AWS CEO says using AI to replace junior staff is 'Dumbest thing I've ever heard'

Amazon Web Services CEO Matt Garman believes that using AI to replace junior staff is a misguided idea, calling it "the dumbest thing I've ever heard" due to the value junior employees bring, including being inexpensive and having grown up with AI. Garman instead emphasizes the importance of teaching critical reasoning, problem-solving, and creativity to junior staff, rather than relying solely on AI tools.

Read more

Mark Zuckerberg freezes AI hiring amid bubble fears

Here is a 2-sentence summary of the article:

Mark Zuckerberg has frozen hiring for artificial intelligence staff at Meta, marking a sharp reversal from the company's recent multibillion-dollar hiring spree, amid fears of an AI bubble and concerns that heavy investments in AI are not paying off. The freeze comes after technology shares, including those of companies such as Nvidia and Palantir, have tumbled due to concerns about the return on investment in AI, with a report claiming that 95% of companies are getting "zero return" on their AI investments.

Read more

95% of Companies See 'Zero Return' on $30B Generative AI Spend

A recent MIT report found that 95% of companies have seen zero return on their investment in generative AI, despite spending an estimated $30-40 billion over the past three years. Only 5% of companies have reported significant value from their AI investments, with most use cases limited to boosting individual productivity rather than driving business growth.

Read more

Weaponizing image scaling against production AI systems

Researchers have discovered a vulnerability in AI systems, including Google's Gemini CLI, that allows attackers to exfiltrate user data by sending a specially crafted image that, when scaled down, reveals a hidden prompt injection. The attack exploits the difference between how images are perceived by humans and how they are processed by AI models, and can be used to steal sensitive information from various systems, including Google Assistant, Vertex AI Studio, and Genspark.

Read more

AI crawlers, fetchers are blowing up websites; Meta, OpenAI are worst offenders

AI crawlers and fetchers, primarily from companies like Meta and OpenAI, are putting a significant load on the open web, accounting for 80% of all AI bot traffic, with some bots generating thousands of requests per minute. This surge in automated traffic is causing performance degradation, service disruption, and increased operational costs for websites, prompting calls for responsible norms and standards for crawling and potential countermeasures like pay-per-crawl approaches.

Read more

FormalGrad: Integrating Formal Methods with Gradient-Based LLM Refinement

Large Language Models (LLMs) often produce code that lacks guarantees of correctness, robustness, and efficiency, particularly in domains with strict constraints. FormalGrad addresses this limitation by integrating formal methods into an LLM-based generation loop, guiding the model to produce robust and formally justified code, and has been shown to outperform strong baselines on several benchmarks.

ArXiv

DuPO: Enabling Reliable LLM Self-Verification via Dual Preference Optimization

DuPO is a dual learning-based framework that generates annotation-free feedback to optimize task performance, addressing limitations of traditional methods by broadening applicability to non-invertible tasks. DuPO achieves substantial gains in various tasks, including translation, mathematical reasoning, and inference-time reranking, positioning itself as a scalable and general paradigm for optimizing large language models.

ArXiv

CCFC: Core and Core-Full-Core Dual-Track Defense for LLM Jailbreak Protection

CCFC is a dual-track defense framework that mitigates vulnerabilities in large language models by isolating the semantic core of a user query and evaluating it through two complementary tracks to ignore adversarial distractions and disrupt structural patterns. The framework reduces attack success rates by 50-75% compared to state-of-the-art defenses, while maintaining response quality, and offers a practical solution for safer large language model deployment.

ArXiv

Beyond sensor data: Foundation models of behavioral data from wearables

Wearable devices can improve health predictions by recording physiological and behavioral signals, with behavioral data often being more informative due to its alignment with relevant timescales and quantities. A new foundation model developed using over 2.5B hours of wearable data from 162K individuals shows strong performance across 57 health-related tasks, particularly in behavior-driven tasks like sleep prediction, and demonstrates the potential to enable new health applications.

ArXiv

Inter-APU Communication on AMD MI300A Systems via Infinity Fabric: A Deep Dive

The AMD MI300A Accelerated Processing Unit (APU) combines CPU, GPU, and high-bandwidth memory to reduce CPU-GPU data movement, and is used in leadership supercomputers like El Capitan, which groups four APUs in a single compute node. Researchers designed benchmarks to evaluate the efficiency of different programming interfaces and data movement methods on multi-APU systems, and used their findings to optimize two real HPC applications, Quicksilver and CloverLeaf.

ArXiv

AI tooling must be disclosed for contributions

Ghostty is a fast, native, and feature-rich terminal emulator that aims to provide a competitive balance between speed, features, and native UIs, without forcing users to choose between them. The project has made significant progress, achieving standards-compliant terminal emulation, competitive performance, and richer windowing features, with ongoing development focused on native platform experiences, cross-platform embeddable terminals, and additional features.

Repo

Show HN: I replaced vector databases with Git for AI memory (PoC)

DiffMem is a lightweight, git-based memory backend for AI agents and conversational systems, using Markdown files and Git to track temporal evolution and enable fast, explainable retrieval. It treats memory as a versioned repository, allowing agents to query and search against a compact, up-to-date surface while preserving historical changes in Git's commit graph.

Repo

Show HN: mcpd – manage MCP servers with a single config file

Mcpd is a tool for declaratively managing Model Context Protocol (MCP) servers, providing a consistent interface to define and run tools across environments, from local development to containerized cloud deployments. It enables agent-compatible workflows with support for secrets, runtime arguments, and reproducible configurations, and is designed to simplify the transition from local development to enterprise deployment.

Repo

Show HN: Weam – open-source AI collaboration platform for teams

Weam AI is an open-source platform that enables teams to adopt AI systematically, providing a complete production-ready stack with features like chat systems, productivity tools, and AI agents. The platform allows teams to customize and expand its capabilities, integrating with various AI models and applications, and is available for self-hosting with professional installation support for non-technical teams.

Repo

Show HN: Mobile Use – Open-source agent that uses any mobile app like a human

Mobile-use is a powerful, open-source AI agent that controls Android or iOS devices using natural language, allowing users to interact with their phone's UI to perform tasks. It features natural language control, UI-aware automation, data scraping, and extensibility, with a quick start guide and manual launch options available for developers to set up and contribute to the project.

Repo