OpenAI Faces Scrutiny Over Potential Violations of California's AI Safety Law In a developing story, the AI research company OpenAI is facing allegations that its latest coding model, GPT-5.3-Codex, may have violated California's new AI safety law, SB 53. The law, which went into effect in January, requires major AI companies to publish and adhere to their own safety frameworks to prevent catastrophic risks from their models. According to the AI watchdog group Midas Project, OpenAI failed to implement the necessary safeguards outlined in its own safety framework when releasing the high-risk coding model. OpenAI CEO Sam Altman has acknowledged that the model is the first to hit the "high" risk category on the company's internal Preparedness Framework, meaning it has the potential to facilitate significant cyber harm, especially if used at scale. An OpenAI spokesperson disputed the watchdog's claims, stating that the company is "confident in our compliance with frontier safety laws, including SB 53." However, the case could set a precedent as a potential first test of the new California law, potentially exposing OpenAI to millions of dollars in fines if found in violation. The controversy surrounding GPT-5.3-Codex highlights the growing concerns over the cybersecurity risks posed by advanced AI models. Security experts have warned about the potential dangers of OpenClaw, an open-source autonomous AI agent developed by Peter Steinberger, which gives users largely unfettered power to customize and interact with computer systems, raising significant security concerns. Meanwhile, OpenAI has been vocal about its ambitious plans for the future of AI research. The company has released a detailed timeline for its work on artificial general intelligence (AGI), aiming to develop an AI system with "research intern" capabilities by 2026 and a fully autonomous AI researcher by 2028. This vision has been met with both excitement and skepticism, as some former OpenAI researchers, like Jerry Tworek, have raised concerns about the current limitations of AI models in learning from mistakes and adapting to new challenges. As the AI landscape continues to evolve, the OpenAI case highlights the growing need for robust regulatory frameworks and safety measures to ensure the responsible development and deployment of these powerful technologies. The outcome of this dispute may have far-reaching implications for the AI industry and the future of AI governance.
Meta's Ambitious AI Expansion Raises Concerns Over Costs and Copyright Infringement In a series of recent developments, tech giant Meta has been making significant strides in the field of artificial intelligence (AI), but not without facing some challenges and controversies. According to reports, Meta has broken ground on a massive $10 billion AI data center in Indiana, which is expected to create 4,000 construction jobs and 300 operational positions. However, the long-term job creation is likely small compared to the public subsidies the company may receive, raising questions about the true economic impact of such projects. Furthermore, Meta's AI ambitions have led to a copyright infringement lawsuit filed by adult film producers Strike 3 Holdings and Counterlife Media. The complaint alleges that Meta has downloaded at least 2,396 of their films since 2018 to aid in the training of its AI models, including the Meta Movie Gen and Large Language Model (LLaMA). The adult producers fear that this training may result in AI models that can create similar "Hollywood grade" films at a lower cost, potentially undermining their business. Alongside these developments, Meta is also making efforts to integrate its products with AI-powered platforms, such as ChatGPT and Microsoft Copilot. The company's chief technology officer, Jetan Chowk, has stated that AI is "the future of how people will shop," and JD Sports, a British sportswear retailer, has announced plans to allow customers to make one-click purchases through these AI platforms. These moves by Meta highlight the company's quest for profits and dominance in the AI space, which may come at a cost. The massive investments in data centers and the potential copyright infringement issues raise concerns about the company's practices and the broader implications for the industry. As Meta continues to expand its AI capabilities, it will be crucial for the company to navigate these challenges while maintaining transparency and addressing the concerns of both its customers and the broader public.
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.
