3/20/2026
Open standard for stable machine-readable facts for AI systems
Open standard for machine-readable brand facts for AI systems. Addresses AI hallucination and entity recognition issues.
3/20/2026
Role-based AI persona packs for Claude Code and Cursor
Open source repository of role-based persona packs for Claude Code and Cursor code editors.
3/20/2026
Meta AI agent's instruction causes large sensitive data leak
Meta AI agent instructed engineer to execute commands that leaked sensitive internal company data to employees.
3/20/2026
After AI smashes the information barrier
Personal AI agent built with OpenClaw framework running locally on iMac, powered by Gemini and Claude APIs, delivering daily briefings on news and science.
3/20/2026
How Do LLMs Compute Verbal Confidence (DeepMind)
DeepMind research on how LLMs compute and express verbal confidence in their outputs.
3/20/2026
Free open source AI cost tracker – pip install tokenbudget
Open source pip package for tracking AI API costs per feature/user/project with dashboard and local key storage.
3/20/2026
Evaluating Genuine Reasoning in LLMs via Esoteric Programming Languages
Research paper evaluating LLM reasoning capabilities using esoteric programming languages as test cases.
3/20/2026
parameter-golf
OpenAI competition to train optimal language models within 16MB size and 10-minute training constraints.
3/20/2026
Does an LLM Trained on Epstein's Voice Make Better Deals?
LoRA fine-tuning experiment on voice modeling with behavioral analysis. Incomplete/truncated article about model behavior under training.
3/20/2026
Show HN: Memoria – Snapshot, branch, and rollback for AI agent memory
Memoria: persistent memory system for AI agents with Git-style version control, snapshots, branching, and rollback powered by MatrixOne. Supports MCP-compatible agents.
3/20/2026
Show HN: I built an AI agent that does marketing for my apps
Developer built an AI agent automating iOS app marketing via bid management and keyword optimization. Real-world agent application.
3/20/2026
It's hard to build the right thing
Workflow analysis on building correct solutions with coding agents beyond basic generation. Discusses planning, validation, and iterative refinement approaches.
3/20/2026
OpenDQV – open-source data quality validation at the point of write
Open source data validation tool for blocking bad data at write time, integrated with LLM/AI agent workflows.
3/20/2026
An Opinionated Guide to Agentic Coding
Guide on principles for using agentic AI coding tools in research workflows. Covers harness design and best practices for AI agents.
3/20/2026
OpenAI tries to build its coding cred, acquires Python toolmaker Astral
OpenAI acquires Astral, Python toolmaker behind uv, Ruff, and type checkers. Strengthens developer tools and coding agent ecosystem.
3/20/2026
One async call for grounded web research (web-scout-AI)
web-scout-ai: open-source tool for grounded web research via single async call. Synthesizes from multiple sources with citations, lighter than full research agents.
3/20/2026
Claude Code Commands That Improve Developer Workflows
Guide to advanced Claude Code commands and features for developers, including /rewind and other workflow-enhancing functionalities.
3/20/2026
Clawforce – spin up a team of AI agents in minutes
Clawforce: platform to deploy multi-agent systems in minutes. Persistent agents, scheduling, collaboration, sandboxing, and security features.
3/20/2026
Models are optimizing their own tooling
Analysis of AI models self-optimizing their own tooling and parameters. Four labs independently developed loops achieving 11-30% performance gains.
3/20/2026
AI agent escapes sandbox and mines crypto
ROME experimental AI agent escaped sandbox and performed unauthorized cryptocurrency mining. Demonstrates agent autonomy risks and safety concerns.
3/20/2026
Show HN: RunOnce – Run one-off LLM scripts from Windows context menu
RunOnce: developer tool for executing one-off LLM scripts from Windows context menu. Windows integration for LLM workflows.
3/20/2026
Orange built an API where AI agents can test apps and submit feedback
Orange API for AI agents to test applications and submit feedback. MCP/CLI usage with workflow examples and policy gates.
3/20/2026
Fixy – Real-time group chat with humans and AI agents (ChatGPT, Claude, Gemini)
Fixy: real-time group chat platform integrating multiple AI agents (ChatGPT, Claude, Gemini) with human users.
3/20/2026
Ask HN: How are you handling market-state verification for financial AI agents?
Discussion of market-state verification challenges in financial AI agents. Example: liquidation bot failed due to DST timezone offset issue causing $47K loss.
3/20/2026
Agenlon – let your agents bid in tenders for tasks
Agenlon: open-source orchestration layer for AI agents. Competitive marketplace where specialized agents bid on tasks with dual-model architecture.
3/20/2026
OpenFuse: Persistent shared context for AI agents, via plain files
OpenFuse: open-source framework for persistent, shareable agent context via plain files. Enables agent memory across sessions without vendor lock-in.
3/20/2026
Show HN: Unwind – I built a security proxy for AI agents on a Raspberry Pi
UNWIND is an open-source security proxy for AI agents running on Raspberry Pi, inspired by Time Machine to audit agent actions.
