24H Flash Updates

According to monitoring, mathematician Levent Alpöge from Anthropic used Claude Fable to discover a three-dimensional counterexample to the Jacobian Conjecture. This problem, proposed in 1939, remained unsolved for decades and was listed among the major mathematical challenges of the 21st century. The polynomial mapping provided by Claude Fable has a Jacobian determinant identically equal to −2, yet maps three distinct inputs to the same output, thus proving it non-invertible—thereby refuting the conjecture in three dimensions and higher. The two-dimensional version remains unresolved. Subsequently, OpenAI researcher Aaron Lou conducted an offline experiment using internal Codex, which independently discovered an essentially identical counterexample and derived the proof process. The original text also notes that AI can be leveraged to parallelize the search for counterexamples, vulnerabilities, and novel insights in open problems; however, it may also mass-produce seemingly rigorous but erroneous proofs, increasing the burden on academic verification.

Reports indicate that Samsung's stock price rose after the company established a robotics business division, signaling Samsung's move to advance robotics-related operations within a more clearly defined organizational structure. The original information primarily focuses on Samsung's newly established robotics division and market reaction, without detailing specific business scopes, product plans, or investment scales.

It is reported that Lu Siyuan, head of Xpeng Automotive's AI Infrastructure Department, is set to leave the company and is currently in the process of handover. His next move will be to OpenAI, where he will participate in the development of embodied intelligence robots. At Xpeng, Lu was responsible for the entire AI infrastructure pipeline from model training to mass vehicle deployment, directly managing a team of around 200 people, with work spanning training frameworks, GPU clusters, self-developed chip compilers, model quantization optimization, and in-vehicle deployment.

Given Xpeng’s in-house chip strategy, both compilers and inference runtime systems must be built internally, making this role critical to core infrastructure capabilities. With no successor identified yet, the team is expected to be split into three or more independent units. OpenAI’s official website is currently recruiting for robotics software, simulation, and firmware engineers. The company states that its robotics team is developing general-purpose robots capable of operating in real-world environments, integrating models, hardware, and software seamlessly.

According to monitoring, Kimi K3 has achieved a comprehensive net improvement of 9.62% on the Arena Agent leaderboard, ranking 4th, following Claude Fable 5, Claude Opus 4.8 Thinking, and GPT-5.6 Sol. The leaderboard is calculated based on real user tasks and tool invocation records, averaging five net improvement metrics to determine the composite score. Kimi K3 has currently accumulated 8,344 test sessions, with a net improvement of 14.42% in user-confirmed success metrics—ranking first—and a 20.62% improvement in the ratio of praise over complaints—ranking third. In contrast, its error correction execution ranks 14th, and Bash error recovery ranks 17th.

It is reported that Oracle may face a collateral bill of approximately $7 billion related to its Wisconsin data center project. The matter involves financial arrangements tied to data center construction and potential guarantee pressures, with amounts reaching several billion dollars. Data centers are critical infrastructure for AI computing power and cloud services; associated capital expenditures, financing conditions, and collateral requirements directly impact project execution costs and corporate balance sheet management.

According to reports, a senior executive from Tencent Cloud recently stated at a forum during the World Artificial Intelligence Conference that Tencent will scale up deployment of domestically-produced computing power in China to further reduce AI inference costs. The executive also mentioned that the company expects to deploy NPO (Near-Package Optics) super nodes by Q4 2026. Regarding related infrastructure development, Tencent Cloud has called for global and domestic alignment on NPO industry standards. This statement touches upon AI inference cost optimization, domestic computing power deployment, and evolution of optical interconnect infrastructure.

According to reports, Sysdig researchers linked a second attack on the same Langflow server to the previously disclosed AI Agent-driven threat actor JADEPUFFER and discovered the deployment of a new Go-compiled ransomware, ENCFORGE. The attack continues to exploit the unauthenticated /api/v1/validate/code endpoint in Langflow versions prior to 1.3.0; the vulnerability, CVE-2025-3248, carries a CVSS score of 9.8 and was added to CISA’s Known Exploited Vulnerabilities catalog on May 5, 2025.

ENCFORGE targets AI infrastructure files such as model weights, vector indexes, and training datasets, defaulting to encrypting approximately 180 file extensions including PyTorch, TensorFlow, SafeTensors, ONNX, GGUF, FAISS, and Parquet, using AES-256-CTR encryption and appending a .locked suffix. Researchers noted that the sample does not include capabilities for data exfiltration, cloud storage, or Tor-based payment portals, and the ransom contact email matches those used in prior activities. Additionally, the attackers leveraged exposed Docker sockets to create privileged containers and mount the host filesystem, attempting to break out from the container to execute encryption on the host system.

