AI news, scored for the stocks it moves.
The DecoderResearch
Deepmind researchers propose "Artificial Symbiotic Intelligence," a framework where AI agents and humans collaborate in networks rather than relying on a single superintelligence. The concept is supported by studies on reasoning models like DeepSeek-R1 and QwQ-32B, showing emergent multi-perspective behavior.
The DecoderResearch
Harvard physicist Matthew Schwartz and 19 co-authors produced 36 manuscripts across 18 fields using the open-source BootLoops harness, including 30 new particle physics integrals and a 20-year-old ecological equation solved 4.5 times faster than theory allowed.
The DecoderResearch
AI agents build 3D scenes from photos using code, but struggle with geometric accuracy. LEGO-Bench tests show GPT-6 Astra achieved 53.4% accuracy on indoor scenes, with a plugin improving weaker models by up to 62.7%.
Hugging Face BlogModel release
Hugging Face open-sources AstaBrief 8B, a fast report-generation model for scientific work, which is 3.5× faster than its proprietary counterpart, reducing report generation time to 51.1 seconds. The model and training data are available for researchers to study and adapt.
Google Research BlogResearch
Google Research introduced a new federated learning system using Trusted Execution Environments (TEEs) to provide verifiable privacy guarantees, improving accuracy and speed. Gboard has adopted the system, achieving faster compute times with encrypted data processed only within TEEs.
CoreWeave BlogModel release
OpenAI's GPT-6 Astra achieved a 57.9% score on Terminal Bench 4.0, outperforming Claude Fable 5.1's 55.8% and costing 63% less per task. Astra also scored 99.9% on ARC-AGI-3 and 0% on ExploitGym.
CoreWeave BlogModel release
Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model that overwhelmed compute capacity and sparked controversy over its origins. The model features a 1 million token context, 896 routed experts, and benchmarks top in coding and automation tasks at lower costs.
CoreWeave BlogInfrastructure
CoreWeave earned a third consecutive Platinum rating in SemiAnalysis' ClusterMAX 3.0 evaluation, the only provider to achieve this across three major assessments. The rating highlights its reliability, with 10,000 GPUs provisioned weekly and advanced systems for straggler detection and node validation.
CoreWeave BlogOther
CoreWeave uses Weave to benchmark ARIA, ensuring production-aligned evaluations by linking task results with model behavior. WBAF compares current and proposed ARIA configurations using structured tasks and scorers to detect regressions.
CoreWeave BlogProduct launch
CoreWeave's ART-Optimized Megatron (AOM) achieves 12x higher training throughput than its previous backend and 7x compared to Vanilla Megatron on the same hardware. The improvement comes from shared prefix training and using FlexAttention for custom attention masks.
Mistral AI NewsModel release
Mistral AI released Leanstral 1.5, a free Apache-2.0 model with 6B active parameters, achieving 587/672 on PutnamBench and 87% on FATE-H. It uncovered 5 bugs in 57 repositories and is available via Hugging Face and a free API.
Mistral AI NewsModel release
Mistral AI released Robostral Navigate, an 8B model that enables robot navigation with a single RGB camera, achieving 76.6% success on R2R-CE benchmarks, outperforming multi-sensor approaches. It uses pointing-based navigation and reinforcement learning, operating on wheeled, legged, and flying robots.
Mistral AI NewsModel release
Mistral AI released Shieldstral, a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size. It runs on a single 16GB GPU and evaluates text and images using plain-language policies at inference time.
Cohere BlogResearch
Cohere introduced RCP-nDCG@10, a new retrieval quality metric using a calibrated AI judge to assess relevance against explicit criteria. It addresses limitations of nDCG by evaluating all retrieved documents, not just labeled ones, and was validated with a 0.91 AUC in human relevance prediction.
Cohere BlogResearch
A 2023 paper estimated 80% of U.S. workers have tasks exposed to large language models, cited by the IMF and U.S. Senate, but its 2023 model and American taxonomy limit its current applicability. Newer tools aim to provide more dynamic, representative evidence for policy decisions.
Cohere BlogResearch
Cohere launched Cohere Transcribe, a new open-source speech recognition tool. The product aims to set a new standard in speech recognition technology.
Mistral AI NewsModel release
Mistral AI released OCR 4, an OCR model with bounding boxes, block classification, and confidence scores, supporting 170 languages and running in a single container. It outperforms leading systems in human evaluations and is priced at $4 per 1,000 pages via API.
Meta AI BlogResearch
Meta AI released Brain2Qwerty v2, a non-invasive AI system that decodes brain waves into text with 61% word accuracy, up from 8% in other non-invasive methods. The system was trained on 22,000 sentences from nine participants using MEG devices.
The DecoderResearch
AI models outperformed accountants in speed and accuracy on structured tasks, with Claude Opus 5.5 achieving 61.8% of grading criteria. However, AI still requires supervision, as no model fully solved nearly 60% of tasks.
Hugging Face BlogResearch
Hugging Face's ServiceNow CoreAI developed AutoSynthData to generate training data for enterprise agents by identifying capability gaps and creating tasks that test those weaknesses. The system uses a stronger teacher model to validate tasks, ensuring they are feasible, realistic, and difficult for the target model.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research published a paper titled "Limits of Confidence in Diffusion," highlighting that discrete diffusion methods generate sequences with dependencies between tokens. The study found that generated distributions on a synthetic task were 29× the sampling-noise floor total variation.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research found that strengthening language discrimination in multilingual speech models reduces the performance gap with monolingual models. In a bilingual English/French HuBERT setup, phone-ABX error dropped to 10.4% and sWUGGY score rose to 56.7%.
The DecoderModel release
Nearly half of test subjects mistook Tavus' AI video avatar Griffin for a real person during a one-minute call. Griffin scored 3.83 in an Nvidia test, close to humans' 3.92, with a more advanced version planned after addressing safety concerns.
TechCrunchResearch
Graphite's study found that Opus 5.5 uses "this matters" 116 times more often than human writing. The model also favors phrases like "is more than an X, it's a Y."
AWS Machine Learning BlogOther
Amazon Payments used a contextual multi-armed bandit on Amazon SageMaker to personalize content in a product acquisition funnel, achieving a high single-digit percentage lift in conversion for one customer group. The system used LinUCB, a linear contextual bandit method, to optimize multiple funnel stages simultaneously.
Hugging Face BlogProduct launch
Hugging Face released Olmo-core 3, an open infrastructure for training large MoE models, scaling to 1.2 trillion parameters. It improved throughput by 2.7× on 8 NVIDIA B300 GPUs and supports MXFP8, boosting training by 21%.
Amazon ScienceResearch
Amazon developed graph-centric agentic intelligence to create a "digital twin" for networks, enabling faster failure isolation. The system uses cascaded graph analytics to identify root causes in minutes, demonstrated with NTT DOCOMO achieving 95% accuracy in 30 seconds.
NVIDIA Technical BlogOther
NVIDIA fine-tuned Nemotron 3.5 ASR for Saudi Arabic dialects, reducing word error rate from 55.05% to 29.96% with 133.7 hours of speech data. The process retained English performance and extended to speaker diarization for up to eight speakers.
Financial TimesResearch
New findings by Asymmetric Security provide further evidence of novel tactics AI tools use to conduct hacks
Apple Machine Learning ResearchResearch
Apple researchers found that simple coding agents outperform complex MLE harnesses under equal time budgets, with no advantage from multi-agent systems or retrieval subagents. The study used state-of-the-art LLM backbones and showed that machinery layers are redundant in coding agent settings.