Moonshot AI Raises $3.5 Billion at a $35 Billion Valuation
Chinese AI startup Moonshot AI (月之暗面) has completed a $3.5 billion funding round, bringing its valuation to $35 billion, Bloomberg reported. The round exceeded the company's initial fundraising target.
Moonshot AI is best known for its Kimi family of large language models and recently drew considerable attention with the release of Kimi K3, a 2.8-trillion-parameter open-weights model. The new funding confirms the company's standing as one of the best-funded AI players in China and further solidifies its position against U.S. rivals such as OpenAI and Anthropic.
Microsoft's Community Library Creator Puts AI Data Representation in Communities' Hands
Microsoft Research has announced Community Library Creator, a platform that enables communities to actively shape how they are represented in AI models.
How a particular community is depicted in AI-generated images and text depends heavily on training data. Traditional data construction processes, however, have struggled to reflect the voices of the people being represented. Community Library Creator provides a mechanism for communities to define for themselves what appropriate representation looks like and to structure those standards as data that AI models can learn from.
Anja Thieme of Microsoft put it this way: “There is no single absolute right answer to how people should be represented. It is something that must be collectively defined and negotiated.” The project is drawing attention as a technically grounded effort to support fairness and inclusion in AI.
Microsoft's Mage-VL: Codec-Style Architecture for Real-Time Visual Understanding
Microsoft has also released Mage-VL, a codec-native streaming multimodal foundation model, on Hugging Face. It is a compact 4B-parameter model, notable for training its vision encoder from scratch.
Mage-VL tackles the “Moravec's paradox” that plagues conventional vision-language models (VLMs): they excel at complex reasoning, while simple real-time perception remains slow and computationally expensive. Rather than decomposing video into uniformly sampled frames, Mage-VL mirrors the structure of video codecs, splitting the stream into anchor frames (I-frames) and predicted frames (P-frames), and keeping only the patches in the regions where a codec would allocate bits — that is, where change actually happens.
The design promises efficient image and video understanding in streaming settings. The paper is also registered on Papers with Code, and validation by the research community seems likely to follow.
Encore AI Raises $30 Million for AI Agents That Learn from Customer Interactions
Encore AI has raised $30 million, TechCrunch reports. The company is building a platform that analyzes customer calls, messages, and CRM data to identify effective sales practices and turns them into playbooks for AI agents.
Where existing AI sales assistants run on generic scripts, Encore AI's differentiator is that it learns its approach from real-world successes. As AI increasingly blurs the boundary between customer support and sales, this real-data-driven approach looks highly practical.
Teacher Arrested Amid Anti-AI-Data-Center Protests as Social Friction Deepens
Tom's Hardware reports that in the United States, a teacher was arrested at a public hearing over an AI data center construction project — for applauding opposing comments. The gigawatt-scale data center project was ultimately approved, overriding local opposition.
Community concerns over the rapid buildout of AI infrastructure are mounting, centered on power consumption and environmental impact. The incident stands as a symbolic example of the friction beginning to surface between the AI industry's growth and the rights of local communities.
“Uncensored” LLMs Turn More Optimistic — Testing Reveals an Unexpected Side Effect
An interesting set of test results was shared in the LocalLLaMA community on Reddit: “uncensored” models do not merely stop refusing to answer questions — the model's overall attitude changes as well.
The tester found that a de-censoring technique known as abliteration does more than remove refusal behavior; it also affects the model's personality traits. Specifically, after de-censoring, models became more optimistic in tasks such as stock price prediction, producing more “it will go up” forecasts and fewer hedging phrases like “might” or “possibly.” Actual predictive accuracy, however, did not change, remaining on par with a coin toss.
In short: more confident, but not more correct. The results have sparked debate in the community as evidence that de-censoring has unintended effects not only on safety but also on a model's output bias.