4/15/2026
Show HN: Libretto – Making AI browser automations deterministic
Libretto: Open-source tool that generates deterministic browser automation scripts from AI agents instead of runtime prompts, enabling inspection and debugging.
4/15/2026
Why build on a single AI provider despite the risk? 3 AI models debated it
Discussion of risks and rationale for building applications dependent on single AI provider.
4/15/2026
WhatRuns extension malicious update: Full URL and AI chat exfiltration [video]
Security report: WhatRuns browser extension compromised to exfiltrate URLs and AI chat data.
4/15/2026
LeftGlove – MCP server that makes websites usable by AI agents (npx leftglove)
Open source MCP server simplifying website interaction for AI agents by filtering DOM noise and providing stable selectors.
4/15/2026
Show HN: L6e – Give your Agent a budget (save tokens, get smarter results)
Open source MCP server implementing token budgeting for AI agents to control costs and prevent context window bloat.
4/15/2026
Show HN: I built an AI music tagger for DJs, fusing metadata, audio DSP, and ML
Developer built AI-powered music tagging tool combining metadata lookup, audio DSP, and LLM processing for DJ library management.
4/15/2026
Show HN: SmokedMeat, like Metasploit, but for CI/CD (open-source)
Open-source red team framework for CI/CD security testing and vulnerability demonstration.
4/15/2026
LegendreGPT: Compressing a transformer into 15.7 MB with Legendre polynomials
LegendreGPT: transformer compression using Legendre polynomial coefficients, reduces 22 layers to 6 coefficient matrices achieving 15.7MB model size.
4/15/2026
Why I Am Doing This: Origin Story of Project-AI (Constitutional AI Governance)
Academic paper on constitutional framework for AI governance and enforcement mechanisms.
4/15/2026
Coding evals are broken. CI is green while AI code quality goes unmeasured
Technical analysis of broken AI code evaluation methods. SWE-bench marked contaminated; standard metrics fail to measure AI code quality impact.
4/15/2026
40% Increased Throughput 16.8% Less Energy for AI (Verified via ZKP)
Brief title about AI throughput and energy optimization improvements verified via zero-knowledge proofs.
4/15/2026
Show HN: Evo – parallel autoresearch experiments for Claude Code and Codex
Evo: Claude Code plugin for automated codebase optimization against metrics, built on Karpathy's autoresearch framework.
4/15/2026
Ask HN: Which LLM model and agentic CLI are you using for local development?
Discussion thread surveying local LLM models, hardware setups, and agentic CLI tools for development.
4/15/2026
Show HN: A semantic flow tool for embeddings
Experimental tool for computing and visualizing flow fields in semantic embedding spaces from trajectory data.
4/15/2026
Ask the Zuckbot? Meta Reportedly Building AI Clone of Mark Zuckerberg
Report that Meta is building an AI avatar of Mark Zuckerberg trained on his mannerisms and statements for employee interactions.
4/15/2026
A 164-parameter architecture beats a 6.5M transformer on SCAN by 94 points
HWTA: discrete routing architecture with 164 parameters outperforms 6.5M parameter transformer on SCAN compositional reasoning benchmark by 94 points.
4/15/2026
Five Things I Learned About Making AI Coding Agents Work
Analysis of AI coding agent behavior degradation over time, citing Meta/Harvard research on root causes beyond model limitations.
4/15/2026
GitHub – Heyitsnoah/Claudesidian
Open-source Obsidian vault configured for Claude Code integration, enables AI-powered note-taking and project management.
4/15/2026
Show HN: Compile English specs into 22 MB neural functions that run locally
ProgramAsWeights compiles English function descriptions into 22MB local neural functions with deterministic output, no API keys required.
4/15/2026
Reports of RAG's death have been greatly exaggerated
Technical post discussing RAG vs fine-tuning approaches for LLMs, referencing Andrej Karpathy's markdown wiki idea and introducing Atomic tool.
4/15/2026
Mapping GPUs to LLMs (and back): A bandwidth-based estimator for local inference
Technical guide for selecting GPUs and LLM models for local inference, addressing hardware-model compatibility for cost-effective local deployment.
4/15/2026
Show HN: IDE like, local API client
Postchi is an IDE-like local API client built with Tauri and React for focused API development.
4/15/2026
Agentic Engineering Methodology – Structured AI-Assisted Dev (Karpathy, Osmani)
Methodology by Karpathy and Osmani for structured AI-assisted development using agentic approaches. Title only, no details.
4/15/2026
KillBench shows every major LLM is biased against Americans
Claims major LLMs show bias against Americans. Title only, no substantive content provided.
4/15/2026
Show HN: ClaudeWatch – macOS menu bar usage monitor for Claude with notch pet
ClaudeWatch is macOS menu bar tool for tracking Claude Code token usage, rate limits, and costs with notch pet.
