Google DeepMind Announces AlphaGenome 2 for Cell-Specific DNA Variant Prediction

Runway · XReported by 1 other source

Runway Makes Praxis-1 Available with Apache 2.0 Open Weights

Runway is making Praxis-1 available to everyone as a world action model that turns video pretraining into control for real robots. It ships with open weights under Apache 2.0 and can be run and fine-tuned on users’ own hardware.

• vLLM-Omni Adds Day-One Support for Runway’s Praxis-1vLLM · X

QbitAI

Baidu Introduces ERNIE 5.5 Flash Preview for End-to-End Coding

Baidu introduced ERNIE 5.5 Flash Preview ahead of its Baidu World conference and said the model will soon be open source. The article reports 280 billion total parameters, 14 billion activated per inference, a 1 million token context window, and generation above 200 tokens per second. Its demonstrations include a browser-playable Backrooms game and a Blender-built Suzhou garden, while Baidu says it trained the model with 14 million executable code tasks.

The DecoderReported by 5 other sources

Google Introduces Gemini Decision 1 for Scoring 1,000 Questions in One Pass

Google DeepMind is introducing the Gemini Decisions API, powered by Gemini Decision 1, which processes up to 1,000 independent closed-ended questions in one pass across text, images, audio, video, and PDFs. It returns probabilities with confidence values in 150 to 250 milliseconds. The API is available through Google AI Studio for $0.10 per million input tokens, with no output-token charge; integrations with OpenRouter, LangChain, Vercel, and other platforms are rolling out.

Gemini App · X

Gemini Is Now Available on Apple CarPlay

Google says Gemini is now available on Apple CarPlay. Drivers can tap the Gemini app icon on the CarPlay screen or side rail to continue a conversation from their phone, explore a topic, or brainstorm aloud with Gemini Live.

The Decoder

Anthropic Says AI Cut Claude’s Switch to AMD GPUs to Three Days

Anthropic says it used its Mythos model to port Claude to AMD GPUs in three days, with almost no human intervention, compared with six to twelve months using traditional engineering approaches. The figures refer to prompt processing on AMD MI355X GPUs and come from an investor presentation that has not been independently verified. Anthropic says it allocates compute across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs.

Latent Space

Mercor Raises $750 Million in Series D at a $20 Billion Valuation

Mercor raised $750 million in a Series D at a $20 billion valuation after reaching $3 billion in annualized revenue, according to the AINews roundup. The article also reports that the company has 5 million experts and quotes a statement that it spends three times as much on inference as on employee salaries.

Ars Technica · AI

Big Walk’s AI-Assisted Bug Fix Divides Indie Game Award Jurors

House House disclosed that it used an LLM sparingly to help address technical challenges and bugs in Big Walk, including a Unity shadow issue; the developers say the resulting code accounts for well under 0.1% of the game. The disclosure has divided Independent Games Festival jurors, with some rejecting any game that uses generative AI and others arguing that limited, transparent use differs from relying on AI to create a game’s art or core systems. The debate has also drawn criticism from fans and industry observers.

Google ResearchFirst-party

Google Research Presents Wearable-Based CKM Risk Profiling for Personalized Coaching

Google Research introduces an investigational framework that uses wearable data to estimate a coachable health age and simulate personalized 12-week health trajectories. Its Sensor World Model was trained on 2.28 million person-weeks of data from roughly 100,000 consenting Fitbit users; in a blinded study, lifestyle coaches preferred counterfactual-grounded plans in 71.4% of evaluations. The post says the models were evaluated offline using retrospective cohorts and are research prototypes, not clinical diagnostic tools.

Baoyu

Stanford CS329Z Publishes Agent Engineering Lectures and Two Assignments

Stanford’s CS329Z course is making its Engineering AI Agents materials public, with lectures appearing daily on Stanford Online’s YouTube channel and two MIT-licensed assignments on GitHub. The first assignment has students build an agent harness from scratch and includes a test suite; the second teaches students to create and evaluate a Terminal-Bench-style task. The post also describes the assignments’ API costs and requirements.

Baoyu

A Post Compares Google Docs’ Cloud Shift With Claude Artifacts for Agents

Baoyu argues that Google Docs changed software from a single-computer model to a cloud model, while Claude Artifacts points to software made for agents, with people reviewing and refining the results. He concludes that changing who uses software also changes its form.