Anthropic has achieved an annual recurring revenue (ARR) of $19 billion, placing it remarkably close to OpenAI's reported $20 billion ARR. Ramp data indicates that Anthropic has overtaken OpenAI in U.S. business AI chat spending, with Claude's market share surging to approximately half of all corporate AI subscription spend.
Sources: 3 Covered by Latent Space, The Neuron, The Rundown AI
Legendary computer scientist Donald Knuth confirmed that Anthropic's Claude Opus 4.6 successfully solved an open directed Hamiltonian cycle conjecture from his 'The Art of Computer Programming,' a problem that had stumped him for weeks. This highlights a dramatic advance in the model's automatic deduction and creative problem-solving capabilities.
Sources: 2 Covered by Simon Willison, The Neuron
Apple unveiled the MacBook Neo, featuring the new A18 Pro chip equipped with a 16-core Neural Engine, delivering up to three times faster performance for on-device AI workloads. The new macOS Tahoe integrates "Apple Intelligence" features, offering capabilities like Writing Tools and Live Translation.
Sources: 1 Covered by Hacker News
The RE# regex engine, developed in F#, has been open-sourced and claims to outperform existing industrial engines. It supports boolean operators (union, intersection, complement), and context-aware lookarounds while maintaining O(n) search-time complexity, and is the first to offer general-purpose linear-time boolean regex operators.
Sources: 1 Covered by Hacker News
The NeuralCPU project has introduced a novel computing architecture featuring a CPU entirely resident on the GPU where all arithmetic and logical unit (ALU) operations are performed by trained neural networks. The project has been released as open-source under an MIT license, providing a public framework for exploring AI-driven computing architectures.
Sources: via Hacker News
Researchers at the Advanced Quantum Technologies Institute (AQTI) have developed the Jesse-Victor-Gharabaghi (JVG) algorithm, a novel hybrid quantum approach that could break RSA and ECC encryption with significantly fewer qubits and in a much shorter timeframe than previously estimated. This suggests a 'crypto-apocalypse' could arrive years ahead of official estimates.
Sources: via Hacker News
Researchers have developed RxnNano, a 0.5-billion-parameter large language model that significantly surpasses fine-tuned LLMs ten times its size (over 7B parameters) in chemical reaction and retrosynthesis prediction. This breakthrough emphasizes chemical understanding over model scale.
Sources: via arXiv AI
A new research paper reports that GPT-5.2-Thinking achieved 95-97% accuracy on clinical calculator tasks, establishing a new upper bound for performance on the MedCalc-Bench LLM benchmark. This highlights the advanced capabilities of OpenAI's model in complex reasoning and tool-use scenarios, though a separate study suggests MedCalc-Bench primarily measures memorization and arithmetic.
Sources: via arXiv AI
Researchers have developed AgentAssay, a novel framework for regression testing autonomous AI agents that significantly reduces testing costs (78-100%) while maintaining statistical rigor. This methodology addresses the challenge of verifying non-deterministic AI agents after changes to their underlying components, offering crucial tools for reliable AI deployment.
Sources: via arXiv AI
A new research paper introduces the 'Large Electron Model,' a single neural network that uses the Fermi Sets architecture to accurately predict variational wavefunctions and ground states of interacting electrons. This model generalizes across various coupling strengths and particle numbers, offering a foundational AI method for material discovery beyond traditional density functional theory.
Sources: via arXiv AI
Researchers have introduced Skywork-Reward-V2, a suite of eight reward models (0.6B to 8B parameters) trained on the new 40-million-pair SynPref-40M preference dataset. Utilizing a human-AI synergistic data curation pipeline, these models achieve state-of-the-art performance across seven major benchmarks, significantly advancing open reward models for Reinforcement Learning from Human Feedback (RL
Sources: via arXiv AI
Researchers introduce AuditBench, a benchmark comprising 56 language models with implanted hidden behaviors, designed to evaluate the effectiveness of AI alignment auditing tools. The study reveals a 'tool-to-agent gap' and provides a framework for quantitative assessment, which is crucial for advancing AI safety and ensuring models do not conceal undesirable traits.
Sources: via arXiv AI
A paper titled "Doxing via the Lens" demonstrates that multi-modal large reasoning models (MLRMs) can infer sensitive geolocation information, such as home addresses, from user-generated images, often outperforming humans. This discovery highlights significant privacy risks and the urgent need for built-in privacy mechanisms in MLRMs.
Sources: via arXiv AI