Tech Giants Lobby to Prevent State AI Regulations; MIT Finds 95% of Generative AI Projects Yield Zero Return
Key Takeaways
- Tech giants lobby to shift AI regulation from state to federal level in the US.
- MIT report: 95% of generative AI pilot projects see 'zero return' on investments.
- New Relic: ChatGPT dominates LLM token usage with over 86% market share.
- Meta freezes AI hiring after recruiting over 50 researchers from OpenAI, Google.
- Article warns of 'AI agent washing,' where simple chatbots are misleadingly labeled as 'agents'.
Top Stories
Tech giants lobby to prevent US states from regulating AI.
Tech giants like OpenAI, Meta, and Google are lobbying the White House and Congress to shift AI regulation from state to federal level. They aim to prevent individual US states from regulating their AI businesses, fearing it could hinder growth.
MIT: 95% of generative AI projects see 'zero return'.
On August 20-21, 2025, an MIT report revealed that 95% of generative AI pilot projects from enterprises see 'zero return' on investments. The study analyzed over 300 initiatives, citing strategy, culture, and integration as primary issues.
ChatGPT dominates LLM token usage at 86%, New Relic report shows.
On August 21, 2025, New Relic released its "AI Unwrapped: 2025 Impact Report," revealing that ChatGPT dominates with over 86% of LLM token usage. The report is based on anonymized data from 85,000 active customer accounts.
Meta freezes AI hiring after recruiting from OpenAI, Google, others.
On August 21, 2025, Meta Platforms froze AI hiring after recruiting over 50 AI researchers and engineers from companies like OpenAI and Google. The hiring freeze coincides with a broader restructuring.
Article warns of 'AI agent washing,' misleading AI labeling.
The article discusses 'AI agent washing,' a trend where companies label simple chatbots as 'agents' without substantial agentic capabilities. This practice confuses users and slows meaningful progress in AI adoption.
AI Breakthroughs
MedReseacher-R1 achieves new state-of-the-art results on medical benchmarks.
Researchers introduced MedReseacher-R1, a medical deep research agent, featuring a novel data synthesis framework and private medical retrieval engine. The MedReseacher-R1-32B model demonstrated exceptional performance, setting new state-of-the-art results on medical benchmarks.
DeepSeek launches V3.1, open-source AI model rivaling GPT-5.
On August 19-21, 2025, DeepSeek launched its V3.1 model, an open-source AI rival to GPT-5, on Hugging Face. The model has 685 billion parameters and a 128k tokens context window, optimized for Chinese chips.
LookOut predicts collision-free trajectories from egocentric observations.
Researchers introduce LookOut, a framework for predicting collision-free future trajectories from egocentric observations. It reasons over temporally aggregated 3D latent features and models geometric and semantic constraints.
New model-driven approach introduced for training Vision Foundation Models.
Jiabo Huang, Chen Chen, and Lingjuan Lyu introduced a new model-driven approach for training Vision Foundation Models (VFMs) through joint knowledge transfer and preservation. Their VFM outperforms existing data-centric models across four fundamental vision tasks.
Models introduced to address human-agent behavioral differences.
A paper introduces the 'Network Behavior Lifecycle' model and the 'Human-Agent Behavioral Disparity (HABD)' model. The HABD model examines behavioral differences between humans and agents across 5 dimensions to address challenges in trust, responsibility, ethics, and security.
Cognitive Surgery framework awakens Territorial Awareness in LLMs.
Researchers propose Cognitive Surgery (CoSur) framework to awaken Implicit Territorial Awareness (ITA) in Large Language Models (LLMs). CoSur comprises 4 modules and shows improved performance in IPP scenarios.
New approach improves logical structure understanding of arguments.
A paper published on arXiv proposes an approach to improve the logical structure understanding of natural language arguments. The authors introduce an intermediate representation called the Representation of the Logical Structure (RLS) of the argument.
D²-LIO enhances LiDAR-inertial odometry in complex environments.
Researchers propose D²-LIO, an enhanced LiDAR-inertial odometry framework that addresses LiDAR feature degeneracy. The framework includes adaptive outlier removal and scan-to-submap registration, improving robustness and accuracy.
MCP-Universe benchmark evaluates LLMs in realistic, real-world tasks.
On August 20, 2025, Salesforce AI Research introduced MCP-Universe to arXiv, a benchmark evaluating LLMs in realistic tasks via interaction with real-world MCP servers. The benchmark includes 6 core domains spanning 11 MCP servers and 231 tasks.
MoCHA-former enhances video demoiréing with adaptive transformers.
Researchers propose MoCHA-former, a Moiré-Conditioned Hybrid Adaptive Transformer for video demoiréing. It addresses limitations in spatially varying artifact strength and outperforms prior methods.
YOLO-GL enhances multi-target detection with feature fusion.
Researchers propose YOLO-GL, an enhanced detection network for efficient multi-target detection in large fields of view via hierarchical feature fusion. YOLO-GL achieves state-of-the-art performance with [email protected] of 74.3%.
