An Open Source Alternative to GPT-4 Vision is Coming Soon

This article introduces LLaVA 1.5, a multimodal language model currently under development in the open-source community. It integrates multiple generative AI components, achieves high computational efficiency post-tuning, and can attain high accuracy across various tasks. LLaVA 1.5 employs CLIP as its visual encoder and utilizes the open-source LLaMA language model, connected through an MLP connector. It can outperform other open-source models in multimodal benchmarks with just approximately 600,000 training samples and one day of training time. Despite its usage limitations, LLaVA 1.5 represents the innovative direction of the open-source community and holds the potential to drive the development of open-source large models, offering users more convenient and efficient generative AI tools.

Civil Aviation Enters a New Era of Digital Intelligence: Five Departments Jointly Deploy and Promote the Application of Industry Large Models

Five departments jointly issued the "Action Plan for 'Artificial Intelligence + Transportation'", focusing on the intelligent transformation of civil aviation, encouraging the development of industry-specific large models to empower core aspects such as route planning, safety monitoring, and flight scheduling.

Domestic Large Models Trigger a New Wave of IPOs, Jieyue Star Pursues Hong Kong's Largest AI IPO in Recent Years

Chinese AI startup Jieyue Star is expected to submit its IPO application to the Hong Kong Stock Exchange this week, with a valuation as high as $12 billion, potentially becoming one of the largest IPOs in Hong Kong in recent years. The company was founded by Jiang Daxin, former vice president of Microsoft, in 2023. It has recently appointed Yin Qi of Megvii Technology as chairman and is accelerating its capital market strategy.

A-shares Enter the Era of Large Model Pricing! Zhipu Plans to Raise 15 Billion Yuan to Compete for a Listing on the Science and Technology Innovation Board

Chinese AI giant Zhipu accelerates its return to the A-share market, announcing on June 1st the recommendation to apply for a listing on the Science and Technology Innovation Board, planning to raise no more than 15 billion yuan, establishing a dual listing platform "A+H", issuing between 9.0988 million and 38.769 million new A-shares, to strengthen the long-term development capital foundation.

ByteDance's Doubao to Launch Paid Features, Accelerating the Commercialization of Large Models

ByteDance's AI product 'Doubao' plans to launch a paid subscription feature in late June, marking a shift from 'free acquisition' to 'paid monetization' for domestic large models. This move indicates that one of the largest-scale applications of large models has entered the deep waters of commercialization, accelerating the consumer-level AI market's departure from the free era. The pricing of Doubao's paid mechanism is affordable and relatively low, aiming to promote sustainable business model development.

How Can Large Models Truly Say Goodbye to Hallucination? Metacognition May Be the Key

The issue of 'large model hallucination' (outputting factual errors) is particularly severe in high-risk fields such as healthcare and law. The two mainstream methods in the industry to combat hallucination—expanding training data and setting up defense mechanisms—have limitations: data cannot cover all facts, and defense mechanisms often lead to AI being overly cautious.