Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A researcher applied activation patching to a DNA sequencing basecaller, a deep learning model, to understand its internal mechanisms, finding MLP dominance in early and late layers and concentrated self-attention activi…
A developer explains the sparse Mixture of Experts (MoE) architecture used in models like Mixtral, DeepSeek-MoE, and Grok-1, detailing how the router selects which experts to activate per token and why load-balancing is …
Kim Hatton, Global Financial Services Marketing Leader at Databricks, discusses her career journey from marketing regulated financial products to leading data platform marketing, highlighting how Databricks' Unity Catalo…
A Nature preprint published June 13, 2026, introduces CURE, a graph unlearning framework that uses contrastive representation editing to remove the influence of target nodes, edges, or features from a trained graph neura…
Rudrite Research published a comprehensive comparison of AI and ML approaches, covering 14 side-by-side analyses of techniques such as Transformers vs Mamba, FlashAttention vs PagedAttention, and PPO vs DPO vs GRPO. The …
Researchers led by Damani et al. introduced a method to train language models to express their uncertainty by adding a calibration reward to reinforcement learning from verifiable rewards (RLVR). The approach, detailed i…
A new interactive visual explainer from Rudrite Research demonstrates that large language model hallucinations are a predictable outcome of training and grading systems that reward confident guessing, based on a 2025 arX…
Researchers Liu et al. published a paper on arXiv 2025 introducing ProRL, a method using prolonged reinforcement learning with KL resets to expand reasoning boundaries in AI models. An interactive visual explainer of the…
Zhao et al. published GenPRM, a generative process reward model that reasons and runs code to verify each step, achieving state-of-the-art performance where a 7B parameter model outperforms a 72B parameter model. The pap…
Nvidia's new GB300 NVL72 system achieves 61,400 concurrent AI agents per megawatt, a 20x improvement over the prior-generation H200. The Blackwell Ultra rack-scale system, validated using the AgentPerf benchmark, is alre…
Researchers argue that parametric machine learning systems fail at latent learning—acquiring information not immediately relevant but useful for future tasks—contributing to data inefficiency compared to natural intellig…
A YouTube video titled 'Erdős Problems and Speculations about the Power of AI Models' explores connections between unsolved problems in mathematics and the capabilities of artificial intelligence, discussing how AI might…
A developer argues that error analysis—reading AI outputs and categorizing failures—is the most critical step in building effective evaluations, far more valuable than dashboards or generic metrics. The process involves …
GitHub Copilot CLI released a smarter subagent delegation feature that reduced tool failures per session by 23% and improved user wait times by up to 5% at P95, with no quality regression. The improvement makes the AI ag…
NVIDIA's Blackwell architecture leads the inaugural AgentPerf benchmark for agentic AI. AMD opens pre-orders for its Linux-friendly Ryzen AI Halo developer platform featuring the Strix Halo APU. Linux 7.2 introduces GPU …
LMCache introduces a novel KV cache optimization layer to accelerate LLM inference, enabling faster local deployment on consumer hardware. AllenAI releases olmo-eval, a workbench for evaluating open language models throu…
Nvidia's Blackwell architecture achieves 20 times more AI agents per megawatt than its Hopper generation, according to the new AgentPerf benchmark. The efficiency leap, driven by FP4 precision and advanced NVLink designs…
Flux, a Boston-based engineering intelligence startup, raised $5 million in seed funding led by Calibrate Ventures with participation from True Ventures and Glasswing Ventures. The company will use the capital to acceler…
Zyphra released ZONOS2, an open-weight real-time voice-cloning model under the Apache 2.0 license, including model weights on Hugging Face and inference code on GitHub. The company also introduced ZTTS1 Eval, a new bench…
Nobel laureate Geoffrey Hinton predicts AI will surpass humans in mathematics within 10 years, arguing math is a closed system like chess or Go. Hinton, who resigned from Google in 2023 over AI safety concerns, now estim…