Daily AI News | Google Reshuffles DeepMind Leadership as Senior Researchers Depart
# Daily AI News | Google Reshuffles DeepMind Leadership as Senior Researchers Depart
## Lead Story
### Google Reshuffles DeepMind Leadership as Senior Researchers Depart
Google said Demis Hassabis will leave the Google DeepMind CEO role to become the lab's chair and Alphabet's chief scientist, while CTO Koray Kavukcuoglu takes responsibility for day-to-day management. Google chief scientist Jeff Dean and several veteran researchers will depart to form science-automation startup Discovery Loop, which Google plans to back.
**AI take:** The research chain of command and the core talent base are changing together; Gemini's delivery pace is the next practical test.
## More News
### Meta Launches Muse Code for Large Software Repositories
Meta launched its first coding agent, Muse Code, alongside an updated Muse Spark 1.2 model. Available for macOS and Linux, the agent is designed to run complex tasks asynchronously in the background across large repositories, putting it in direct competition with Claude Code and Codex.
**AI take:** Long-task reliability, permission controls, and real pricing will matter more than launch-day benchmark comparisons.
### Anthropic Forms an Internal AI Chip Design Team for Claude
Anthropic is hiring engineers for an internal AI chip design team that will co-design hardware and Claude models for better speed and energy efficiency. The company has primarily relied on accelerators supplied by Nvidia, Google, Amazon, and other outside partners.
**AI take:** Custom silicon could lower long-run inference costs, but moving from a new team to deployed chips is a multiyear project.
### Qwen-Image-3.0 Opens to All Users on the Qwen AI Platform
Alibaba made Qwen-Image-3.0 available to all users on the Qwen AI platform and exposed an API. The release supports complex layouts, prompts of roughly 4,500 tokens, and text rendering down to 10 pixels, with text-to-image API pricing starting at RMB 0.18 per image.
**AI take:** Broad access matters more than rankings; stability on dense text and editing tasks will determine practical value.
### Tencent Releases the Hy ASR 3.0 Preview Speech Model
Tencent Hunyuan released Hy ASR 3.0 preview, extending speech recognition from literal transcription toward contextual meaning while improving dialect and complex-scene handling. The preview debuted in Tencent Yuanbao, with selected capabilities also opened to developers.
**AI take:** Enterprises will need to test noise, accents, and specialist vocabulary before treating contextual recognition as production-ready.
### Dreamina Adds Maya and Blender Tools for Seedance 2.5
ByteDance's Dreamina integrated Seedance 2.5 and released native Maya and Blender plugins together with smart-editing tools. The update brings video generation into 3D production workflows instead of limiting it to one-off generation in a web or mobile interface.
**AI take:** Native plugins reduce adoption friction, but shot consistency and editability will decide whether studios use them in final production.
### US Appeals Court Vacates Injunction Against Perplexity's Shopping Agent
The US Court of Appeals for the Ninth Circuit vacated Amazon's preliminary injunction against Perplexity's Comet shopping agent and remanded the case. On the current record, the court said the user accessed Amazon's servers with the agent as a tool, making Amazon unlikely to prove that Perplexity itself performed the access under the relevant computer-access statutes.
**AI take:** The ruling leaves room for user-directed browser agents, but the court expressly declined to create a general legal rule for all agents.
### Wayve and Uber Win Approval for Supervised Autonomous Rides in London
Wayve and Uber received Transport for London licensing for supervised passenger rides in autonomous vehicles. The approved vehicles still require safety drivers, moving the companies from testing and sign-up plans into a regulated passenger-service stage.
**AI take:** This is supervised access, not approval for driverless scale; safety evidence and further regulation remain hard gates.
### Zoox Will Start Charging for Las Vegas Robotaxi Rides on August 10
Amazon-owned Zoox said it will begin charging fares on August 10 for rides in its purpose-built Las Vegas robotaxis, which have no steering wheel or pedals. The service had offered free rides and is moving into paid operation after receiving a federal exemption last week.
**AI take:** Paid rides are a commercialization milestone; service area, fleet scale, and sustained safety performance come next.
### Doubao Rolls Out SeedRealtime, a Native Full-Duplex Audio-Video Model
ByteDance's Seed team launched SeedRealtime on August 5 with audio, video, and text inside one native architecture. It can watch, listen, and speak during a continuous interaction, including proactive prompts, turn timing, and background-noise handling, and is now fully rolled out in the Doubao app.
**AI take:** Native audio-video full duplex moves interaction beyond alternating turns toward continuous perception; latency, false triggers, and privacy controls are the real tests.
### Sand AI Open-Sources the 114B MAGI-2 Preview Audio-Video Model
Sand AI released inference code and weights for MAGI-2 Preview. The model has 114B total parameters with about 6B active per token, supports text-to-video and image-to-video with generated sound, produces 10-second clips, and uses a two-stage process to reach 1080p. The official repository uses the Apache 2.0 license.
**AI take:** Sparse activation lowers active compute, but eight Hopper GPUs and roughly 307GB of weights keep local deployment demanding; reproducible code and weights are the meaningful part of this release.
### U.S. Framework Reportedly Excludes Open-Weight Models from Voluntary Pre-Release Testing
Reuters and Axios reported that a White House voluntary pre-release testing framework would primarily cover closed, frontier models that may pose national-security risks, while excluding open-weight models. The framework is not public, and its definitions and boundaries remain unclear.
**AI take:** Exclusion lowers a pre-release gate for open weights, but the unpublished definitions, scope, and post-release misuse controls remain unresolved.
### Alibaba Cloud Launches One Key MCP for Shared Authentication and Billing
Alibaba Cloud introduced One Key MCP in Model Studio, allowing developers to use one API key across an initial catalog of 14 partners in six fields. It works with Qoder, Codex, Claude Code, and Cursor, while access, calls, and billing are managed from one entry point.
**AI take:** A shared key removes integration friction but increases the blast radius of one credential, so granular permissions, rotation, and auditing still matter.
### Tencent Expands Global Access to Hy3
Tencent expanded its 295B-parameter Hy3 model to developers, businesses, and users worldwide through WorkBuddy, Tencent Design Miora, and TokenHub. WorkBuddy is free globally through August 31. This is a new availability stage after the July launch, not a second model release.
**AI take:** Distribution is the material change; regional terms, API stability, and pricing after the promotion are the next things to watch.
## Rumors
No rumors met the publication threshold in this edition.
