During federal testimony, Elon Musk admitted that xAI partially distilled OpenAI models and expressed regret over his early $38 million donation to OpenAI, calling himself "a fool." He also revealed a $97.4 billion Musk-led bid for OpenAI's assets, indicating significant ongoing legal and financial drama.
Sources: The Neuron
Microsoft, Google, Meta, and Amazon collectively spent approximately $130 billion on AI infrastructure in Q1 2026, nearly doubling Q1 2025 spending, as demand for AI services outstrips their current capacity. Google Cloud, in particular, reported that revenue would have been higher if they could meet demand, while Microsoft's AI business hit a $37B annual run rate and AWS's custom chip business passed a $20B run rate.
Sources: The Neuron
Versions 2.6.2 and 2.6.3 of the 'lightning' deep learning framework on PyPI were compromised in a supply chain attack, injecting obfuscated JavaScript that steals credentials, authentication tokens, and cloud secrets, and propagates through npm by infecting other packages. This incident significantly impacts AI/ML developers using the framework for tasks like LLM fine-tuning and diffusion models, requiring immediate remediation and credential rotation.
Sources: Hacker News
OpenAI's GPT-5.5 has demonstrated top-tier performance in long-horizon cyber tasks, becoming the second model to complete multi-step cyber-attack simulations end-to-end, showing rough parity with Claude Mythos Preview. This materially changes the earlier narrative that Anthropic had a unique lead in offensive cyber automation, and OpenAI also paired this with a product-side security release: Advanced Account Security for ChatGPT.
Sources: Latent Space
Sam Altman and Noam Brown highlight AI inference compute as a critical, undervalued resource, with Intel's CEO noting rising CPU demand for AI. NVIDIA's Jensen Huang declares the 'inference inflection' has arrived, with AI compute demand increasing by a million times in two years, leading to a positive flywheel for AI development.
Sources: Latent Space
AutoRound has launched an advanced quantization toolkit for Large Language Models and Vision-Language Models, enabling high accuracy at ultra-low bit widths (2-4 bits, FP8, INT2-mixed) with minimal tuning. This toolkit, detailed in the SignRoundV2 paper, significantly enhances LLM deployment efficiency and has been widely integrated into major AI frameworks including Hugging Face Transformers, vLLM, LLM-Compressor, and SGLang. Its effectiveness is highlighted by a DeepSeek-R1 model retaining 97.9% accuracy using AutoRound's INT2-mixed quantization.
Sources: Hacker News
Hugging Face reports that the cost of evaluating advanced AI models, particularly agentic and training-in-the-loop benchmarks, has escalated significantly, often surpassing training costs. This surge, evidenced by benchmarks like HAL and OpenAI's MLE-Bench costing thousands to tens of thousands per evaluation, is driven by evaluation complexity, limited compression effectiveness, and the need for multiple reruns due to agent unreliability, creating a substantial barrier for AI development and accountability.
Sources: Hugging Face Blog
AI company Anthropic acquired the Bun JavaScript runtime in December 2025, a prominent project known for its performance and use of AI assistance, indicating Anthropic's potential interest in integrating AI capabilities into developer tools or infrastructure.
Sources: Simon Willison
Amazon Web Services (AWS) has released a comprehensive guide on Reinforcement Learning with AI Feedback (RLAIF), also known as LLM-as-a-judge, positioning it as the leading method for efficient LLM alignment. The guide details six steps for implementing reward functions, including selecting judge architectures, configuring models via Amazon Bedrock (recommending specific LLMs like Nova and Claude), and building resilient systems, with a practical application shown in automated legal contract review.
Sources: AWS ML Blog
Fintech company Sun Finance collaborated with the AWS Generative AI Innovation Center to develop and deploy a new AI-powered identity verification (IDV) and fraud detection pipeline. This partnership resulted in significant improvements in accuracy and processing efficiency for loan applications.
Sources: AWS ML Blog
Amazon Quick now features an agentic conversational AI assistant, enabling business users to perform self-service data analytics and access business intelligence insights through intuitive natural language queries. This capability is enhanced by integrated knowledge bases for contextual awareness and foundational improvements like native Apache Iceberg support in Amazon S3 Tables, with AWS also releasing a reference architecture for building these AI-powered lakehouse solutions.
Sources: AWS ML Blog
Stripe has launched a broad suite of AI-powered products and features designed to support automated commerce and enhance platform security in the AI era. Key innovations include AI wallets for agent-driven purchases, an agentic execution environment, real-time streaming payments, and advanced usage-based billing. Additionally, Stripe significantly bolstered its fraud prevention with AI-powered Radar enhancements and a Managed Risk API to combat new forms of abuse exacerbated by AI.
Sources: Keep Bookmarks
Google CEO Demis Hassabis emphasized the critical need for the West to develop a strong open-source AI stack to maintain competitiveness against China. He also posited that edge models, due to their on-device presence, should be open-source by nature.
Sources: The Neuron
DeepSeek V4 has incorporated Compressed Sparse Attention and Heavily Compressed Attention, significantly reducing KV cache memory by up to 98% on long-context tasks. This architectural advancement improves efficiency for processing extended inputs in LLMs.
Sources: The Neuron
Anthropic has made its models available in public beta, expanding access to its AI capabilities. Concurrently, OpenAI has deployed its frontier models to vetted "critical cyber defenders," enhancing cybersecurity capabilities with advanced AI.
Sources: The Neuron