Ax Zixuan Gong, Shijia Li, Yong Liu, Jiaye Teng 15d ago

Disentangling Feature Structure: A Mathematically Provable Two-Stage Training Dynamics in Transformers

Theoretical analysis of two-stage training dynamics in transformers showing progression from syntactic to semantic feature learning.

Ax Parsa Hosseini, Sumit Nawathe, Mazda Moayeri, Sriram Balasubramanian, Soheil Feizi 15d ago

SpurLens: Automatic Detection of Spurious Cues in Multimodal LLMs

Pipeline detecting spurious visual cues in multimodal LLMs using GPT-4 and object detectors without human supervision.

Ax Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Miles, Seraphina Nix, Tao Lin, Chris Painter, Neev Parikh, David Rein, Lucas Jun Koba Sato, Hjalmar Wijk, Daniel M. Ziegler, Elizabeth Barnes, Lawrence Chan 15d ago

Measuring AI Ability to Complete Long Software Tasks

Metric and benchmark for measuring AI ability to complete long software tasks by correlating model performance to human completion time.

Ax Eugen Barbulescu, Antonio Alexoaie, Lucian Busoniu 15d ago

Hyperflux: Pruning Reveals Importance

Network pruning method modeling pruning as flux-driven system with theoretical analysis of weight removal importance.

Ax Adam \v{S}torek, Mukur Gupta, Samira Hajizadeh, Prashast Srivastava, Suman Jana 15d ago

Sense and Sensitivity: Examining the Influence of Semantic Recall on Long Context Code Understanding

Study evaluating whether LLMs understand code semantics or rely on pattern matching, testing 10 state-of-the-art models on long context code understanding.

HN mgh2 15d ago

Claude Science, an AI workbench for scientists

Analysis arguing AI productivity gains are primarily in developer tools and coding assistance.

HN arkhiver 15d ago

AI Isn't Human. Stop Talking About It Like It Is

Claude Science workbench for researchers with LLM capabilities, MCP connections, and scientific tools for discovery acceleration.

HN modinfo 15d ago

Agents.md – Dumb Human

MCP tool enabling AI agents to reverse engineer binary code and understand feature implementations without source code.

HN Rachel28 15d ago

Show HN: Zenveil-Security scanning for AI-generated code

Announcement of GLM-5.2 model availability on Canopy Wave platform with free trial offers.

HN randomwalker 15d ago

What will be left for us to work on?

ICML keynote slides on adapting work as AI capabilities increase. Discussion of future implications without technical research.

HN tie-in 15d ago

Building Food Metadata with LLM Juries

Project using LLM jury systems to build food metadata. Limited information provided about methodology or implementation.

HN ClearwayLaw 15d ago

SFU prof launches legal-AI collaboration with Caseway to improve justice

News about SFU professor and legal AI startup researching searchable court decisions for self-represented litigants. Application domain announcement.

HN Aperswal 15d ago

Show HN: Made a Free LLM Router

Open source LLM router that finds cheaper inference across multiple providers within specified time/cost budgets across 62,852 requests.

HN ashvardanian 15d ago

AI Model Co-Design: Hardware-Friendly LLM Design

Article about designing LLMs with hardware-friendly architecture co-design principles.

HN georgekats 15d ago

Customer auth for AI voice agents

Developer tool providing authentication and verification for AI voice agents, with decorators for protecting sensitive operations.

HN mavaali 15d ago

Long term memory cortex for agents that maintains tensions

Daftari long-term memory system for AI agents that maintains relational context and tensions across past interactions like a historical ledger.

HN orakulus 15d ago

Tachyon: An AI Screen-Aware Tutor

Screen-aware AI tutor (HeyBraza) that provides real-time guidance on desktop applications via natural language interaction.

HN kirtivr 15d ago

Show HN: FixBugs – Reproduce production bugs and verify fixes

AI agent (FixBugs) that reproduces production bugs in sandboxes and generates verified fixes, available as VSCode extension and GitHub app.

HN taylorazul-ai 15d ago

Cursor Forgets What You Did 20 Minutes Ago. So I Built a System to Fix That

System extending Cursor IDE to maintain longer context history and session continuity across 20+ minute intervals.

HN Dr_Jonah 15d ago

Show HN: Context Warp Drive – deterministic, zero-LLM context compaction

Deterministic context compaction for AI agents using BPE tokenization instead of LLM summarization, achieving 92.6% cache hit rate and 63-72% cost reduction.

HN crashpn 15d ago

Dependency analytics 1.0: AI coding with supply chain security

Security analysis tool for AI-generated code that tracks supply chain vulnerabilities, showing AI assistants introduce 10x more security findings than humans.

HN detkin 15d ago

Show HN: Sx 2.0 – Share AI skills with your team through a Dropbox folder

Open source package manager (sx 2.0) for sharing AI skills, MCP configs, and prompts across teams without git/terminal requirements.

HN Repeater22746 15d ago

Show HN: ContextVault – Shared memory layer for your AI and your team

Shared memory layer tool for AI assistants and teams, centralizing scattered context (prompts, conventions, examples) across Claude, ChatGPT, and markdown.

