4/20/2026
Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations
Empirical study of LLM robustness to perturbations in chain-of-thought reasoning across five corruption types.
4/20/2026
ConFu: Contemplate the Future for Better Speculative Sampling
ConFu: Speculative decoding technique improving LLM inference speed by enhancing draft model quality for token verification.
4/20/2026
Beyond Augmented-Action Surrogates for Multi-Expert Learning-to-Defer
Multi-expert learning-to-defer framework addressing architectural limitations in expert selection and gradient distribution during training.
4/20/2026
COMPOSITE-Stem
COMPOSITE-STEM: 70 expert-written benchmark tasks in physics, biology, chemistry for evaluating AI agent reasoning on scientific discovery.
4/20/2026
The Amazing Agent Race: Strong Tool Users, Weak Navigators
The Amazing Agent Race benchmark with 1,400 directed acyclic graph tool-use puzzles for evaluating LLM agent navigation complexity.
4/20/2026
A Triadic Suffix Tokenization Scheme for Numerical Reasoning
Triadic Suffix Tokenization scheme improving LLM numerical reasoning by preserving digit structure and magnitude markers.
4/20/2026
KMMMU: Evaluation of Massive Multi-discipline Multimodal Understanding in Korean Language and Context
Korean-language multimodal benchmark with 3,466 questions covering nine disciplines and cultural context.
4/20/2026
Correcting Suppressed Log-Probabilities in Language Models with Post-Transformer Adapters
Post-transformer adapter technique to correct suppressed factual probabilities in aligned LLMs using 0.02% parameters.
4/19/2026
Swiss AI Initiative
Swiss government funding initiative for foundation model research and open science artifacts. €10M GPU hours available for core ML and applications.
4/19/2026
Headless Everything for Personal AI
Essay on headless APIs for personal AI agents to interact with services (passports, hotels, banking, shopping). Practical agent applications.
4/19/2026
Show HN: Auxx.ai – Customer Support CRM (Mix of Attio and N8n)
Auxx.ai: customer support CRM combining Attio and n8n automation. Self-built for managing support messages and customer data.
4/19/2026
Show HN: Agentjail – Self Hosted Freestyle.sh
Agentjail: minimal Linux sandbox for executing untrusted code in agents and build systems. Rust core in production, TypeScript SDK/UI open-source.
4/19/2026
Show HN: Nyx – multi-turn, adaptive, offensive testing harness for AI agents
Nyx: open-source autonomous testing harness for AI agents. Detects logic bugs, reasoning failures, edge cases, jailbreaks via red-teaming.
4/19/2026
How are you handling security for AI agents that use MCP tools?
Discussion on security practices for AI agents using MCP tools, including data sanitization, tool output monitoring, and isolation strategies.
4/19/2026
The demand for local AI could shape a new business model for Apple
Analysis of Apple Silicon optimization for local LLM inference and AI agents. Hardware demand surge driven by on-device AI model capability.
4/19/2026
OpenMythos – open-source Mythos alternative
Open-source theoretical implementation of Claude Mythos model using Recurrent-Depth Transformer architecture with looped recursion.
4/19/2026
Ask HN: Is giving AI agents DB access the new BI-tool problem?
Hacker News discussion comparing AI/ML team database access challenges to traditional BI tool implementations, exploring read replicas and data governance.
4/19/2026
Autoloom – Autonomous AI Agent built on tinyloom
Autoloom: lightweight autonomous AI agent (~1500 lines) built on tinyloom with state management, cron scheduling, TUI, and webhook server.
4/19/2026
Show HN: Lmcli v0.5.0 – A useable tool for LLM minimalists
lmcli v0.5.0: minimalist CLI tool for LLM interaction written in Go with agentic tool-calling support.
4/19/2026
Drawmode: MCP server for Excalidraw architecture diagrams with auto-layout
Drawmode is an MCP server that generates Excalidraw architecture diagrams with automatic layout.
4/19/2026
The LLM costs are not going up
Analysis arguing LLM operational costs are sustainable and declining, countering common claims about economic viability.
4/19/2026
GEPA prompt optimization: Claude Code Haiku +20% solve rate on new bugs
GEPA prompt optimization improved Claude Haiku's bug-solving rate by 20% on new bugs.
4/19/2026
Getting chat-tuned models to act kinda like base models
Techniques using fake tool calls to prompt-engineer chat-tuned LLMs into base model behavior
4/19/2026
Cardynal – AI support agent for businesses, no code, WhatsApp and web chat
Cardynal: No-code platform for deploying AI support agents with visual workflows, omnichannel inbox, and knowledge base.
