heterogeneous evidence in probabilistic prediction. Formal guarantees for multi-evidence reasoning.
3/18/2026
Context-Length Robustness in Question Answering Models: A Comparative Empirical Study
Empirical study measuring LLM robustness to increasing context length on SQuAD and HotpotQA. Analyzes accuracy degradation with context size.
3/18/2026
CUBE: A Standard for Unifying Agent Benchmarks
CUBE: Universal benchmark standard for AI agents built on MCP and Gym. Addresses fragmentation by allowing benchmarks to be wrapped once and used everywhere.
3/18/2026
Prose2Policy (P2P): A Practical LLM Pipeline for Translating Natural-Language Access Policies into Executable Rego
Prose2Policy: LLM pipeline translating natural-language access control policies into executable Rego code. End-to-end pipeline with test generation and validation.
3/18/2026
Persona-Conditioned Risk Behavior in Large Language Models: A Simulated Gambling Study with GPT-4.1
Empirical study of GPT-4.1 behavior in gambling tasks under different persona prompts. Examines whether LLM risk behavior reflects principled patterns or prompt mimicry.
3/18/2026
Regularized Latent Dynamics Prediction is a Strong Baseline For Behavioral Foundation Models
Regularized latent dynamics prediction as baseline for behavioral foundation models, examining how state feature choice affects task adaptability and reward function expressivity.