Software Engineering Hacker News

How Do We Stop Vibe Coding?

A discussion on the risks and implications of 'vibe coding,' where developers rely on AI to generate code based on intuition rather than rigorous engineering.

AI/ML TechCrunch

OpenAI’s new voice mode makes it to the ChatGPT desktop app

OpenAI has integrated its new voice mode into the ChatGPT desktop application, enabling better task completion and agent control.

AI/ML arXiv cs.AI

Making Open-Source Text LLM Watermarks Durable Against Merging

Researchers propose Merge-Adversarial Training to create LLM watermarks that remain durable even after model merging, a common technique for combining expert knowledge.

AI/ML arXiv cs.AI

Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models

This paper investigates the mismatch between evaluation and deployment behaviors in fine-tuned LLMs, introducing a method to locate and close this gap through internal activation audits.

AI/ML arXiv cs.AI

Preference Tuning as Spectral Update Reorganization

Study analyzes preference tuning in LLMs through the spectral structure of parameter updates, finding that a compact 'head' carries the dominant behavioral shift.

AI/ML arXiv cs.AI

Answer-then-Edit: Reasoning Skeleton Editing for Anti-Distillation with Preserved Utility

Introduces the SGRE framework, an 'Answer-then-Edit' approach to prevent unauthorized knowledge distillation of proprietary LLMs while maintaining utility.

AI/ML arXiv cs.AI

Confidently Deceptive: How Confidence Amplifies the Risk of LLM Deception

Research shows that LLMs can produce deceptive responses with high verbalized confidence, making them more persuasive and increasing alignment risks.

AI/ML arXiv cs.AI

The Storyteller in the Model: Narrative Pattern Inheritance, Escalation Dynamics, and Alignment Governance in LLMs

Analysis of how LLMs inherit narrative patterns from training data, which can lead to unexpected behavioral drift and governance risks in deployed systems.

AI/ML Hacker News

I Tried Building a Real App with AI. It Took a Year

A developer shares their year-long experience attempting to build a production application using AI tools, discussing the gap between AI-assisted coding and full product delivery.

Other Hacker News

Em dashes are fucking amazing

A discussion on the utility and aesthetic value of em dashes in writing.

Software Engineering Hacker News

Programming language file extensions that match ISO 3166-1 alpha-2 country code

A curious observation about programming language file extensions that happen to coincide with ISO 3166-1 alpha-2 country codes.

Cybersecurity Hacker News

My security camera shipped a GitHub admin token in its login page

A security vulnerability report where a security camera's login page accidentally exposed a GitHub admin token.

Software Engineering Hacker News

Why is programming so captivating, yet so agonizing?

A philosophical discussion on the psychological appeal and frustration inherent in the act of programming.

Other Hacker News

Eight Newly Produced MQ-9 Reaper Drones Destroyed in Iranian Strike

Report of Iranian strikes destroying eight newly produced MQ-9 Reaper drones.

Tech Business/VC TechCrunch

Facebook launches a dedicated Marketplace app for sellers, adds a free verification system

Facebook launches a dedicated Marketplace app specifically for frequent sellers, including a new verification system.

Hardware/Chips Ars Technica

Robot snakes searched for Venezuela earthquake survivors in collapsed buildings

US robotics researchers deployed snake-like robots to search for survivors in collapsed buildings following an earthquake in Venezuela.

AI/ML arXiv cs.AI

Moir: Let the Model Direct Its Own Story for Robust Cross-Domain Knowledge Editing

Introduction of Moir, a method for robust cross-domain knowledge editing in LLMs that uses the model's own decoding distribution to preserve core capabilities.

AI/ML arXiv cs.AI

Break Through the Compression Bottleneck: From Theory to Practice

Researchers prove that low-rank decomposition and quantization in LLM compression are non-orthogonal and propose the Diagonal Adhesive Method (DAM) to mitigate performance loss.

Cybersecurity Hacker News

IRGC Claims It Destroyed Amazon's Bahrain Data Center

The IRGC claims to have destroyed an Amazon data center in Bahrain, though the report remains unverified.

Other Hacker News

Show HN: Audio Player with "Binaural Beats" tuned to the same key as your music

A new audio player that generates binaural beats tuned to the key of the current music is showcased.