4/19/2026
The Missing Human Half of AI
AbleCredit case study on building human-centered AI systems for non-technical loan officers in India using WhatsApp/Instagram interfaces.
4/19/2026
Show HN: Alodb – I got tired of pasting my Postgres schema into ChatGPT
Open-source macOS app converting natural language to optimized SQL queries using LLM schema understanding
4/19/2026
MCPs are the music of AI Agents. We built the player
Vinkius Desktop: Native application providing unified MCP server management across multiple AI clients on a single machine.
4/19/2026
Model Council
Debate-style app orchestrating multiple LLMs (GPT, Grok, Claude) with scoring and verification
4/19/2026
Show HN: Infrawise Azure Cloud Optimization
Infrawise: Azure cloud optimization tool using RAG backend and React frontend for AI-generated cost recommendations and resource analysis.
4/19/2026
Ethics of Artificial Intelligence and Robotics
Philosophical article examining ethical implications of AI and robotics development over 70 years. Theoretical rather than technical.
4/19/2026
Auto pentest your LLM endpoint and watch the chat in real-time
Tool for automated penetration testing of LLM endpoints with real-time chat monitoring. Minimal description provided.
4/19/2026
Building Autonomous Business Systems with AI Agents
Free online book on building autonomous business systems with AI agents, including architectural patterns and working code examples.
4/19/2026
FastStair: Learning to Run Up Stairs with Humanoid Robots
Model-free RL approach for humanoid robots to learn stair climbing with safety constraints
4/19/2026
Show HN: A privacy-first, local-LLM note app for iOS (Google Keep alternative)
Remen: iOS notes app with on-device LLM for natural language search and photo sync via iCloud. Privacy-first alternative to Google Keep.
4/19/2026
Opus 4.7 vs. 4.6 after 3 days of real coding side by side from my actual session
Real-world comparison of Claude Opus 4.7 vs 4.6 performance metrics from live coding sessions
4/19/2026
543 Hours: What happens when AI runs while you sleep
Data analysis of 543 autonomous coding hours using Claude Code agent showing productivity metrics: 14,926 prompts, 2,314 sessions, 165 releases.
4/19/2026
Rensei – let agents code 3D models and screenshot them. then 3D print
Rensei: CLI enabling AI agents to generate 3D models iteratively from code with screenshot feedback, then 3D print. Includes skill file for agents.
4/19/2026
The State of LLM Bug Bounties in 2026
Article title about LLM bug bounty landscape in 2026. No content provided.
4/19/2026
CNNs + VLM outperforms pure VLMs on OCR
Study comparing CNN-based OCR with vision language models, showing hybrid approach combining both architectures outperforms either alone.
4/19/2026
Show HN: I built an open source and secure infrastructure for internal apps
RootCX: Open-source production infrastructure for internal AI-coded apps and agents. Includes database, auth, secrets, scheduling, file storage.
4/19/2026
Passmark: Open-source Playwright library for AI regression testing
Passmark open-source Playwright library for AI regression testing, enables automated test validation.
4/19/2026
Sandboxed AI agent orchestration platform
Stub title about sandboxed AI agent orchestration platform; no content provided.
4/19/2026
Scoring 500 Show HN Pages for AI Design Patterns
Analysis of 500 Show HN pages scoring them against 15 common AI design patterns to quantify sterile aesthetic in AI-generated landing pages.
4/19/2026
Clerk – Auto-summarize Claude Code sessions into plain Markdown
Clerk tool auto-summarizes Claude Code sessions to Markdown format for documentation.
4/19/2026
MCP Spine – Middleware proxy for MCP with security, routing, and token control
MCP Spine middleware proxy for Model Context Protocol with security, routing, token control, and compliance between LLM clients and tool servers.
4/19/2026
Generate real Android APKs on-device
iappyxOS: AI-powered on-device Android APK generator without cloud/servers; users describe apps to AI for local building and installation.
4/19/2026
Ask HN: Do you spend more time conversing with AI or humans?
Hacker News discussion about spending more time conversing with AI than humans, reflecting on agentic coding tools and interactions.
4/19/2026
Show HN: verifiable API and MCP calls
PEAC: tool for verifiable API and MCP calls using portable signed records, supports Node.js, Go, Python with offline verification.
4/19/2026
Claude system prompts as a Git timeline
Git timeline visualization of Claude system prompts showing changes between versions; created by converting Anthropic's published prompts into commits.
4/19/2026
Show HN: Libredesk – self-hosted, single-binary Intercom/Zendesk alternative
Libredesk: open-source helpdesk alternative to Zendesk/Intercom built with Go backend and Vue frontend, fully free with no enterprise paywall.