5/16/2026
Ada-MK: Adaptive MegaKernel Optimization via DAG-Based Search for LLM Inference
Research on optimizing LLM inference kernels via DAG-based search for better performance.
5/16/2026
Show HN: Vyvoice: Privacy-first, cross-platform, offline voice transcription app
Cross-platform offline voice transcription app with voice commands and AI agent support. Built for Windows, Linux, macOS. Free tier with premium subscription.
5/16/2026
Show HN: Strava for AI coding – analytics on your Copilot/Claude/Codex usage
Analytics tool for tracking AI coding assistant usage from Copilot, Claude, and Codex. Provides insights on coding patterns.
5/16/2026
Show HN: Infinite Swap – Trade a bottle cap up to a house
Trading game exploring LLM latent space. Players trade items up from a bottle cap to a house, with LLM generating trade offers.
5/16/2026
Show HN: Agenda Intel MD – schemas and CLI to audit LLM strategic-risk briefs
Open-source MCP toolkit and JSON schemas for validating and auditing strategic intelligence agent outputs. Evidence layer for markdown-first reasoning.
5/16/2026
AI research papers are getting better, and it's a big problem for scientists
Commentary on impact of AI-generated research papers on scientific publishing.
5/16/2026
What Sudoku Reveals About AI Reasoning Architectures and the Future
Research on energy-based vs autoregressive models for constraint satisfaction problems. Introduces Kona 1.0, a commercial energy-based reasoning model.
5/16/2026
DeepSeek-V4-Flash means LLM steering is interesting again
Exploration of LLM steering techniques using DeepSeek-V4-Flash model and llama.cpp optimization for running capable models locally.
5/16/2026
Recent Developments in LLM Architectures: KV Sharing, MHC, Compressed Attention
Technical analysis of recent LLM architecture improvements: KV-cache sharing, multi-head compression, and attention optimizations for long-context efficiency.
5/16/2026
GitHub takes aim at Claude Code and Codex with its new Copilot app
News brief on GitHub Copilot app competing with Claude Code and Codex for code generation tasks.
5/16/2026
How to Use LLMs as Teachers?
Analysis of effective techniques for using LLMs as educational tools, critiquing common approaches and examining learning effectiveness.
5/16/2026
Stera: Open-Source Infra That Turns iPhones into Spatial Data for World Models
Open-source end-to-end pipeline converting iPhone data into spatial training data for embodied AI; includes 10M+ frame dataset.
5/16/2026
Can AI Un-Slop Itself?
Discussion of whether LLMs can filter and improve their own outputs; author's experience building a programming language with AI.
5/16/2026
ArXiv to Ban Researchers for a Year If They Submit AI Slop
ArXiv announces year-long author bans for submitting AI-generated research papers with plagiarism, errors, or inappropriate content.
5/16/2026
Keepithub – Physical-world geo-referenced memory for AI agents (B2A marketplace)
Keepithub: Brazilian marketplace where AI agents trade skills and features in real-time. Concept description, limited technical detail provided.
5/16/2026
TokenBBQ – track AI coding token usage across Claude, Codex, Gemini
TokenBBQ: open-source dashboard tool tracking AI coding token usage across Claude, Gemini, Codex and other LLM providers without installation.
5/16/2026
Designing, Refining, and Maintaining Agent Skills at Perplexity
Technical overview of Perplexity's modular Agent Skills architecture for frontier agent products across finance, law, health domains.
5/16/2026
Zero: The Programming Language for Agents
Zero: experimental systems programming language designed for AI agents with native tools, effects, memory management, and compiler output.
5/16/2026
A Single Neuron Is Sufficient to Bypass Safety Alignment in LLMs
Research showing single neuron suffices to bypass LLM safety alignment mechanisms.
5/16/2026
What rebuilding AlphaGo teaches us about self-play, RL, and future of LLMs [video]
Video analyzing AlphaGo reconstruction to extract insights on self-play, reinforcement learning, and LLM development.
5/16/2026
iOS-linuxkit: Optimized ish-ARM64 fork to provide a Linux userland to iOS apps
ios-linuxkit: Linux runtime for iOS enabling shell, compiler, and CLI tooling on iPhone/iPad via ARM64.
5/16/2026
Axiomind: Compiles Daily Notes into Axioms
Axiomind: Protocol for LLM agents to compile daily notes into structured knowledge systems compatible with Obsidian.
5/16/2026
We Built SynapseKit: The Truth About Production LLM Frameworks
SynapseKit: Production LLM framework built from first principles emphasizing minimal dependencies and async-first architecture.
5/16/2026
N8n-MCP: AI assistants can now search and build n8n workflows
n8n-MCP: Model Context Protocol server enabling AI assistants to search and construct n8n workflow automation nodes.
5/16/2026
Show HN: AI/ML benchmark for local LLM inference and XGBoost training on GPU/CPU
Benchmarking tool for local LLM inference and XGBoost training on GPU/CPU with reproducible results and HTML reports.
