Top Researchers

Meta releases Muse Glimmer, a 30B open-weight agentic LLM

Meta's release of Muse Glimmer — a 30B dense open-weight model under Apache 2.0 — is the headline move, marking the company's return to open-weight LLMs with a model built explicitly for agentic use:…

184 themes mapped1048 sourcesActivity 99/100

Next signals to watch

Daily brief

Five signals shaping AI today

Live map

A 60-second overview before you explore the evidence graph.

Daily brief · Aug 18

Five signals shaping AI today

Meta releases Muse Glimmer, a 30B open-weight agentic LLM

4 sources · 1 corroborated

Meta released Muse Glimmer, a 30B dense open-weight model licensed under Apache 2.0, marking a return to open-weight LLM releases for the company. The model supports tool use, multi-step reasoning, and multimodal agentic workflows, and benchmarks at approximately the level of Qwen 3.6-27B. [2][3][4]

Muse Glimmer can be quantized to run within 17GB of memory, and combined with quantization and speculative decoding achieves 50 tokens per second. [4]

U.S. Treasury Secretary Scott Bessent publicly welcomed Meta's release of Muse Glimmer, framing it as a win for American innovation and U.S. leadership in AI, particularly in the open-source domain. [1]

Day-zero support for Muse Glimmer was announced by at least one third-party platform (Actual), indicating immediate ecosystem integration upon release. [3]

Google DeepMind launches SL2T sign-language-to-text model for Pixel 11

4 sources · 1 corroborated

Google DeepMind introduced SL2T (Sign Language to Text), described as a "massively multilingual sign-language-to-text translation model" and the first general-purpose sign-language-to-text system brought into a consumer product, addressing an estimated 70 million Deaf and hard of hearing people who use more than 200 sign languages — a population previously excluded from AI-powered language interfaces. [4]

SL2T powers sign-to-text dictation in Gboard and Live Transcribe on the Pixel 11, starting with American Sign Language (ASL) to English, enabling users to sign to their phone anywhere they would normally type — including web search, drafting messages, and querying Gemini. [2][3][4]

In Live Transcribe specifically, SL2T allows Deaf users to sign responses during conversations instead of typing back and forth; testers reported that signing in ASL is faster, more natural, and more delightful than typing in English. [4]

The feature was built in close collaboration with the Deaf community, according to Demis Hassabis, and additional devices and sign languages beyond ASL-to-English are planned to follow. [1][4]

Simulation Training and Crowdsourced Data Scale Robotic AI Development

4 sources · 1 corroborated

Google DeepMind's Gemini Robotics 2 uses sim-only post-training on a mixed-data base model to handle automotive kitting and other contact-rich tasks, meaning the robot's post-training phase relies entirely on simulated environments rather than real-world data collection. [1]

Axis Robotics' robotic data engine on the Base platform scaled from 1 million trajectories on June 7 to 3 million trajectories, a 3x increase, generated by over 123,000 contributors worldwide. [2][3]

A widely noted challenge in physical AI development is that robots require large volumes of high-quality, diverse data to improve, and static datasets are insufficient because real-world variation demands continuous, dynamic data collection. [4]

DeepSeek forms capital partnership with Unitree Robotics for physical AI

6 sources · 1 corroborated

DeepSeek, the Chinese AI startup, is entering a capital and business partnership with Unitree Robotics (宇樹科技), a humanoid robot developer, with the two companies cooperating toward the realization of AGI with human-equivalent capabilities. The deal was revealed through IPO-related documents disclosed by Unitree, which show DeepSeek investing 140 million (figure truncated in the paywalled source) in conjunction with Unitree's IPO. [1][2][3][4][5][6]

Dyna Robotics releases Dyna-2 World Action Model trained on 1M hours

6 sources · new

Dyna Robotics released Dyna-2, a World Action Model (WAM) explicitly distinguished from Vision-Language-Action (VLA) models; it uses video diffusion as its base architecture to jointly learn future video prediction and actions rather than treating them separately. [2][3][4]

Dyna-2 was pre-trained on over one million hours (described as approximately 170 years) of human first-person activity video, covering tasks such as cooking, tidying, and laundry folding. [5][6][4]

Alongside the model release, Dyna Robotics introduced a laundry-folding robotics benchmark, framing it as a standardized evaluation task for the broader robotics research community. [1]