Yahoo Mail's AI starts hallucinating fake sneaker win emails, the Unsloth Efficient GRPO algorithm slashes VRAM use by 90% for long-context models, and DeepSeek-ai opens its AGI exploration with 5 public repositories in a week.
Magma sets the bar for multimodal AI with unparalleled learning capabilities, DeepSeek's open-source journey unveils AI infrastructure innovations, and WonderHuman crafts 3D avatars from single-view videos for lifelike realism.
Meta unveils LlamaCon, while Microsoft introduces the groundbreaking Majorana 1 quantum chip, and Deep Lake emerges to enhance deep learning applications with non-tabular data handling.
HP acquires Humane's AI software for $116 million, Valve releases Team Fortress 2 code, and the Deep Lake project tackles big data challenges in deep learning.
Kindle devices can now be jailbroken with WinterBreak, Open-CUAK makes managing automation agents open-source at scale, and the EnigmaEval benchmark reveals major limitations in long multimodal reasoning by state-of-the-art models.
Google scraps AI pledges as expressive movements enhance robot design, while ZeroBench challenges LMM visual understanding and DIY GPT-style models become accessible with a new notebook.
The IRS is purchasing a $7 million AI supercomputer from Nvidia for fraud detection, Distributed-Llama connects home devices into a cluster for LLM inference, and research challenges the necessity of hierarchy in HNSW for high-dimensional data.
AI models halt tech adoption with outdated training data, "Matryoshka Quantization" boosts model precision efficiency, and AI operators on Mac perform tasks via OpenAI and GEMINI APIs.
Larry Ellison proposes an AI-analyzed national database including DNA, OpenAI's new GPT-5 model move integrates different intelligence levels for users, and DockerShrink launches an AI tool to reduce Docker image sizes.
A YouTube vulnerability exposing user emails for $10k raises security concerns, Mixture-of-Agents outperforms GPT-4 Omni with 65.1% on AlpacaEval 2.0, and ANEMLL offers a solution to run LLMs on Apple's Neural Engine.
Thomson Reuters prevails in a landmark AI copyright case, Gemini Cursor introduces an AI-driven multimodal desktop tool, and research warns that AI systems like Meta's Llama have crossed a self-replication threshold.
France launches a massive 109 billion euro AI investment, Meta's Llama and Alibaba's Qwen have achieved self-replication, and the Terminatrix RPG showcases a full adventure in under 100 lines of code.
Intel's AI ambitions stumble as it halts Habana's Falcon Shores, Meta's Llama and Alibaba's Qwen reach the self-replication red line, and Klarity tool debugs AI reasoning with dual entropy analysis.
Amazon's revamped Alexa struggles amid structural and technological challenges, a new multimodal LLM enhances mobile UI searches, and FoloUp offers an AI-driven voice interview platform as open-source for hiring processes.
Meta faces copyright infringement charges for torrenting 81.7 TB of data; HippoRAG outshines existing methods with up to 20% improved performance in LLMs; Upsonic revolutionizes AI tasks with an enterprise-ready framework.
Gumroad ceases hiring junior developers due to AI advancements, Meta AI improves LLM reasoning with Latent Tokens, and "Former" offers AI-enhanced SQL editing for easier database management.
Google's AI ethics policy shift allows weapon and surveillance applications, SmolLM2 advances small language model performance with 1.7 billion parameters, and researchers leverage LLMs to extract characters and relationships from entire books.
Google's removal of its AI weapons pledge signals a strategic shift, Modal Labs' DoppelBot mimics Slack users for seamless communication, and Harmonic Loss enhances neural networks with improved interpretability and faster convergence.
EU bans AI with 'unacceptable risk,' OpenEuroLLM pushes for transparent AI in Europe, and Klarity offers insights into LLM output uncertainty.
Gravy Analytics' geolocation data leak exposes invasive tracking, AMD reveals how to run DeepSeek R1 on RyzenAI for enhanced reasoning, and DeltaLLM's method compresses models to 60% of original size while retaining performance.
AI transcribes audio to tasks with TalkNotes, Gradual Disempowerment warns of AI-induced existential risks, and OmiAI merges models for seamless multimodal output.
Google users can bypass AI summaries with a profane trick, "Tulu 3" achieves new heights in open model post-training, and LLPlayer revolutionizes language learning with AI-generated subtitles.
DeepSeek emerges as a competitive open-weight model despite minimal training costs, while novel FP4 quantization improves LLM training efficiency, and Emmetify drastically cuts LLM agent costs with HTML compression.
Nvidia faces a historic $589B loss due to DeepSeek, while Janus Pro advances text-to-image generation capabilities and researchers unveil uSpectre, a new class of microcode-based transient execution attacks.
Strong math reasoning emerges in small models via simple RL, scientists exploit microcode for data leaks with uSpectre, and LLM Reasoner refines step-by-step reasoning for AI.
Read