It was reported that Tencent has transferred relevant operations and part of the team from the QClaw Product Center into Department 6 of Cloud Products, under which QClaw will continue operations and be grouped together with WorkBuddy. Department 6 of Cloud Products was established in April this year, responsible for the development and commercialization of AI-native productivity products such as CodeBuddy and WorkBuddy.

QClaw is built upon OpenClaw, with a product positioning more oriented toward individual users and remote computer control scenarios; WorkBuddy, on the other hand, focuses on workplace and enterprise office environments. This restructuring signifies further organizational consolidation of Tencent’s internal AI agent product lines.

According to Beating monitoring, following the surge in popularity of Kimi K3, debates within the U.S. AI community over open-source models from China have escalated under the Trump administration, with proponents advocating restrictions or even bans pushing to translate their stance into concrete policy. Previously, the U.S. Department of Commerce considered adding several Chinese AI labs to the Entity List, while the White House discussed imposing security liabilities on U.S. companies hosting Chinese models—however, these proposals were shelved last year due to concerns over stifling innovation.

As officials supportive of open competition gradually exit, the influence of national security hardliners has grown, and Kimi K3’s capabilities and momentum have revived these proposals on the policy agenda. Even if the U.S. does not directly announce a ban, procurement restrictions, security advisories, and Entity List threats could compel enterprises to pivot toward alternative models.

According to Beating monitoring, Google is developing the Frozen v2 AI inference chip, with deployment planned as early as 2028. The chip will directly hardwire parts of the Gemini architecture into silicon, reducing computational overhead and data movement during model inference.

Google internal projections indicate that Frozen v2 will achieve 6 to 10 times higher token throughput per watt compared to the latest TPU. Google aims to alleviate the ongoing compute scarcity, a shortage that has already forced Google Cloud to decline certain external customer orders. Frozen v2 will coexist alongside TPUs: while Frozen v2 prioritizes efficiency at the cost of flexibility, TPUs maintain versatility by supporting diverse model workloads.

According to data from Beating, alternative data and market research firm YipitData estimates that Anthropic's annualized revenue has reached $79.5 billion. This figure was $69 billion at the end of June, representing an increase of $10.5 billion within approximately three weeks. On a monthly basis, Anthropic saw increases of $10 billion, $11 billion, $14 billion, and $15 billion from March to June, respectively, with monthly growth consistently exceeding $10 billion. Monthly growth for this month is also expected to surpass $10 billion.

According to Beating monitoring, the two-year-long collective copyright lawsuit against Anthropic has officially concluded, with a U.S. federal judge approving a $1.5 billion settlement—setting a record for the largest known copyright settlement in the United States. This case also marks the first major proceeding in U.S. AI training copyright litigation to complete the full cycle of trial, ruling, and settlement.

The court previously ruled that using books to train Claude constituted fair use; however, Anthropic’s downloading and retention of over 7 million pirated books still amounted to infringement. More than 91% of affected authors and publishers have filed claims for compensation, while some rights holders who withdrew from the settlement remain pursuing separate litigation against Anthropic.

According to data from Beating, Tencent Hunyuan has launched Hyra-1.0, a research-oriented AI agent targeting tasks such as model development, mathematical discovery, and engineering optimization. Hyra employs a lightweight iterative mechanism, enabling multiple Agents to generate solutions in parallel. The system operates within isolated sandboxes, evaluates results, and then writes back code, logs, and feedback into an experience repository;

When no pre-existing evaluator is available, it also synchronously refines the scoring criteria, reducing the likelihood of agents exploiting rule loopholes to artificially inflate scores. Tencent reports that Hyra has achieved record-breaking results on 29 out of 55 collected open mathematical problems, reduced the time for NanoGPT to reach target loss to just 76.4 seconds, and designed a Transformer with only 15 trainable parameters capable of performing addition on 10-digit numbers.

According to reports, during pre-market trading on Monday, Alibaba's U.S. shares rose approximately 5%, placing it among the top gainers in the pre-market session. Meanwhile, U.S. equity index futures showed divergent movements as investors weighed market uncertainty stemming from escalating Middle East conflicts against the backdrop of a dense schedule of major tech company earnings reports this week.

The report highlighted Alibaba as a standout performer in pre-market gains but did not specify the full catalysts at the corporate level. This movement reflects ongoing pre-market volatility in certain large-cap tech-related stocks amid concurrent concerns over geopolitical risks and expectations surrounding upcoming tech earnings releases.