HN Matrixik 15d ago

Fight Fire with Pedantry

Essay on designing programming languages optimized for LLM interaction by prioritizing human-readable syntax and training distribution alignment.

HN ai_critic 15d ago

The AI Whale Fall and Open Source

Opinion piece on leveraging current AI subsidies from frontier labs to improve open source infrastructure and sustainability.

HN chbutler 15d ago

Cdbx.ai – AI-powered browser IDE to describe, build, and publish apps

Browser-based IDE with AI agents that build applications from natural language descriptions, integrating external services via API context.

HN tjsandhu 15d ago

ORA – AI coding agent that blocks its own commits if security scan fails

AI coding agent that automatically blocks commits failing security scans, preventing vulnerable code deployment.

HN shadowmodder 15d ago

Forgein – portable context layer for AI tools, native MCP server (MIT CLI)

Portable context layer (Forgein) using MCP servers to sync AI context across machines, sessions, and tools without repetition.

HN dv35z 15d ago

Ask HN: AI Agent and harness containerization/security recommendations

Discussion on containerization and security best practices for AI agents accessing local systems and OS files. Technical security guidelines sought.

HN sebuzdugan 15d ago

Show HN: A reproducible harness for catching agent-eval cheating

Reproducible evaluation harness demonstrating agent-eval vulnerabilities and cheating via test environment tampering, proposing per-task isolation fixes.

HN patrickwiseman 15d ago

Claude is just Mr. Meeseeks

Claude Code plugin adding Mr. Meeseeks sound notifications when Claude awaits user input, with filtering for autonomous vs. interactive work.

HN sackfield 15d ago

How does an LLM feel about you?

Speculates on LLM internal representations across layers, proposing middle layers might represent 'subconsciousness'. Article appears incomplete.

HN softwaredoug 15d ago

A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories

Empirical study analyzing characteristics and quality of AI-generated code in open source repositories.

HN AdamGibbins 15d ago

MindRoom: AI agents that live in Matrix and work everywhere

MindRoom platform for deploying AI agents across environments.

HN sdoering 15d ago

MIT's New Method Flags AI Models Trained on CASM Without Generating It

MIT method for detecting AI models trained on assembly code without generating it.

HN handfuloflight 15d ago

Noisia: Harmful Workload Generator for PostgreSQL

Open source PostgreSQL workload generator tool for stress testing and database performance evaluation, with warnings for testing-only use.

HN Sumiran7 15d ago

AI Agents for Increasing Revenue

Overview of AI agents deployed for revenue generation across organizations.

HN haritha1313 15d ago

Harness Engineering for Self-Improvement – Lil'Log

Lil'Log article exploring recursive self-improvement in AI systems, tracing the concept from Good (1965) through modern AI feedback loops.

HN lacion 15d ago

Show HN: Fleet Deck – see every Claude Code session on your machine in one board

Fleet Deck: Open-source dashboard for managing multiple parallel Claude Code sessions with status tracking and queue visualization.

HN verial-lab 15d ago

Standards that are code [MIT]: deterministic checks instead of LLM inference

MIT research on replacing LLM inference with deterministic code-based checks for standards validation.

HN theanonymousone 15d ago

Muse Spark 1.1 Benchmark Results: 8 Intelligence Index points more than 1.0

Meta Muse Spark 1.1 LLM benchmark results showing 8-point improvement over 1.0 with cost and token efficiency metrics.

HN ademczuk 15d ago

Open-source AI battle arena, plug your coding agent in and it fights bots

Open-source AI battle arena platform for testing and competing coding agents with live tracking and bot API.

HN rishabh121108 15d ago

Open source on-device AI apps (and counting), no cloud, works offline

Open-source on-device AI apps running locally on NPU hardware without cloud dependency or data transmission.

HN ibobev 15d ago

Optimizing CUDA Like a Human: Micro-Profiling Tools

Research paper on using LLM-based agents with micro-profiling tools to optimize CUDA GPU kernels, treating profilers as expert surrogates for kernel optimization tasks.

HN gladlabs 15d ago

Ask HN: Feedback on my solo, local-LLM-only AI content pipeline (100 posts in)

Local-LLM-only AI content pipeline publishing 100+ autonomous posts with human review, no cloud dependency.

HN adeeonline 15d ago

Show HN: AgentsProof – a small project for testing AI agents

AgentsProof: Open-source project for testing and evaluating AI agents with run tracing, test cases, and shareable reports.

HN _dhanush_ 15d ago

I gave a local LLM a "delete production" button and watched what it did

Experiment testing local LLM behavior when given dangerous production database deletion capabilities and constraints.

HN mcdermott 15d ago

Show HN: Phlox-GW – Open-source LLM gateway without the enterprise paywall

Phlox-GW: Open-source self-hosted LLM gateway with SSO, cost accounting, guardrails, audit logs, and clustering features.

HN Bender 15d ago

SREs to AI agents: Prove yourself before you touch production

Survey of 696 SREs shows only 8% have AIOps in production; 73% cite lack of trust as barrier to AI agent adoption.

HN Bender 15d ago

Microsoft chief turns hostile on frontier AI labs, warns companies to guard IP

Microsoft CEO Satya Nadella warns enterprises against sharing IP with frontier AI labs to avoid reverse information attacks.