September 18, 2025

AI

A new World Trade Organization study predicts artificial intelligence will significantly impact global trade by 2040, potentially increasing trade volumes by 34-37% and boosting GDP by 12-13%. AI-driven mechanisms such as translation tools and automated paperwork are expected to ease trade friction, but inclusive policies are crucial to avoiding exacerbating economic inequalities.

Amazon has introduced an AI agent to help third-party sellers manage their online businesses more efficiently by automating tasks such as inventory coordination and business growth plans. The agent will work with Seller Assistant, a tool launched last year, to free up time for sellers to focus on product innovation and customer relationships.

Alibaba's Tongyi Lab has released an open-sourced large language model, Tongyi-DeepResearch-30B-A3B, with approximately 30.5 billion parameters. The model features a mixture-of-experts design, enabling high throughput while maintaining strong reasoning performance. Optimized for multi-turn research tasks, it supports applications requiring deep information-seeking capabilities, such as searching and synthesizing evidence effectively.

IBM has released Granite-Docling-258M, an open-source vision-language model for end-to-end document conversion. The model extracts layout elements like tables, code, and equations with high accuracy, producing a structured representation instead of lossy Markdown output. Available on Hugging Face, the model provides an enterprise-ready solution for document conversion tasks on Apple Silicon.

Researchers from the Allen Institute for Artificial Intelligence have introduced Fluid Benchmarking, an adaptive Large Language Model (LLM) evaluation method that replaces traditional accuracy-based evaluations with a two-parameter IRT ability estimation and item selection process. This approach aims to improve LLM evaluation robustness and reliability by focusing on informative questions that reflect a model's current ability.

A trust gap in payment systems widens when shopping agents auto-purchase higher-tier plans instead of user intent. This issue raises questions about responsibility and hinders agent-led checkout adoption. Google's Agent Payments Protocol (AP2) aims to address this by creating an open specification for interoperable, agent-initiated payments across merchants and wallets securely.

Google Research has released TimesFM-2.5, a time-series foundation model with 200M parameters and a 16K context length for probabilistic forecasting. The model now surpasses zero-shot foundation models on various accuracy metrics in the GIFT-Eval leaderboard, demonstrating significant progress in time-series forecasting, which involves predicting future values in a sequence of data points accurately.

Stanford University researchers have introduced MedAgentBench, a new benchmark suite designed to assess large language model (LLM) agents in healthcare settings. The virtual EHR environment allows AI systems to engage in multi-step clinical tasks, marking a significant shift towards evaluating real-world scenarios and practical applications of healthcare AI agents more comprehensively.

MoonshotAI has open-sourced a lightweight middleware called checkpoint-engine, aiming to improve large language model (LLM) deployment by efficiently updating model weights across thousands of GPUs without disruption. Designed for reinforcement learning and RL with human feedback, the library addresses one of LLM deployment's key bottlenecks, enabling frequent updates without sacrificing inference performance.

OpenAI has released GPT-5-Codex, an enhanced version of the GPT-5 model designed to improve reliability and autonomous behavior within the Codex ecosystem. The primary objective is to transform Codex into a reliable teammate by optimizing its capabilities for "agentic coding" tasks, expanding its accessibility across various development workflows.

Researchers at Meta Reality Labs and Carnegie Mellon University have developed MapAnything, an end-to-end transformer architecture that generates factored metric 3D scene geometry from images and sensor inputs. This breakthrough enables over-the-top generalization of 3D vision tasks without specialized pipelines, marking a significant advancement in the field with multiple applications.

A recent Pew study reveals a growing skepticism towards Artificial Intelligence's impact on daily life among Americans, with 50% expressing concern rather than excitement. Concerns center around AI's effects on creative thinking and relationships, as well as its role in personal decisions, with many preferring AI focus on practical applications like data analysis.

Google is facing a lawsuit from Penske Media over its AI Overviews feature, which provides quick summaries to users. The company's policy executive Markham Erickson claims users prefer these over traditional search results, sparking controversy, particularly given allegations that Google is using AI to lift journalism without permission or proper attribution.

Google has laid off hundreds of AI workers, including over 200 contractors, amid a fight over pay and working conditions. The move follows attempts to unionize despite facing retaliation from outsourcing company GlobalLogic. Workers claim poor treatment, stress, and lack of job security, with some threatened with job loss if they didn't meet unrealistic targets.

A tutorial guides readers in building a sophisticated voice AI agent using freely available models from Hugging Face, seamlessly integrating Whisper for speech recognition, FLAN-T5 for natural language reasoning, and Bark for speech synthesis through transformers pipelines. The goal is to create an advanced voice AI agent that can run smoothly on Google Colab.

Tesla has agreed to settle a wrongful death lawsuit from a fatal 2019 crash involving its Autopilot system, which struck and killed 15-year-old Jovani Maldonado. This settlement comes amid growing lawsuits over Autopilot crashes, including a recent $243 million award in Florida, as the company faces nearly 60 deaths linked to Autopilot or Full Self-Driving since 2021.