AI News

Anthropic released a novel Natural Language Autoencoder (NLA) that converts digital 'activations' inside its language model Claude into human-readable text, addressing the challenge of interpreting model internal states. This technology opens new doors for model interpretability, making AI's 'thinking processes' more transparent.....

Fields Medalist Professor Gowers solved a long-standing combinatorial math problem in one hour using the unreleased ChatGPT 5.5 Pro, challenging the traditional belief that large models merely memorize rather than truly understand advanced mathematics, sparking a reevaluation of AI's mathematical capabilities.....

Elon Musk's legal battle with OpenAI opened in a California court, with co-founder Brockman testifying for two consecutive days. His private diary from around 2017 was presented as key evidence, revealing early power struggles and strategic shifts within the company.....

A TechPowerUp survey of nearly 20,000 gamers shows 58% oppose AI altering game visuals like lighting, textures, and character faces, preferring original graphics. Only 8% believe DLSS 5 renders better than native.....

Fields Medalist Timothy Gowers shares his experience using the unreleased ChatGPT 5.5 Pro for mathematical research. The AI solved an open combinatorial problem within an hour, sparking academic interest in AI's mathematical capabilities. Gowers notes that large language models have evolved from finding known information to solving open problems.....

The domestic large model sector has witnessed a capital storm, with DeepSeek (Deep Seek) initiating its first major funding round targeting an amount of 5 billion yuan, which, if successful, would break the industry record. Notably, the lead investor is not a venture capital firm or an internet giant, but the founder Liang Wenfeng himself, who has invested the highest personal amount, demonstrating his strong confidence in the company.

DeepSeek launched a large-scale beta of its 'Image Recognition Mode' on May 9, marking its entry into multimodal interaction. After a limited gray test in late April, most accounts can now access it via a standalone entry in the chat interface. Though still in beta, it's listed alongside 'Fast Mode' and 'Expert Mode,' signaling multimodal understanding as a core direction.....

SoftBank Group reduced its planned $10 billion loan secured by OpenAI shares by 40% to $6 billion, with a two-year term and a one-year extension option. The financing cut reflects creditor concerns over pricing stability, disposal efficiency, and liquidity of unlisted OpenAI shares.....

Oracle laid off 20,000 to 30,000 employees via email on March 31. Affected workers only realized their unemployment when unable to access VPN. Severance terms were controversial, requiring employees to waive litigation rights and receive only 40% of compensation in the first year.....

A technical report from XiYu Technology reveals that the M2 series model fails to accurately output specific names like 'Ma Jiaqi' due to a 'token offset' issue caused by the tokenizer. The model splits the name into 'Ma' and 'Jiaqi', compressing the vector space and causing recognition bias. This exposes a common yet subtle flaw in current large model training, affecting precise generation of specific names.....

DeepSeek will release an updated V4.1 model in June, accelerating release frequency to catch up with competitors. Despite praise for technical depth, the company lagged with no new models for 140 days while rivals released about 50. The V4, originally scheduled for February 2026, was delayed to April 24 due to hardware migration issues and launched as a preview.....

The MiniMax M2 series model exhibits anomalies when outputting 'Ma Jiaqi', caused by low-frequency token degradation during post-training. The tokenizer splits 'Ma Jiaqi' into [Ma, Jiaqi], where 'Jiaqi' was merged as an independent token (ID 190467) due to high pre-training frequency but was 'squeezed' in subsequent training, leading to output errors. The official technical report has revealed this mechanism.....

OpenAI launches Codex Chrome extension for Mac and PC, streamlining browser tasks. It boosts web efficiency by integrating with GitHub, Slack, and Figma, filling gaps in previous sandbox and plugin approaches.....

Baidu launches the new generation Ernie Bot 5.1, featuring 'multi-dimensional elastic pre-training' technology, emphasizing high cost-effectiveness and strong search integration. Now available on Baidu Qianfan Model Plaza and Ernie Bot official website, it offers open experience for enterprise users and developers, showcasing a breakthrough in domestic large models amid AI competition.....

On May 8, China Mobile launched the first ultra-large-scale AI model service platform in China, integrating over 300 mainstream AI models, making it the most extensive integrated platform domestically. It pioneers a 'token-intensive' operational model for intelligent matching, marking a significant breakthrough in China's AI infrastructure and service capabilities.....

In the first quarter of 2026, Sanqi Interactive achieved a significant gain of 359 million yuan from its investment in Zhipu, an AI company that went public, with its stock price surging. This represents a 981.35% year-on-year increase. This achievement highlights the recognition of AI assets by the capital market, marking a phased success in the company's competitive positioning in the field of artificial intelligence, and continuing to build a portfolio of investments in large AI models.

StepAudio 2.5 Realtime, a new real-time voice model by StepStar, is now fully online for developers. It enhances paralinguistic perception, character customization, and dialogue capabilities for more realistic interactions, innovatively processing tone, speed, pauses, and sighs to boost conversational naturalness.....

OpenAI launches Codex for Chrome extension, marking AI's first deep integration into web operations. With authorization, Codex reads web pages, manipulates internal systems, and fills complex forms like a human, transitioning AI from dialogue boxes to system-level applications.....

Google's mobile Gmail app recently added an 'AI Inbox' shortcut at the bottom, simplifying access to the 'Help me write' generative AI feature to boost reach and accelerate AI adoption in daily productivity. Available on Android and iOS, this update optimizes user interaction logic.....

Cloudflare announces layoffs of approximately 20% (1,100 employees) despite revenue growth. CEO Matthew Prince cites AI-driven efficiency gains, not performance issues, enabling leaner operations. Similar to Meta and Microsoft, the company optimizes workforce during high growth. Cloudflare's cybersecurity services cover millions of websites globally.....