Reports indicate that Chinese quantitative hedge funds experienced significant drawdowns last week, with domestic market declines deepened further following global semiconductor stock sell-offs. Affected by market volatility, multiple quantitative products targeting index enhancement struggled, among them a quant fund managed by the founder of DeepSeek, which aims to outperform the CSI 1000 Index, recording a drawdown of approximately 16%. This event illustrates how the global downturn in semiconductor stocks, once transmitted to China’s local markets, not only dampened trading sentiment for related equities but also amplified net asset value volatility in quantitative strategies during rapid market selloffs.

It is reported that Prysmian has signed an agreement with Molex, with a maximum value of €5.5 billion, covering the supply of data center optical cables. The core focus of this agreement is optical communication cables tailored for data center environments, serving as foundational infrastructure supporting server clusters, network interconnectivity, and high-bandwidth data transmission.

As AI training, inference, and cloud services demand increasingly higher capabilities in intra-data center and inter-facility connectivity, the importance of underlying network components such as optical cables continues to rise within compute infrastructure. With the transaction cap reaching €5.5 billion, this underscores the substantial scale of related supply chain orders, positioning Prysmian and Molex’s collaboration around the evolving connectivity needs of data centers.

According to reports, Trend Micro analyzed 200 Gemini CLI session logs from March 19 to April 21, 2026, revealing that a Russian-speaking threat actor known as "bandcampro" abused Google's open-source Gemini CLI to conduct cyber operations and control a small botnet. The actor leveraged AI to crack passwords, set up residential proxies, compromise WordPress merchants, and plan phone-based cryptocurrency scams targeting elderly individuals in the United States and Canada. Logs indicate the actor also used AI to migrate C&C servers, gain control over eight computers at a dental clinic, and access the OpenDental database. The entire C&C operation was orchestrated through approximately 3 plaintext files totaling around 5KB. The AI proactively offered 59 improvement suggestions without being prompted and completed the migration to a new architecture along with error remediation within six minutes.

It was reported that DeepSeek offered a pre-tax daily salary of 5,500 RMB to a Tsinghua University Yao Class intern, amounting to a monthly salary exceeding 120,000 RMB based on a 22-day work month. DeepSeek has not yet publicly confirmed this compensation information.

In publicly listed internship positions, DeepSeek’s typical daily pay ranges from 500 to 1,000 RMB. Regarding the issue of salary inversion triggered by high-paying recruitment, internet rumors suggest that if a new hire’s salary surpasses that of senior team members, multiple senior employees must evaluate and score the candidate, with approval only granted upon achieving a passing threshold. However, these details remain unverified and are currently circulating solely as online speculation.

According to reports, Tian Yuandong, co-founder of Recursive and former director of Meta FAIR research, recalled that Wu Yuxin, co-founder of Moonshot AI, was once his intern in 2015. To persuade his superiors to hire Wu Yuxin, Tian spent an hour arguing on the rooftop of a Facebook office building, ultimately betting his own reputation on the decision.

Subsequently, Wu joined FAIR, where he co-developed Group Normalization with He Kaiming and created Detectron2. Today, he is involved in the development of Kimi’s foundational model at Moonshot AI. Tian also emphasized that in the era of large models, technical doers who personally inspect data, write code, and rapidly iterate through experiments are more critical than ever—even founders must be hands-on.

According to market analyst qinbafrank, merely releasing more large models, disclosing higher token usage, or continuing to increase capital expenditures is no longer sufficient to alleviate market concerns about AI investment returns. He believes the market is now focused on a set of fundamental signals that corroborate one another, including sustained expansion of AI revenue streams, concurrent growth in enterprise customers, production-grade workloads, paid seats, renewal rates, and average spend per customer—while demand remains robust even after excluding a few major clients such as OpenAI and Anthropic.

Meanwhile, incremental gross profit from AI must continue to grow faster than rising costs tied to depreciation, energy, network infrastructure, and talent acquisition. As unit inference costs decline, total gross profit should keep increasing. Investors will also closely monitor whether order backlogs can accelerate into revenue over the next 12 to 24 months, whether utilization rates and customer consumption exceed initial commitments following the rollout of new compute capacity, and whether in-house chip development and model optimization deliver verifiable economic benefits. For the upcoming second-quarter earnings report, the market anticipates AI and cloud revenue exceeding expectations, margins remaining broadly stable, capital expenditures under control, and free cash flow ceasing its deterioration; if revenue quality, gross margin, and cash flow fail to improve, concerns may intensify further.