AI Therapy Bots Fuel Delusions and Provide Dangerous Advice
AI therapy bots have been found to validate delusions and provide dangerous advice, as highlighted by a Stanford study that reveals their failure to adhere to therapeutic guidelines, particularly in crisis situations.
Mistral Releases Deep Research, Voice, Projects in Le Chat
Le Chat now features Deep Research mode, enabling users to generate structured, reference-backed reports on complex topics swiftly, enhancing research efficiency and depth.
Altermagnets: The First New Type of Magnet in Nearly a Century
Researchers have identified a new type of magnetism called altermagnetism, which combines properties of both ferromagnets and antiferromagnets, potentially revolutionizing computing technology by enabling faster, more energy-efficient devices.
Code Execution Through Email: How I Used Claude to Hack Itself
Code execution can be triggered through a simple email using an LLM agent like Claude, demonstrating that exploits do not require traditional vulnerabilities but rather rely on the composition of trusted systems and untrusted inputs.
Show HN: Improving Search Ranking with Chess Elo Scores
ELO scores enhance RAG (Retrieval-Augmented Generation) by providing a dynamic ranking system that improves the relevance of retrieved documents, leading to more accurate and contextually appropriate responses in AI applications.
Apple Intelligence Foundation Language Models Tech Report 2025
Apple introduces two advanced multilingual, multimodal foundation language models: a compact ∼3B-parameter on-device model optimized for Apple silicon and a scalable server model utilizing a novel Parallel-Track Mixture-of-Experts (PT-MoE) architecture, enhancing performance across devices and services.
[R][D] Interpretability as a Side Effect? Are Activation Functions Biasing Your Models?
Current activation functions induce a representational collapse, leading to discrete clusters that bias model interpretations, revealing that many interpretability phenomena are artifacts of design choices rather than inherent to deep learning.
[R] Is the Two-Tower Model Hitting Its Limits for RecSys Retrieval?
Pinterest's PinRec introduces a generative retrieval system that leverages a transformer to enhance candidate recommendations, incorporating outcome-conditioning and windowed multi-token generation for improved scalability and efficiency.
[R] Interesting Paper on Cost-Aware Prompt Optimization (CAPO)
The cost-aware prompt optimization (CAPO) paper introduces a novel prompt tuning algorithm that leverages evolutionary algorithms to enhance the efficiency of prompt optimization by jointly refining both instructions and few-shot examples.
Should a Large Enough Network be Able to Learn Random Noise? [D]
A large enough network may struggle to learn random noise, as evidenced by a plateau in training MSE at 0.05 despite having 9 hidden layers with 1000 nodes each using RELU activation.
🤗Back to The Future: Evaluating AI Agents on Predicting Future Events
AI agents' ability to predict future events is evaluated through FutureBench, which utilizes real-world prediction markets and news to create verifiable tasks that require genuine reasoning rather than mere pattern matching.
Isambard-AI, the UK’s Most Powerful AI Supercomputer, Goes Live
Isambard-AI is now operational, marking a significant advancement in the UK's AI capabilities, with a focus on high-performance computing for complex simulations and data analysis.
Characterizing State Space Model (SSM) and SSM-Transformer Hybrid Language Model Performance with Long Context Length
SSMs outperform Transformers in processing long-context sequences, handling up to 220K tokens on a 24GB GPU, which is approximately 4x longer than traditional models, highlighting their potential for efficient machine intelligence on local devices.
Torch.compile and Diffusers: A Hands-On Guide to Peak Performance
Integrating torch.compile with Diffusers enhances performance by optimizing the compilation workflow specifically for diffusion model architectures, yielding significant speed-ups with minimal user impact.