4/15/2026
Show HN: Voiden – API Workflows as Markdown Files (Obsidian × Curl)
Voiden is an open-source offline API tool that executes API requests as Markdown files with Git versioning, combining Obsidian-style workflows with curl functionality.
4/15/2026
ModelCascade – Route LLM calls to your own GPU first, cloud second
ModelCascade is open-source router for LLM calls that handles 74% of requests locally on GPU, escalating to cloud API when needed.
4/15/2026
Show HN: SafeWeave – 8 security scanners as an MCP server inside your AI editor
SafeWeave is an MCP server integrating 8 security scanners into AI editor for code analysis.
4/15/2026
LLM pricing is 100x Harder than you think
Discussion of LLM API pricing complexity and hidden variables. Title only, no substantive content.
4/15/2026
Project Think: building the next generation of AI agents on Cloudflare
Cloudflare Project Think: framework for building next-generation AI agents on their platform. Title only, no technical details.
4/15/2026
GPT-5.4 pro solves erdos problem #1196
GPT-5.4 Pro claims solution to Erdos problem #1196 with short proof; community verification requested.
4/15/2026
Friday, self-evolving assistant, only CC $100 plan, no agent framework
Friday is a self-evolving AI assistant running 24/7 on personal machines via Claude Code CLI and Telegram, learning autonomously.
4/15/2026
Show HN: A book that builds GPT-2, Llama 3, DeepSeek from scratch in PyTorch
Book implementing GPT-2, Llama 3, and DeepSeek architectures from scratch in PyTorch with progressive examples and real weight loading.
4/15/2026
I ran 5 social engineering attacks on AI. The failure modes are human
Analysis of LLM jailbreak attempts as social engineering rather than code exploits, examining failure modes through psychological manipulation lens.
4/15/2026
Vibe Coding Fails
User documents patterns in maintaining Claude-based agent loops with approval workflows and identifies failure modes in vibe coding.
4/15/2026
Playmakerly – AI agent that runs football prediction leagues in Slack
AI agent managing football prediction leagues within Slack. Demonstrates practical agent application for sports forecasting.
4/15/2026
Show HN: A /speak slash command for Claude Code that reads replies aloud
Claude Code /speak command that reads AI assistant replies aloud using system text-to-speech on macOS.
4/15/2026
Why the World Needs Open Source AI
Opinion on necessity of open source AI development. Advocacy piece for open AI model accessibility.
4/15/2026
Are the Costs of AI Agents Also Rising Exponentially? (2025)
Analysis of whether AI agent operational costs are rising exponentially. Cost economics for deployed AI agents.
4/15/2026
Show HN: RAG chunking playground: visualize how your docs get split
Interactive tool for visualizing RAG document chunking strategies. Developer tool for optimizing retrieval-augmented generation systems.
4/15/2026
Making prompts longer did not help. Making the task contract explicit did
Benchmark study on Python code tasks shows explicit task contracts improve LLM performance better than longer prompts alone.
4/15/2026
Show HN: LLM-primer – pre-warmed Claude Code session pool, zero startup wait
LLM-primer: pre-warmed Claude Code session pool eliminating 30-60s startup latency. Developer tool for maintaining persistent agent contexts.
4/15/2026
MCP as Observability Interface: Connecting AI Agents to Kernel Tracepoints
Using Model Context Protocol as observability interface to connect AI agents to kernel tracepoints. Technical approach for agent monitoring and debugging.
4/15/2026
Ask HN: Why/how do you use LLMs to draft tech blogs?
Survey asking how developers use LLMs to draft technical blogs. Community research on LLM content creation practices.
4/15/2026
1-week challenge using resilient LLMs (starts this Sunday)
ResilientLLM library for production-ready LLM integration handling failures, rate limits across multiple providers. Developer tool for reliable agent and LLM applications.
4/15/2026
Show HN: Lazyagent – TUI for to watch all your AI coding agents
Lazyagent is a terminal UI tool for monitoring multiple AI coding agents (Claude Code, Codex, OpenCode), displaying their events and tool calls in a unified interface organized by working directory.
4/15/2026
Show HN: Chat-rs, yet another LLM provider
Chat-rs: Rust LLM inference provider with streaming, tool calling, model routing, and human-in-the-loop. Open source developer tool with agent support.
4/15/2026
Show HN: Agent first social media scheduling
Agent-first social media scheduling tool. Application of AI agents to content management automation.
4/15/2026
AI learns language from skewed sources. That could change how we speak and think
Analysis of how LLM training on written text skews language representation compared to unscripted conversation. Research on language model training data bias.
4/15/2026
Show HN: Monogate – EML operator family, hybrid framework, 108-node sin(x)
36-hour implementation extending arXiv:2603.21852 symbolic regression paper. Creates hybrid EML operators achieving 52-74% node reduction with machine-precision results.