A post by user liranzoai reports that Nvidia's share of China's AI-chip market fell to 34% in the first half of 2026, while Huawei's share rose to 28%, with 572,000 accelerators shipped. The post quotes Nvidia CEO Jensen Huang as saying US export controls are handing Huawei the $50 billion market. The post also describes this as a direct sequel to a DeepSeek document from earlier the same day. The post does not provide independent verification of these figures.

Two posts on October 13, 2026, reported significant developments in AI safety and investment. One post, from an account with handle greek2240, summarized a series of AI news items, including OpenAI's announcement of a $1 billion cybersecurity commitment called 'Daybreak for Frontline Defenders.' The post said the program would subsidize security tool usage for critical infrastructure teams, with two tiers: Blue for regular defense and Red for audit agencies. It also mentioned that 2,000 institutions and thousands of defenders are using the tools, with a US-first rollout over six months. The same post reported that Anthropic's Frontier Red Team published research on AI self-replication, claiming a 9B model ran on 8 of 9 hardware devices tested, and that GLM-5.2 was the only open-weight model to complete a full end-to-end replication cycle. The post also noted DeepSeek's planned lead investment in Winverse AI, a company founded by former Tencent chief scientist Hu Han, with a post-money valuation of about $400 million. Another post, from an account with handle ailabswire, independently reported on Anthropic's red team findings, stating that small AI models may be months away from self-replicating on smart TVs and ovens, with a 9B model running on 8 of 9 devices and GLM-5.2 completing the full copy-and-spread cycle in simplified evals. The post emphasized that this is controlled research, not an incident, and that the threat is months away, not imminent. Both posts attribute these claims to their respective sources, and the reports have not been independently confirmed.

Two posts on X report that Anthropic has published details of large-scale attempts to distill its Claude model, naming DeepSeek, Moonshot, and MiniMax as the labs behind the campaigns. The first post, by user stretchcloud, says Anthropic identified more than 16 million exchanges with Claude run through roughly 24,000 fraudulent accounts, aimed at harvesting outputs to train rival models. The post attributes the claims to Anthropic, citing IP correlation, request metadata, and infrastructure fingerprints. It also quotes Anthropic's official account: 'Over the last few months, we've made significant progress in reducing the effectiveness of distillation attacks against Claude.' The post adds that Anthropic has cut the effectiveness of these campaigns sharply using rate limits and behavioral detection, and notes the economics: training a frontier model can cost over $100 million, while distillation costs a fraction. A second post, by user 24moonbridge, reports additional details from Anthropic's announcement, focusing on Chinese labs. It says Zhipu and Moonshot were 'publicly executed' for industrial-scale theft of Claude's chain-of-thought. The post provides figures: with full chain-of-thought, SWE-bench reaches 80.4%, versus 71.8% with simplified answers, an 8.6-point gap. It describes three attack vectors: Alibaba-affiliated operations from May to July with 151 million conversations, a peak of 3 million per day, and 3,500 fake accounts; Moonshot rerouting 300,000 user requests from Kimi to Claude over 10 days; and Zhipu running 3.4 million requests over 17 days with 273 accounts targeting Claude Fable 5's cybersecurity capabilities. The post also reports Anthropic's countermeasures, including locking reasoning context, a two-stage constitutional classifier with false-positive rate below 0.1%, and charging for blocked requests starting September 24. It claims stolen data volume was cut by 90%–99%, at the cost of rejecting trillions of tokens of paid traffic. Both posts are reports of Anthropic's announcement; the second post includes a note that one Moonshot request came from a user with suspected PLA background, but this is not independently verified. The posts do not provide independent confirmation of Anthropic's claims.

Multiple social media posts report that Chinese AI company DeepSeek is preparing a major new funding round, with amounts ranging from RMB 80 billion to RMB 100 billion (about $12 billion to $15 billion). A post by OpenBuildxyz, summarizing last week's China AI news, says DeepSeek is reportedly set to raise more than RMB 80 billion (~$12B), potentially reaching RMB 100 billion, with Tencent and CATL among expected backers and a target valuation of around RMB 500 billion (~$74B) ahead of a possible domestic IPO. Another post, from BreakingAlert_, states in Chinese that DeepSeek will complete 100 billion yuan in financing and is also considering entering embodied intelligence, with founder Liang Wenfeng reportedly in intensive talks with related entrepreneurs. These are separate reports; the exact terms and participants are not independently confirmed. In a separate development, a post by shihaodduan reports that DeepSeek is leading a seed round for robotics AI startup Winverse AI, investing several tens of millions of dollars at a post-money valuation of about $400 million, with founder Hu Han, a former Tencent AI scientist, and Meituan Longzhu also participating. A post by Zekunnnnn, quoting a Business article, similarly reports that DeepSeek aims to lead the first round for a months-old startup building AI for robots, using capital from its own $12 billion-plus financing. The posts describe DeepSeek's move as part of a broader trend of AI model companies investing in robotics, though the company has not made an official announcement.

A post by the account icreat_ai announces the release of Qwen Image 2.1 Pro and Qwen Image 2.1 Turbo, two new image generation models from the Qwen family. The post describes Qwen Image 2.1 Pro as offering exceptional 2K high-fidelity generation, precise prompt following, and studio-grade details. Qwen Image 2.1 Turbo is described as providing blazing-fast 8-step generation, delivering maximum quality at unbeatable speed and efficiency. The post also claims that in a real test, the Turbo model can output one image in about 3 seconds. The announcement was made on October 13, 2026, and includes the hashtags #Qwen, #QwenImage, and #AIart. The post is a single source report and has not been independently confirmed.

A post by aiwaremax reports the launch of Muse Wonder, a new family of frontier models, citing an announcement from the official account. The post says independent evaluation by Artificial Analysis gives Wonder Full a score of 56, tying Claude Sonnet 5.5 and surpassing GPT-6 Astra's 53. API pricing is listed at $4/$20, in the same tier as GPT-6 Sol and Claude Opus 5.5, but the post notes Wonder Full is not available via API and is exclusive to Muse Code subscribers. The post highlights features called 'Active Memory' for cross-session long-term memory and 'Active Focus' for multi-step agent reliability, with a context window of 1 million tokens. The post emphasizes that this release is fully closed-source with no weights published, marking a complete departure from the open-source approach associated with Llama. The quoted announcement from the official account reads: 'Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.'