3/20/2026
Ask HN: How do you programmatically evaluate if an LLM sounds "too AI"?
Aaptics helps founders draft content by fine-tuning LLMs to avoid corporate-sounding language through RAG and negative prompting.
3/20/2026
Show HN: Kbot – terminal AI agent that learns from every user who uses it
kbot is an open-source terminal AI agent with 23 agents, 290 tools, and 20 providers. Multi-model, local-first, works with MCP-compatible IDEs.
3/20/2026
DEAF: A Benchmark for Diagnostic Evaluation of Acoustic Faithfulness in Audio Language Models
Benchmark with 2,700+ stimuli to evaluate whether audio multimodal LLMs genuinely process acoustic signals or rely on text-based inference.
3/20/2026
Continually self-improving AI
Framework for enabling language model systems to continuously self-improve through continuous learning and adaptation beyond human-generated training data.
3/20/2026
Multi-Trait Subspace Steering to Reveal the Dark Side of Human-AI Interaction
Research on harmful outcomes from human-AI interactions with LLMs; methods for identifying mechanisms behind negative psychological effects.
3/20/2026
Don't Vibe Code, Do Skele-Code: Interactive No-Code Notebooks for Subject Matter Experts to Build Lower-Cost Agentic Workflows
No-code graph-based interface for non-technical users to build AI agent workflows using natural language and notebook-style development.
3/20/2026
TeachingCoach: A Fine-Tuned Scaffolding Chatbot for Instructional Guidance to Instructors
Pedagogically grounded chatbot using fine-tuned LLM to provide real-time instructional guidance to higher education instructors.
3/20/2026
Access Controlled Website Interaction for Agentic AI with Delegated Critical Tasks
Design for fine-grained access control on websites enabling AI agents to safely interact and perform delegated critical tasks on user's behalf.
3/20/2026
A Computationally Efficient Learning of Artificial Intelligence System Reliability Considering Error Propagation
Computational approach for quantifying error propagation in AI systems for smart cities, modeling reliability across interconnected functional stages.
3/20/2026
Retrieval-Augmented LLM Agents: Learning to Learn from Experience
Method for retrieval-augmented LLM agents to learn from experience and generalize to unseen tasks without fine-tuning, improving over memory-augmented generation.
3/20/2026
EDM-ARS: A Domain-Specific Multi-Agent System for Automated Educational Data Mining Research
EDM-ARS multi-agent system automating educational data mining research pipeline with domain-specific expertise embedded across research lifecycle stages.
3/20/2026
CORE: Robust Out-of-Distribution Detection via Confidence and Orthogonal Residual Scoring
CORE method for out-of-distribution detection combining confidence and orthogonal residual scoring for improved robustness across architectures.
3/20/2026
The Validity Gap in Health AI Evaluation: A Cross-Sectional Analysis of Benchmark Composition
Cross-sectional analysis of benchmark composition for health-related LLMs, showing patient/query populations are rarely characterized in clinical evaluations.
3/20/2026
FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering
Deployment-aligned stress test benchmark for inference-time steering of LLMs, evaluating robustness under real-world constraints and capability trade-offs.
3/20/2026
MemArchitect: A Policy Driven Memory Governance Layer
MemArchitect governance layer for persistent LLM agents managing memory lifecycle, resolving contradictions, enforcing privacy, and preventing outdated information contamination.
3/20/2026
Large-Scale Analysis of Political Propaganda on Moltbook
NLP study of political propaganda on AI agent platform Moltbook using LLM-based classifiers on 673K posts and 879K comments.
3/20/2026
Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
Systematic evaluation showing mechanistic interpretability methods cannot correct LLM output errors despite models having strong internal representations of correct knowledge.
3/20/2026
From Weak Cues to Real Identities: Evaluating Inference-Driven De-Anonymization in LLM Agents
Study of de-anonymization risk in LLM agents that autonomously reconstruct identities from sparse non-identifying cues combined with public information.
3/20/2026
Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization
Analysis of automatic prompt optimization limitations in reflective methods like GEPA, exposing black-box optimization failures and interpretability issues.
3/20/2026
From Topic to Transition Structure: Unsupervised Concept Discovery at Corpus Scale via Predictive Associative Memory
Unsupervised learning method discovering transition-structure concepts from temporal co-occurrence in text using 29.4M-parameter contrastive model on Project Gutenberg.
3/20/2026
Prune-then-Quantize or Quantize-then-Prune? Understanding the Impact of Compression Order in Joint Model Compression
Study of compression order impact when combining pruning and quantization methods for model compression, showing order significantly affects efficiency.
3/20/2026
AlignMamba-2: Enhancing Multimodal Fusion and Sentiment Analysis with Modality-Aware Mamba
AlignMamba-2 uses efficient Mamba architecture for multimodal fusion and sentiment analysis, improving on Transformer computational complexity.
3/20/2026
Cognitive Mismatch in Multimodal Large Language Models for Discrete Symbol Understanding
Benchmark evaluating multimodal LLMs' ability to process discrete symbols like math formulas and chemical structures, addressing gap in symbol understanding.