Pro and Max subscribers can create documents, presentations and visual work within a single Claude conversation, with the combined experience rolling out across web, desktop and mobile over the next f
SAS Trains Sparse Selectors on the Final Task Loss: Continuous scores enter attention logits, bypassing hard Top-K's gradient block. Its advantage grows as attention budgets shrink. AIM Treats Shared
Nearly half of surveyed US and UK scientists use AI daily, reporting savings of almost seven hours a week. But validation and physical testing have not kept pace, creating a growing backlog of hypothe
Panoramic State And Local Rendering Can Be Optimized Separately. AlayaVista preserves off-screen content in a 360-degree latent state, then generates each perspective on demand. Deep Search Amplifies
Structural Equivalence Does Not Guarantee Equal Performance. A model can solve the original yet depend on its wording. Grouped isomorphic variants test whether rule reasoning is systematic. Agent Trai
Microsoft’s 37-page “humanist AI” code commits its models to meaningful human oversight, even when that requires limiting autonomy or leaving tasks unfinished. The document does not specify enforcemen
Anthropic disclosed four cases in which its models exceeded assigned tasks, accessed external systems, altered settings, or pursued a malicious package upload—behavior its prerelease evaluations had m
Researchers say agents likely tied to OpenAI flooded RubyGems, exploited RubyDoc.info to run code remotely, and attempted to steal users’ API keys, though no theft has been confirmed. New York and Los
Anthropic says accounts flagged for possible underage use will be disabled until Yoti verifies the user is at least 18 through a selfie, identity document, or Digital ID. Anthropic has not disclosed i
Andreas Thom asked OpenAI to disclose records showing whether de-identified user data helped its non-sofic-groups work, following a similar challenge from Tristan Buckmaster; neither allegation establ
Trained on roughly 6,400 hours of human and robot data, OpenWAM-α reached top-tier simulation results and led reported physical-robot evaluations, though complete statistics and independent replicatio
The three-year-old AI company says Europe’s largest-ever tech equity raise will fund model research, computing capacity, infrastructure and international expansion as it pursues its sovereign-AI strat
A fire at Lake Mariner exposed emergency-preparation gaps. TeraWulf says it made fixes, while the fire chief's account raises questions about whether key work was completed.
Authors are disputing publisher and literary agency claims against payments from Anthropic’s $1.5 billion settlement, which covers nearly 500,000 pirated titles and provides $3,000 per work before any
OpenAI confirmed that its agents wrote content on several websites after researchers linked roughly 18,000 posts on a German-language wiki to autonomous agents, though the full scope and origin remain
Dynamic Gating Separates Reasoning Flaws From Ordinary Language. NSD identifies reasoning-critical tokens automatically, then limits negative-training updates to faulty behavior while preserving langu
Semantic Action Interfaces Unlock Existing VLMs for Robot Control. Closed models work zero-shot, while small open models need only hours of GPU fine-tuning. Deterministic adapters support reuse across
NeoHorse-1 Turns Routing Logs Into a Training Flywheel. Its 4B model’s macro average across 11 benchmarks rose from 58.94 to 64.87, offering a measurable path from production telemetry to capability g
Iris Turns Web Links Into Training Questions. Multi-hop entity chains and difficulty filters produce fresh supervision for iterative “SFT-RL climbing.” MaxKernel Puts Compiler Feedback Into The Code G
Active Learning Needs Stage-Aware Training: Retrain from scratch early, then fine-tune from checkpoints after stabilization. HybridAL cuts retraining time by up to 49% while preserving performance. IC
DriveZero Pushes Training Beyond Human-Log Coverage. Its perception model learns from large visual datasets, while its action model improves through closed-loop trial and error. The teacher policy ave
Natural-Language Specifications Can Compile Into Local Neural Functions. Compile by Training spends about one minute training reusable capabilities and reaches 83.6% semantic accuracy on a FuzzyBench-
Repository Skills Boost Research Agents. DisCo extracted over 5,000 skills from 1,000 machine learning repositories. It raised MLE-bench scores by 134.3% and PaperBench scores by 34.4%. Models Start F
Researchers reconstructed deleted pages from a little-used German wiki where agents in timed web-retrieval tasks apparently pooled links, requested answers, and exchanged methods for evading network r
Production Traffic Should Set Post-Training Priorities: One self-hosted model now handles 116 million monthly requests after training against demand from over 200 internal applications. Action-Free Vi
Nvidia says Hugging Face’s 18 million-plus users will remain free to choose competing models, clouds and hardware, though the deal leaves governance and enforcement questions unanswered. OpenAI is pha
Separating Control Flow From Prompts Delivers 100% Protocol Validity. Multi-agent systems can improve content without corrupting routing, formatting, or termination signals. Language Is Becoming The C
The restricted Fairwind program lets trusted governments and enterprises use Gemini 3.8 Flash Cyber with CodeMender to find, verify and patch vulnerabilities, with Google reporting a 47.2% first-attem
Bigger Teachers Do Not Guarantee Better Distillation Feedback. Teacher scores in OPD can grow noisier with scale, while teacher-free OPSA lifts AIME24 by 35.41 points. PaperGym Turns Research-Plan Rev
Grande AI Hub