5/16/2026
Show HN: How-to-train-your-GPT. Every line commented
Interactive 12-chapter textbook teaching language model architecture and training from scratch with 7500+ commented lines covering attention mechanisms.
5/16/2026
We Pay Salesforce 83% More Than Last Year. But Stopped Using Notion
SaaStr analysis of how AI agents change B2B software economics and seat-based licensing models. Reports running 20+ agents with 3 humans.
5/16/2026
Uniting biological toolkits for a new approach to ALS
Co-Scientist bridges interdisciplinary research at MIT/Boston Children's Hospital, helping teams explore ALS by connecting different biological toolkits.
5/16/2026
Experience Layer for AI
CortexDB v1 launches memory/experience layer for AI agents with 93.8% benchmark performance and 53 integrations. Five-layer memory model.
5/16/2026
Goodbye Travel Agents, Hello AI Agents
Travel narrative demonstrating practical AI agent capabilities: itinerary rewriting, cross-referencing sources autonomously, real-time adaptation.
5/16/2026
Frontier AI has broken the open CTF format
Opinion piece on how frontier AI systems are disrupting competitive cybersecurity CTF competitions and skill measurement.
5/16/2026
A programming language made for agents
Zero programming language designed for human-AI agent collaboration. Systems language with explicit effects, predictable memory, structured compiler output.
5/16/2026
EY retracts study after researchers discover AI hallucinations
EY withdraws published study after discovering AI hallucinations in research methodology or findings.
5/16/2026
Tarvex ZM-1 – A compiler-free weight-stationary inference accelerator
Hardware inference accelerator for AI models without compiler requirement or weight movement. Built in 3 months.
5/16/2026
Show HN: Nexa-Gauge – LLM eval framework, now with self-hosted model support
Nexa-Gauge: Python evaluation framework for LLM/RAG systems with caching, cost estimation, and structured metrics. Self-hosted model support.
5/16/2026
Show HN: Plan-Graph based code generation with LLMs
Plan-graph based LLM code generation: creates implementation graphs instead of markdown, implements components iteratively. Novel LLM architecture approach.
5/16/2026
Jane Street's approach to AI adoption throughout their SDLC [video]
Jane Street video discussing AI adoption across their software development lifecycle. Practical industry insights.
5/16/2026
A Compression Tool for LLM Reads. Est. 60-95% Fewer Tokens
Headroom: compression tool reducing LLM token usage by 60-95% for agent reads, RAG, logs, and conversation history. Supports MCP protocol.
5/16/2026
LiteLLM Agent Platform: Run Claude Code/Codex On-Prem Sandboxes and Vaults
LiteLLM Agent Platform: Self-hosted infrastructure for running coding agents (Claude Code, Codex) in isolated sandboxes with credential vaults.
5/15/2026
We built an agent specialized for CI (using the 3 Claude models)
Mendral: CI-specialized agent using Claude models optimized for diagnosing CI failures, fixing flaky tests, and catching regressions.
5/15/2026
ExploitGym: Can AI agents turn bugs into exploits?
ExploitGym framework for developing AI agents that convert software bugs into exploits. Research on arXiv exploring agent capabilities in security.
5/15/2026
Gemini 3.5: frontier intelligence with action
Gemini 3.5 model family optimized for agentic workflows and complex long-horizon tasks. 3.5 Flash released with frontier performance for agents and coding.
5/15/2026
Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
Orthrus-Qwen3: Dual-architecture framework achieving 7.8× speedup in token generation with diffusion models while maintaining autoregressive output fidelity.
5/15/2026
Git is unprepared for the AI coding tsunami
GitHub infrastructure challenges from AI agent activity surge; developer discusses migration to alternative platforms.
5/15/2026
A tmux layer for coordinating multiple free tier coding-agent CLIs
Endy: tmux control plane coordinating multiple free-tier coding-agent CLIs (Gemini, OpenCode, CommandCode, Hermes) to hand off tasks.
5/15/2026
Show HN: A seed prompt that bootstraps a custom knowledge-base system
Custom knowledge-base system bootstrapped via LLM seed prompts, converting markdown directories to searchable HTML interfaces.
5/15/2026
Different models solve number-theory race problem
Benchmark comparing different LLM models solving number-theory problems with real-time competitive ranking by solution speed.
5/15/2026
Wikipedia: Writing articles with LLMs
Wikipedia policy prohibiting LLM-generated article content except for copyedits and translations to maintain content quality standards.
5/15/2026
HWE Bench: A new unbounded Benchmark for LLMs (GPT 5.5 is on top)
HWE Bench: benchmark evaluating LLM performance designing RISC-V CPUs with formal verification; GPT-4 designs outperform reference implementations.
5/15/2026
The Coding Harness Behind GitHub Copilot in VS Code
GitHub Copilot in VS Code: Architecture of the coding harness layer managing context, tools, agent loops, and tool interpretation.