The reproduction of state-of-the-art multimodal LLM pre-training faces barriers at every stage of the pipeline, including high-quality data filtering, multimodal data mixture strategies, sequence packing techniques, and training frameworks. We introduce Open-Qwen2VL, a fully open-source 2B-parameter Multimodal Large Language Model pre-trained efficiently on 29M image-text pairs using only 220 A100-40G GPU hours. Our approach employs low-to-high dynamic image resolution and multimodal sequence packing to significantly enhance pre-training efficiency. The training dataset was carefully curated using both MLLM-based filtering techniques (e.g., MLM-Filter) and conventional CLIP-based filtering methods, substantially improving data quality and training efficiency. The Open-Qwen2VL pre-training is conducted on academic level 8xA100-40G GPUs at UCSB on 5B packed multimodal tokens, which is 0.36% of 1.4T multimodal pre-training tokens of Qwen2-VL. The final instruction-tuned Open-Qwen2VL outperforms partially-open state-of-the-art MLLM Qwen2-VL-2B on various multimodal benchmarks of MMBench, SEEDBench, MMstar, and MathVista, indicating the remarkable training efficiency of Open-Qwen2VL. We open-source all aspects of our work, including compute-efficient and data-efficient training details, data filtering methods, sequence packing scripts, pre-training data in WebDataset format, FSDP-based training codebase, and both base and instruction-tuned model checkpoints. We redefine "fully open" for multimodal LLMs as the complete release of: 1) the training codebase, 2) detailed data filtering techniques, and 3) all pre-training and supervised fine-tuning data used to develop the model.

Text semantic segmentation involves partitioning a document into multiple paragraphs with continuous semantics based on the subject matter, contextual information, and document structure. Traditional approaches have typically relied on preprocessing documents into segments to address input length constraints, resulting in the loss of critical semantic information across segments. To address this, we present CrossFormer, a transformer-based model featuring a novel cross-segment fusion module that dynamically models latent semantic dependencies across document segments, substantially elevating segmentation accuracy. Additionally, CrossFormer can replace rule-based chunk methods within the Retrieval-Augmented Generation (RAG) system, producing more semantically coherent chunks that enhance its efficacy. Comprehensive evaluations confirm CrossFormer's state-of-the-art performance on public text semantic segmentation datasets, alongside considerable gains on RAG benchmarks.

The frenzy to create Ghibli-style AI art using ChatGPT's image-generation tool led to a record surge in users for OpenAI's chatbot last week, straining its servers and temporarily limiting the feature's usage. The viral trend saw users from across the globe flood social media with images based on the hand-drawn style of the famed Japanese animation outfit, Studio Ghibli, founded by renowned director Hayao Miyazaki and known for movies such as "Spirited Away" and "My Neighbor Totoro." Average weekly active users breached the 150 million mark for the first time this year, according to data from market research firm Similarweb.

Want to talk with Copilot? If you do, Microsoft is testing an intuitive little tweak to allow you to do so. On March 10, Microsoft launched a new shortcut that is rolling out as part of the Copilot app on Windows: a "long press" combination of Alt SPACE as a press-to-talk shortcut to open the Copilot app. But who would remember that? Now, Microsoft is testing a smarter alternative as part of the Windows Insider Beta Channel, launching a "press-to-talk" voice chat by holding down the Copilot key for two seconds instead.

ChatGPT's built-in image generation feature is now available to everyone. OpenAI CEO Sam Altman said last week that the company is delaying its rollout to free tier "for a while," because the tool was way more popular than they had expected. But the company made the feature available to free users over the weekend, allowing them to generate images from within ChatGPT and without having to switch to OpenAI's DALL-E generator. Prior to its rollout to the free tier, the tool was only available to Plus, Pro and Team subscribers. Altman previously said that free users will get a limit of three images per day.

My in-laws own a little two-bedroom beach bungalow. It's part of a condo development that hasn't changed much in fifty years. The units are connected by brick paths that wind through palm trees and tiki shelters to a beach. Nearby, developers have built big hotels and condo towers, and it's always seemed inevitable that the bungalows would be razed and replaced. But it's never happened, probably because, according to the association's bylaws, eighty per cent of the owners have to agree to a sale of the property.

The fear that artificial intelligence (AI) will replace millions of jobs is widespread. But equally, in today's tough job market, not using AI wisely as part of your search could mean you miss out. You can use AI models such as ChatGPT and Perplexity to research employers, competitors and industry trends before applying for a job. Hannah Salton, a careers coach, says some of her clients have successfully used AI to find out more about companies, allowing them to "gain insights into culture, competitors and market positioning. It can also help identify SMEs [small and medium-sized enterprises] to apply to or network with."

Fox News anchor Bret Baier has the latest on the Murdoch Children's Research Institute's partnership with the Gladstone Institutes for the'Decoding Broken Hearts' initiative on'Special Report.' The Trump administration recently asked American developers, including OpenAI, for input on what the U.S. needs to do to stay ahead in the global AI competition. We believe that preserving AI's ability to learn should be at the top of the list. Today, artificial intelligence is poised to scale human ingenuity itself–the sum of our freedoms to learn and know, think, create, and produce. Humans have never created a technology that can do as much to advance education, science, and discovery–and we're already seeing its benefits.

OpenAI on Monday said it raised 40 billion in a new funding round that valued the ChatGPT maker at 300 billion, the biggest capital-raising session ever for a startup. The infusion of cash comes in a partnership with Japanese investment giant SoftBank Group and "enables us to push the frontiers of AI research even further," the San Francisco-based company said in a post on its website. "Their support will help us continue building AI systems that drive scientific discovery, enable personalized education, enhance human creativity, and pave the way toward AGI (artificial general intelligence) that benefits all of humanity," the company said.

OpenAI said it had raised US 40bn in a funding round that valued the ChatGPT maker at 300bn – the biggest capital-raising session ever for a startup. It comes in a partnership with the Japanese investment group SoftBank and "enables us to push the frontiers of AI research even further," OpenAI announced, adding it would "pave the way toward AGI (artificial general intelligence)" for which "massive computing power is essential". SoftBank said it wanted to realise "artificial super intelligence" (ASI) surpassing human intelligence and OpenAI was the partner closest to achieving that goal. SoftBank is to put 10bn at first into OpenAI and 30bn more by the end of 2025 if certain conditions are met. Also on Monday, OpenAI announced it was building a more open generative AI model as it faces growing competition in the open-source space from DeepSeek and Meta.