Aramco Ventures leads $800M round for the AI neocloud powering enterprise training and inference on open models. 1,500+ customers including Cursor and Cognition.
President Lee Jae Myung announced a 911 trillion won plan anchored by Samsung and SK Hynix building 4-6 new fabrication plants, a chip packaging cluster, and 550 trillion won in AI data center investment by 2029.
88 bionic joints, 300+ micro-expressions, emotional LLM with 500ms response — and 13,361 pre-orders confirmed across Lite/Pro/Ultra variants priced from 119,800 to 990,000 RMB.
Meta-benchmark from a systematic review of 3,101 papers covering 78 datasets across 6 creativity domains. Factor analysis reveals a single 'c' factor explaining 81.5% variance across frontier LLMs, separable from general knowledge and reasoning.
Evaluates code patches without Docker or execution using repository exploration for evidence gathering. Outperforms strongest open-source verifier by 14.3 AUC points, enabling environment-free RL post-training that reaches 62.0% on SWE-bench Verified on a 9B model.
LLMs fine-tuned with GRPO for materials science hypothesis generation via conceptual recombination. Achieves 40-65% improvement over base models with ~2-3x greater semantic diversity and explicit causal graph construction.
Mid-training paradigm converting evolutionary search trajectories into supervision over 156K trajectories across 10 domains. Fine-tuned 2B-9B LLMs achieve 10.22% average improvement on 22 held-out tasks, matching SOTA on circle-packing and Erdos minimum-overlap problems with test-time RL.
RL with verifiable rewards that adaptively switches between imitating references and self-improvement per instance. Improves average SRxSim from 0.0959 (GRPO) to 0.1773 on TOMG-Bench MolOpt with significant gains on LogP, MR, and QED.
Nous Research ships major self-learning and improvement capabilities, clears all P0/P1 bugs, and thanks ~300 contributors. Update via `hermes update`.
Together AI reports rapid adoption of NVIDIA's open model, described as fast, efficient, and customizable, gaining community traction over alternatives in days.
New open-source framework for training Deep Research Agents using agentic reinforcement learning. Released with benchmarks and initial results.
NVIDIA makes undisclosed equity investment and signs technical collaboration to deploy Cosmos world foundation models across Verkada's 2.4 million connected security devices at 30,000 organizations. Reports 68% improvement in spatial-temporal search precision.
Elon Musk confirms Optimus production on the former Model S/X line. Initial output slow due to 10,000 unique parts; target cost of goods sold is $20,000 per robot at scale.
Qualcomm expands into AI datacenter silicon, adds acquisition of Modular for a CUDA-compatible software stack. Microsoft and Meta sign multi-generation deals. Also debuts a rack server for AI inferencing.
Prosecutors raid 12 sites including Super Micro Computer offices, expanding from 3 to 9 suspects. Case involves alleged shipment of Nvidia GB300-based AI servers in violation of US export controls. NT$9 million in cash and 50 servers seized.
Strategic funding framework grows from $5B to $25B to finance fuel-cell power projects for AI data centers globally, offering islanded onsite power as alternative to utility grid interconnection.
Anthropic's new Sonnet offers near-Opus-level reasoning, tool use, and coding at lower prices: $2/M input, $10/M output tokens.
Mobile iOS app brings Cursor's AI coding agent to phones. Topped Product Hunt as #1 product of the day.
Desktop app brings real-time topic updates, Google Tasks/Keep integration, with upcoming multi-step task assignment from phone.
Swiss AI lab offers browser-based access with no artificial usage caps, positioning itself as civic AI infrastructure built in Europe.
