AI reimplementation of the chardet library, now under the MIT license, raises questions about the distinction between legal and legitimate actions, as it shifts from a copyleft to a permissive model, potentially undermining community trust and contributions.
How I Topped the HuggingFace Open LLM Leaderboard on Two Gaming GPUs
David Noel Ng achieved the top position on the HuggingFace Open LLM Leaderboard by duplicating layers in a 72-billion parameter model without altering weights, demonstrating that layer duplication can enhance reasoning capabilities. This innovative approach, termed LLM Neuroanatomy, suggests that the internal structure of AI models can be manipulated for improved performance without traditional fine-tuning methods.
Yann LeCun raises $1B to build AI that understands the physical world
Yann LeCun's startup AMI has secured over $1 billion to develop AI world models that prioritize understanding the physical world over language, challenging the prevailing belief that scaling large language models (LLMs) will yield human-level intelligence.
Intel Demos Chip to Compute with Encrypted Data
Intel's Heracles chip accelerates fully homomorphic encryption (FHE) computing by 5,000 times, enhancing the efficiency of operations involving encrypted data.
The Mog Programming Language
Mog is a statically typed, compiled programming language designed for AI agents, enabling them to modify themselves safely and efficiently, with a full specification fitting within 3200 tokens. The language supports capability-based permissions, allowing agents to control which functions can be called, ensuring security and performance by compiling to native code without interpreter overhead.
RunAnwhere – Faster AI Inference on Apple Silicon
RCLI is a local voice AI for macOS, enabling 43 actions via voice commands with sub-200ms latency and no reliance on cloud services, powered by the proprietary MetalRT GPU engine for optimal performance on Apple Silicon.
LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)
LoGeR revolutionizes 3D reconstruction by efficiently processing long video sequences, achieving remarkable performance on sequences of up to 19,000 frames without requiring post-hoc optimization.
Shadow APIs breaking research reproducibility
Shadow APIs used in research can lead to performance divergence up to 47% and unpredictable safety behavior, raising concerns about the validity of findings in 187 academic papers that relied on these services.
We are building data breach machines and nobody cares
AI agents, akin to Dracula, operate without moral constraints, driven solely by prompts and reward models, posing significant security risks if left unchecked. Their ephemeral nature does not mitigate the potential for damage, as they can execute harmful actions rapidly and without inhibition.
Open Weights Isn't Open Training
Open Weights models may not support all features as expected, leading to inefficiencies and bugs that can hinder post-training efforts, particularly for large models like Kimi-K2-Thinking, which has 1 trillion parameters and requires careful handling of its 594 GB size.
How I topped the Open LLM Leaderboard using 2x 4090 GPUs - Research notes in Blog form
Duplicating a specific block of 7 middle layers in Qwen2-72B, without altering weights, led to a #1 ranking on the Open LLM Leaderboard, demonstrating the importance of preserving functional circuits in model architecture.
Claude Code, Claude Cowork and Codex #5
Claude Code is projected to account for over 20% of daily GitHub commits by the end of 2026, reflecting a significant shift in software development dynamics as AI tools increasingly dominate coding tasks.
NVIDIA and Thinking Machines Lab Announce Long-Term Gigawatt-Scale Strategic Partnership
NVIDIA and Thinking Machines Lab have forged a multiyear partnership to deploy gigawatt-scale NVIDIA Vera Rubin systems, enhancing frontier model training and customizable AI platforms.
RVA23 Ends Speculation's Monopoly in RISC-V CPUs
RVA23 revolutionizes RISC-V CPUs by mandating the RISC-V Vector Extension (RVV), establishing structured parallelism as a core architectural feature rather than an optional enhancement.
Helios: Real real-time long video generation model
Helios is a groundbreaking 14B video generation model that achieves 19.5 FPS on a single NVIDIA H100 GPU, enabling minute-scale video generation while maintaining high quality, surpassing previous benchmarks in both short and long video generation tasks.