APCP framework proposed for agentic AI in collaborative learning.
On August 20, 2025, Lixiang Yan submitted a paper proposing the APCP framework for agentic AI in human-AI collaborative learning. The framework charts the transition from AI as a tool to AI as a collaborative partner.
Context-Aware Memory model improves long-context speech synthesis.
Zhipeng Li et al. published a paper introducing a Context-Aware Memory (CAM)-based model for long-context speech synthesis. The CAM block integrates long-term memory and local context details to guide sentence-level speech synthesis.
Diffusion models perform well with fewer latent states, study finds.
Researchers challenge the assumption that diffusion models require many latent states, showing that models trained over a small number (T ~ 32) match the performance of models trained over a much larger number (T ~ 1,000).
FBI framework enables visuotactile imitation learning for manipulation.
Researchers introduce FBI, a visuotactile imitation learning framework for dexterous in-hand manipulation. FBI dynamically fuses tactile interactions with visual observations through motion dynamics.
FedEve mitigates client drift in cross-device federated learning.
Researchers propose FedEve, a method to mitigate client drift and period drift in cross-device federated learning. FedEve uses a predict-observe framework to reduce the variance of model updates.
LLMs generate code with bugs, vulnerabilities, and code smells.
A study evaluated the code quality and security of LLMs using Java assignments, finding that LLMs generate functional code but also introduce bugs and security vulnerabilities. The research highlights that LLM-generated code requires verification.
MISS improves retrieval recommendation with multi-modal indexing.
Researchers propose MISS, a multi-modal tree indexing and searching method with lifelong sequential behavior for retrieval recommendation. MISS aims to improve retrieval model performance by leveraging extensive user and item information.
Industry Watch
Amazon bets on AI agents, hiring ex-OpenAI researcher Luan.
Amazon is focusing on AI agents, with David Luan, formerly of OpenAI and Adept, leading Amazon’s AGI research lab. The goal is to build a universal AI teammate for knowledge workers, allowing task delegation to AI systems.
FieldAI raises $405M, valuing robotics startup at $2 billion.
FieldAI, a California-based robotics startup, raised $405 million in two rounds, valuing the company at $2 billion. The funding will accelerate global growth of AI-driven robotics for industrial use.
Wisenut to unveil new products optimized for AI agent construction.
와이즈넛 will hold the 'WISE Edge' event on September 17th, 2025, at Yangjae El Tower, and will unveil new products optimized for building AI agents. Key products to be released include WISE LLOA, WISE Agent Labs, and WISE Agent Sphere.
Proton upgrades Lumo AI assistant with faster, smarter responses.
Proton has upgraded its privacy-focused Lumo AI assistant, introducing version 1.1 on August 21, 2025. The update promises faster and more intelligent responses, with a 200% improvement in reasoning.
Real-World AI
Halo X smart glasses launch with AI, raise privacy concerns.
Former Harvard students launched Halo X smart glasses, featuring a display and AI capabilities using Google's Gemini and Perplexity. The $249 glasses record audio and display real-time information, but lack recording indicators, raising privacy concerns.
Deep learning model achieves 96.40% accuracy in detecting weeds.
A deep learning model, fine-tuned MobileNetV2, achieved 96.40% accuracy in detecting weeds in Teff crops using UAV imagery. This enhances weed management and improves Teff crop productivity in Ethiopia.
Anthropic supports higher education with AI tools and resources.
Anthropic is supporting higher education through AI tools and resources, including a Higher Education Advisory Board and Claude for Education. They launched three AI Fluency courses for students and faculty.
Zoox partners with The Routing Company for robotaxi routing.
Zoox has partnered with The Routing Company to enhance its robotaxi service with advanced routing software. The collaboration aims to improve the efficiency and scalability of Zoox's robotaxi network, planning paid public rides in Las Vegas.
DualNILM framework recognizes appliance states, identifies energy usage.
Researchers present DualNILM, a deep multi-task learning framework for appliance state recognition and injected energy identification in NILM. It integrates sequence-to-point and sequence-to-sequence strategies within a Transformer-based architecture.
ChatGPT generates isomorphic physics problems with prompt-chaining.
A method for generating isomorphic physics problems using ChatGPT with prompt-chaining and tool use is presented. This approach allows for precise control over structural variations and supports diverse contextual variations.
SeeMe detects facial movements in unresponsive brain injury patients.
Researchers developed SeeMe, a computer vision tool that detects covert voluntary facial movements in unresponsive brain injury patients. SeeMe identifies low-amplitude movements in response to auditory commands.
AI methods developed for autism spectrum disorder (ASD) diagnosis.
Researchers have developed methods for autism spectrum disorder (ASD) diagnosis. One approach is ASDFormer, a Transformer-based architecture using fMRI data, achieving state-of-the-art diagnostic accuracy.
AI to redefine higher education with AI professors, personalized learning.
On August 20, 2025, AI's impact on higher education was discussed, highlighting AI professors, personalized education, and improved affordability. AI will offer personalized learning experiences and objectively assess student competence.