Sam Altman joined OpenAI, Anthropic, AWS, Google, Microsoft and Oracle in a joint call urging urgent global action on AI-powered cyber defense, warning the window to harden systems is limited.
DeepMind's new model adds 40-second scene extension, first/last frame control and 3-second video references at up to 4K, and is live across the Gemini API, AI Studio, Cloud, Flow and the Gemini app — ranked #1 on the Text-to-Video Arena.
After a $175M debut backing Harvey, Cursor and others, OpenAI is readying a $400M fund for application-layer AI startups, per The Wall Street Journal.
A two-stage SFT data selection method (trajectory screening, then segment-level loss weighting) beats full-data training using just the top 10% of trajectories, with relative gains up to 12.2% on SWE-Bench Pro and 24.2% on SWE-Bench Verified.
The first video encoder trained under LeJEPA's collapse-free objective uses a single encoder plus projector and one hyperparameter, matching or beating V-JEPA 2 at matched epochs with 5.6-20.8x less compute and topping the strongest video baseline by 7.6 points on ImageNet-1K at matched FLOPs.
Cross-embodiment world models trained on internet-scale heterogeneous video — reconciled via end-effector poses, language and latent actions — approach or beat single-embodiment SOTA on DROID, with code and models open-sourced.
A theoretical and empirical study shows ES yields broader reasoning coverage and better Pass@K than GRPO, whose entropy collapses in training, then proposes a sequential GRPO-ES hybrid that combines GRPO's Pass@1 strength with ES's Pass@K gains.
Distills matching rollouts while penalizing disagreeing ones via grouped RL, letting TTPO match label-supervised OPSD on five competition benchmarks without any labels and lifting Qwen3-1.7B from 38.0% to 45.2% in test-time training.
Cache-DiT, step reuse, sparse attention and fused kernels deliver a ~1.95x lossless MiniMax-H3 speedup — and up to 6.24x with quality trade-offs — on 8xH200 GPUs.
portal.nousresearch.com now consolidates active free models, promos and a sortable list of all available models in one place.
The agent can now use a managed copy of an existing Chrome profile — logins and cookies included — to browse and act as the user.
The OpenAI-backed humanoid maker is in talks with SoftBank at roughly $6B — about 40% below 1X's earlier $10B target — which would give SoftBank control of the home-oriented NEO humanoid line.
Commerce's Bureau of Industry and Security is examining 47 shipments from a Kuehne+Nagel freight unit that allegedly routed Super Micro servers with Nvidia chips to China via Southeast Asia and Hong Kong; Kuehne+Nagel stock fell up to 4.2%.
The second consumer robot from Hugging Face, after the Reachy Mini, targets developers and educators who want programmable physical AI on a budget.
Each agent gets its own cloud computer and codebase, reachable from Slack, email, iMessage, browser or terminal, with routines that keep running when your laptop is closed.
The Rust-native gateway manages self-hosted, frontier and open models with under 1 ms overhead and no token markup, optionally mining OTel traces to route each request to the optimal model or train a custom one.
Qoder's plan-execute-verify coding platform went GA on August 27, with 500 daily Qwen credits for individuals and 50% off Qwen models for enterprise until September 1.
Neural acoustic modeling turns MIDI into editable modeled performances for strings, woodwinds and brass, running fully locally with no GPU required.
An open-source tool for defining and running test scenarios to evaluate and harden agent behavior before shipping.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
