OpenAI Scientist's Latest Talk: GPT-4 is About to Reach a Turning Point, 1000x Performance Surge is Inevitable!

Vision-R1: Reinforcing Visual Localization with RL, Achieving 50% Performance Boost

Recently, a collaborative effort between the Institute of Automation, Chinese Academy of Sciences, and the CASIA-Zidong Taichu team introduced Vision-R1, a novel method leveraging R1-like reinforcement learning to significantly enhance visual localization capabilities. This approach achieved a 50% performance improvement in complex tasks such as object detection and visual localization, surpassing existing state-of-the-art (SOTA) models with over ten times the parameter scale. Currently, large vision-language models typically rely on "pre-training + supervised fine-tuning" to improve responsiveness to user instructions, but this...

IBM Unveils z17 Mainframe: Capable of 450 Billion AI Inferences Daily, 50% Performance Boost

IBM on Monday launched its latest mainframe hardware, the IBM z17. This fully encrypted mainframe, powered by the IBM Telum II processor, is designed for over 250 AI use cases, including AI agents and generative AI applications. While mainframes may be considered legacy technology by some, 71% of Fortune 500 companies still use them, according to sources. According to market research firm Market Research Future, by 2024...

Qwen3 is Coming Soon: Alibaba Cloud's New Model Integrates with vLLM, High Performance Anticipated

Recently, Alibaba Cloud's Qwen series of AI large language models has seen significant progress. Support for its next-generation model, Qwen3, has been officially merged into the vLLM (efficient large language model inference framework) codebase. This news has sparked heated discussions in the tech community, signaling that Qwen3's release is imminent. It is understood that Qwen3 will include at least two versions: Qwen3-8B and Qwen3-MoE-15B-A2B, representing innovative attempts at different scales and architectures, for developers and enterprises.

SenseCore 2.0 Upgrade Imminent: Exponential Growth in Computing Power Expected from SenseTime

SenseTime announced that its 2025 SenseTime Technology Exchange Day will be held on April 10th at 2 PM. SenseTime's SenseCore AI infrastructure is set for a major upgrade, promising exponential growth in its technological capabilities. This upgrade will significantly enhance its capabilities, particularly in AI infrastructure, embodied AI research and development, and the application of industry-specific large models. Since its launch in 2021, SenseCore has successfully integrated nationwide resources as a new type of AI infrastructure.

OpenAI Adds Watermarks to GPT-4o Image Generation, Plus Members Get Exclusive Access

OpenAI is actively testing a new feature to add watermarks to its GPT-4o image generation model. Recently, ChatGPT has garnered significant attention for its powerful image generation capabilities. The model not only accurately generates images with text but also creates highly realistic visual works, even mimicking the artistic styles of renowned animation studios like Studio Ghibli. Image attribution: Image generated by AI, image licensing provider Midjourney. Initially, this feature is reportedly only available to ChatGPT Plus subscribers.

Google Research Shows Synthetic Data Improves Large Model Logical Reasoning Eightfold

Google Releases Gemini 2.5 Pro API, Cheaper Than GPT-4

Google recently quietly released its latest flagship AI model, Gemini 2.5 Pro, causing a stir in the developer community. Google calls it the most intelligent model ever, and this powerful model is now available in public preview via the Gemini API on Google AI Studio and will soon be available on Vertex AI. This move allows more developers to experience top-tier AI capabilities at a highly competitive price.

Meta Accused of AI Model Double Standard: Maverick's Performance Varies Widely Between Evaluation and Public Versions

Meta released its new flagship AI model, Maverick, on Saturday. The model ranked second in the LM Arena benchmark. LM Arena is a testing platform that relies on human raters to compare different model outputs and select their preferences. However, several AI researchers quickly discovered that the version of Maverick deployed to LM Arena appears significantly different from the version widely used by developers. Meta acknowledged in its announcement that the Maverick on LM Arena is an experimental version.

Google Research: Synthetic Data Boosts Large Model Math Reasoning Eightfold

A recent joint study by Google, Carnegie Mellon University, and MultiOn explores the application of synthetic data in training large language models. According to Epoch AI, a research institution focused on AI development, currently available high-quality text training data totals around 300 trillion tokens. However, with the rapid advancement of large models like ChatGPT, the demand for training data is growing exponentially, projected to exhaust existing resources by 2026. Therefore, synthetic data is becoming increasingly crucial.

Meta Announces Open-Source Multimodal Llama 4 with Impressive Performance

Meta, the American tech giant, has released Llama 4, its most powerful open-source AI model. The initial release includes two models: Llama4Scout and Llama4Maverick. Llama4Scout boasts 109 billion parameters, 17 billion active parameters, and 16 experts. Its key feature is support for a 10 million context window, enabling processing of over 20 hours of video, all while running on a single H100 GPU (after Int4 quantization).