DeepSeek plans to train an 8-trillion-parameter model, whose scale is nearly three times that of the largest existing open model.

The Balancer community has proposed a "Fork and Reincarnate" plan, designed to continue Balancer’s technology and ecosystem via an official fork in the event the original protocol faces potential shutdown, with the new protocol positioned as a tokenized stock trading platform. The proposal suggests MAXYZ lead the establishment of a new official fork entity, migrating part of Balancer’s liquidity, team members, partners, users, and intellectual property to the new protocol, while extending the suspension deadline for existing pools and the Balancer Vault to Q2 2027. For funding, the plan allocates Balancer’s remaining ~6 million uncirculated BAL tokens as seed capital for the new project, valued at roughly $690,000 at current prices. In exchange, if the new protocol conducts a future token issuance or other liquidity exit events, the Balancer Treasury will receive 10% of the new protocol’s token’s fully diluted valuation (FDV) upfront. The new protocol will leverage Balancer’s core technologies—including multi-asset pools, dynamic weights, dynamic fees, LVR capture, and permissioned pools—focusing on tokenized stocks and other traditional financial assets on-chain. Using the U.S. stock market’s average daily trading volume of $750 billion as a benchmark, the proposal estimates that if 10% of that volume moves on-chain, with a 1 basis point fee and 1% market share, theoretical annual revenue would reach approximately $27.4 million. Additionally, the proposal intends to grant the new fork a permanent, irrevocable, non-exclusive license for Balancer-related intellectual property; if the original Balancer entity is dissolved, this license will be upgraded to an exclusive assignment of the held intellectual property.

Exclusive: Beating AI News Flash OpenAI has established a "Mathematics and Artificial Intelligence Advisory Group" hosted by Princeton’s Institute for Advanced Study (IAS). The company stated that an internal research model has solved over 100 long-standing open problems spanning multiple mathematical fields, including the previously announced research on the Navier-Stokes Millennium Prize Problem. Most of these findings remain unpublished. The advisory group will assess the significance of these results, their publication approach, and compliance with academic standards. Its first nine members include three Fields Medal winners: Timothy Gowers, Martin Hairer, and Edward Witten. Members will not receive compensation from OpenAI, may publicly criticize the company, and can independently release recommendations. However, the group holds no decision-making power and will not determine the pace of OpenAI’s internal mathematical research. This move is OpenAI’s direct response to recent dissatisfaction from the mathematics community. Earlier, 25 Fields Medal winners including Terence Tao, Yufei Zhao, and Peter Scholze released a joint statement accusing AI companies of framing the solving of major math problems as a benchmark, which has begun to harm mathematical research. The statement acknowledged that large models have advanced rapidly in recent months and can solve several important open problems across fields, but emphasized that mathematics prioritizes the new concepts, methods, and understanding generated from such solutions.

According to EmberCN’s monitoring, a crypto whale or institutional entity has recently liquidated its holdings of 31.4 million CRV tokens held for roughly three years, incurring a total loss of around $4.1 million, equivalent to a 27% loss rate. The address in question withdrew and accumulated the 31.4 million CRV from Binance between 2023 and 2024, with a total investment of approximately $15.13 million, translating to an average cost of $0.48 per token. Over the past two weeks, the address has transferred these CRV to OKX for sale, fetching an average selling price of roughly $0.35 per token.

Tom Lee (@fundstrat)'s #Bitmine keeps buying $ETH, withdrawing another 12,500 $ETH($34.55M) from #Kraken today.

Beating AI News Insight: Alibaba CEO Wu Yongming unveiled the company’s next-phase AI roadmap at the Yunqi Conference. The Qwen team plans to train a new model with 5 trillion to 10 trillion parameters; Pingtouge released the Zhenwu V900; and Alibaba Cloud aims to expand its global data center capacity to over 20GW by 2032. Qwen is exploring Recursive Self-Improvement (RSI), which enables the model to identify its own shortcomings based on real task feedback, design experiments, construct data, and continue training. The new generation of models will focus on enhancing capabilities for complex long-horizon tasks, while advancing multimodal models that integrate understanding and generation. Wu Yongming estimates that the total thinking capacity generated by machines today is less than 3% of humans, and could exceed human levels by 1,000 times in the future. He noted that current AI coding is still similar to the "electric light in 1882"—it mainly replaces existing work, while new products truly belonging to the machine intelligence era have not yet emerged. In the future, complex tasks may be split into tens of millions of subtasks, assigned to millions of Agents for continuous execution. To support this scale of computing demand, Alibaba is also strengthening its chip and cloud infrastructure. The Zhenwu V900 delivers 3 times the performance of the M890, with a single cluster scalable to up to 500,000 cards. Wu Yongming stated that medium- and long-term demand for AI currently exceeds supply, and Alibaba Cloud plans to operate over 20GW of global data centers by 2032.

Beating AI Express (from Dongcha) – AMD’s stock closed up nearly 10% on September 21 to $615.52, hitting an all-time high. The chipmaker’s market capitalization crossed $1 trillion for the first time, making it the fourth U.S. semiconductor firm to reach this valuation, following NVIDIA, Broadcom, and Micron. AMD has rallied roughly 24% over the past five trading days, with its year-to-date gain exceeding 180%. One direct catalyst for the recent surge is the market’s re-evaluation of AI agents’ demand for CPU computing power. Meta’s newly launched personal AI agent Muse quickly gained traction: it amassed 730,000 downloads in just five days and topped the U.S. Apple App Store’s free app chart on September 18. Market participants note that as AI agents become more widespread, server CPUs will see increased workloads in inference, task orchestration, and infrastructure operations. AMD’s second-quarter revenue rose 50% year-over-year to $11.54 billion, with its data center business generating $6.7 billion in revenue, up 107% YoY. Meta currently accounts for around 5.5% of AMD’s total revenue, ranking as the chipmaker’s second-largest customer. As AI applications expand beyond model training to agent deployment, market focus on AI infrastructure is shifting from GPUs to CPUs and other general computing resources.