Jensen Huang resigns as NVIDIA CEO

Defense Factory sends findings from security tools to agents that investigate issues and prepare patches for engineers to review. Daybreak Red Team tests systems for vulnerabilities and can feed its findings into the same workflow.

The agent can prepare an order on a merchant's site. The user approves the final total through Stripe Link, with Visa supporting the payment. Users also approve ad campaigns before launch. Purchasing starts with U.S. Core and Pro users, and ads are available to Pro users worldwide.

The open-source library offers composable interaction behaviors such as popups, selection and typeahead. Coding agents can combine the behaviors they need to build components for a particular project.

When a user selects Astryx, v0 builds with the design system's components, patterns and templates in context. Astryx joins several other options listed on the v0 design systems page.

Eligible Team and Enterprise users can add Claude as a visible participant. It takes notes, answers questions and creates artifacts such as documents, whiteboards and slides during the call.

Decider v1.3 and its Flash variant have downloadable weights. The new locate feature returns matching passages and their character positions. It can run alongside classification and scoring questions on the same input.

Perplexity says its browser agent in Computer and teammates now shows its work in progress. Users can watch and take over the browser when needed.

Paid plans can start a cloud iOS Simulator or Android Emulator on demand. Agents can install and drive apps, with recordings, screenshots and logs available for review.

Cursor reports that Gemini 4 Hydrogen scored 54.3% at its default effort on CursorBench, compared with 39.6% for Gemini 3.8 Flash.

The Grok Bot team says users can connect a bot to a personal or company tailnet to reach private MCP servers and internal tools.

Kody describes a hosted setup for connecting multiple providers' accounts and credentials to one MCP service, where team agents can access authorized infrastructure.

Anthropic's developer team shared a video about releasing new API models, adapting the agent harness after an upgrade and measuring cost per completed task.

Mercor published 26 tasks across software, physical engineering, machine learning and scientific computing. It reported an average score of 54.6% for Fable 5.2 and 44.2% for GPT-6 Astra. An electric aircraft design task has its own scoring rules.

LlamaIndex expanded the page-level test set and added forms. The scoring reflects the impact of each error, so a wrong number in a table can carry a larger penalty than a formatting change that preserves the content.

The paper uses a frozen BF16 teacher to guide a quantized student through corrections and uses execution feedback to select revisions. The student is then trained to run independently.

Featured on Hugging Face Papers, the study extracts about 380,000 records from 2,118 university textbooks, preserving chapter order, proposition-to-proof links and source provenance. The authors plan to release the corpus.

The synthetic stereo videos use multiple virtual camera rigs. The data accompanies the StereoSplat work on reconstructing scenes from stereo video and synthesizing new views.

Open Athena describes experiments and trade-offs behind decisions such as model size and token dropping during expert parallelism.

VoxDance Lab shared its VoxMel work and linked the models. It uses wavelet-factorized autoregressive mel-spectrogram modeling to synthesize speech.

The authors show that the relative performance of optimizers can reverse as batch size changes, even when each optimizer is tuned separately at each size.

The benchmark's creator shared about 8 MB of legacy results, with iteration-level data covering 162 model variants.

Expand generates additional samples based on examples and their topics and domains. Co-founder Sara Hooker says some workflows can grow 20 examples into 100,000, with scale and speed depending on the task.

A teacher describes the material, student level and learning goals, then reviews the generated design. Agents build the simulation, and students practice through role-play or explanation and receive feedback.

The point of homework is learning to think.

François Chollet argues that conversational tools can make personalized tutoring more widely available and should give learners a chance to practice reasoning.

A developer used a four-step LoRA to run MOSS-Image-SFT on an iPhone 18 Pro Max. They reported a generation time of 7.02 seconds for the demo.

The demo sends a short prompt to Baidu's model and returns a single HTML file. Users can select a film, game, 3D scene or report as the output.

A creator cloned a scene and asked Fable 5.2 for a shot of a rider crossing a canyon toward a beach. They say the model found a horse animation and changed the shader.

OpenAI's developer team highlighted Lara Avcı's use of Codex to build sites and games for European tech founders. It says the projects attracted more than 30,000 visitors in under four days.

Raschka raises a training trade-off: directly optimizing the reasoning trace may change how useful it is for observing the model's behavior.