Meta's AI Ambitions: Closed Models and Potential Copyright Challenges In a move that has raised eyebrows in the tech community, Meta (formerly Facebook) is reportedly training its new AI models using rival systems like Gemini, ChatGPT, and Claude. This comes as the company faces legal challenges over its alleged use of copyrighted material to train its own language models. According to reports, Meta's AI division has been working on new models named "Avocado" and integrating technology from its recent acquisition of Manus AI. The company is also testing features like "Fast" and "Thinking" modes, as well as integrations with productivity apps like Gmail and Outlook. The legal battles surrounding Meta's AI training practices are also heating up. In one case, a group of book authors accused the tech giant of using pirated books to train its large language model, Llama. While Meta secured a partial fair use victory, the court emphasized the importance of proving potential market harm from AI-generated content. In a separate lawsuit, adult film producers Strike 3 Holdings and Counterlife Media allege that Meta downloaded at least 2,396 of their films via BitTorrent to aid its AI training. The producers fear that this could lead to AI models capable of producing high-quality adult films at a lower cost, potentially disrupting their industry. These legal challenges underscore the growing tensions between AI development and copyright protection. As tech companies like Meta push the boundaries of what's possible with AI, they are increasingly finding themselves at odds with content creators who want to protect their intellectual property. Meta's response to these issues has been to double down on its AI ambitions. The company is reportedly investing heavily in a new $10 billion data center in Indiana, which it says will help mitigate the energy and water impact of its AI operations. However, critics argue that the long-term job creation from such projects is often small compared to the public subsidies received. As the AI race heats up, Meta's strategy of leveraging rival models and potentially using copyrighted material for training purposes suggests a shift towards a more closed and profit-driven approach. This could have significant implications for the broader AI ecosystem, as well as the rights of content creators. The outcome of these legal battles will likely shape the future of AI development and its relationship with intellectual property.
Anthropic Accuses Chinese AI Labs of Stealing Data from Its Claude AI Model In a major development, Anthropic, the AI company behind the Claude chatbot, has accused three Chinese AI labs - Deepseek, Moonshot, and MiniMax - of conducting large-scale "distillation attacks" to steal data and capabilities from Claude. According to Anthropic, the three labs used over 24,000 fake accounts to fire off more than 16 million queries targeting Claude's reasoning, programming, and tool usage abilities. The attacks were carried out using proxy services to bypass China's access restrictions. Deepseek specifically targeted Claude's reasoning chain, extracting thought processes and censorship-compliant answers on sensitive topics. MiniMax ran the biggest campaign with over 13 million requests and quickly pivoted to a new Claude model within 24 hours when Anthropic released an updated version. Anthropic says it was able to link these distillation attacks to the Chinese firms with "high confidence" through IP address correlation, metadata analysis, and infrastructure indicators. The company also claims that OpenAI and Google have reported similar attempts from Chinese labs. In response, Anthropic is calling on the industry and policymakers to mount a coordinated effort to address these "industrial-scale" data theft campaigns. The company says it will also upgrade Claude's systems to make distillation attacks harder to execute and easier to identify. This development comes as Anthropic is already facing a lawsuit from music publishers who have accused the company of using illegal copies of songs to train Claude. The company is also experimenting with new features, such as giving Claude the ability to end "distressing" conversations as a last resort. Experts say these incidents highlight the growing competition and potential for abuse in the rapidly evolving AI landscape. As AI models become more powerful, the race to extract and leverage their capabilities is intensifying, raising concerns about data privacy, security, and the responsible development of transformative technologies.
Distilling the tool-using capabilities of large language models (LLMs) into smaller, more efficient small language models (SLMs) is a key challenge for their practical application. The predominant approach, supervised fine-tuning (SFT), suffers from poor generalization as it trains models to imitate a static set of teacher trajectories rather than learn a robust methodology. While reinforcement learning (RL) offers an alternative, the standard RL using sparse rewards fails to effectively guide SLMs, causing them to struggle with inefficient exploration and adopt suboptimal strategies. To address these distinct challenges, we propose MENTOR, a framework that synergistically combines RL with teacher-guided distillation. Instead of simple imitation, MENTOR employs an RL-based process to learn a more generalizable policy through exploration. In addition, to solve the problem of reward sparsity, it uses a teacher's reference trajectory to construct a dense, composite teacher-guided reward that provides fine-grained guidance. Extensive experiments demonstrate that MENTOR significantly improves the cross-domain generalization and strategic competence of SLMs compared to both SFT and standard sparse-reward RL baselines.
Gradient-based data attribution methods, such as influence functions, are critical for understanding the impact of individual training samples without requiring repeated model retraining. However, their scalability is often limited by the high computational and memory costs associated with per-sample gradient computation. In this work, we propose GraSS, a novel gradient compression algorithm and its variants FactGraSS for linear layers specifically, that explicitly leverage the inherent sparsity of per-sample gradients to achieve sub-linear space and time complexity. Extensive experiments demonstrate the effectiveness of our approach, achieving substantial speedups while preserving data influence fidelity. In particular, FactGraSS achieves up to 165% faster throughput on billion-scale models compared to the previous state-of-the-art baselines. Our code is publicly available at https://github.com/TRAIS-Lab/GraSS.
Large language models (LLMs) have demonstrated promising performance in generating diagnostic conclusions from imaging findings, thereby supporting radiology reporting, trainee education, and quality control. However, systematic guidance on how to optimize prompt design across different clinical contexts remains underexplored. Moreover, a comprehensive and standardized framework for assessing the trustworthiness of LLM-generated radiology reports is yet to be established. This study aims to enhance the trustworthiness of LLM-generated liver MRI reports by introducing a Multi-Dimensional Credibility Assessment (MDCA) framework and providing guidance on institution-specific prompt optimization. The proposed framework is applied to evaluate and compare the performance of several advanced LLMs, including Kimi-K2-Instruct-0905, Qwen3-235B-A22B-Instruct-2507, DeepSeek-V3, and ByteDance-Seed-OSS-36B-Instruct, using the SiliconFlow platform.
