arch

Zhiyu announced the completion of approximately 5 billion USD in financing, including 2 billion USD in equity placement and 3 billion USD in zero-interest convertible bonds. The funds will primarily be used for the next-generation GLM model, self-training systems, and computing power infrastructure construction.

According to CryptoQuant analyst Darkfost, the recent rise in Bitcoin has driven investor sentiment to its most optimistic level in nearly two years, with related market sentiment indicators briefly breaking 89, entering the "extreme greed" zone, the last time a similar level was seen was in March 2024.Darkfost pointed out that when market sentiment is excessively optimistic or pessimistic, it is often worth paying attention to potential market turning points. Currently, this indicator has started to retreat from its previous euphoric levels, gradually returning to a relatively normal range, while the Bitcoin price is attempting to hold its current level.

Security researcher Shou Chaofan stated that he just purchased about 6TB of Fable model invocation data from a leading Chinese model relay station, which included sensitive credentials such as SSH keys, VPN configurations, Alibaba Cloud keys, and GitLab tokens. He claimed that the keys in this batch of data were sufficient for him to access the servers or internal systems of 19 leading Chinese companies, as well as 7 Chinese and CIS government-related agencies, including Huawei, Xiaomi, NIO, and MiniMax.The model relay station is positioned between users and models like Claude, so requests and responses pass through it first, allowing it to see the complete plaintext. Once developers insert SSH Keys, API Keys, VPN configurations, etc., into the Agent context, if the relay station saves or even sells these records, the company's system keys will also leak out. This is not the first time Shou Chaofan has warned about the risks of relay stations. A paper he participated in tested 428 LLM relay stations in April and found that 9 actively injected malicious code, 17 actually used AWS test keys deliberately placed by researchers to call AWS, and 1 directly transferred ETH from the test wallet.Shou Chaofan is a co-founder of the blockchain security company Fuzzland and has long been engaged in vulnerability and supply chain security research. At the end of March, he also first discovered that the source map in the Claude Code 2.1.88 release package accidentally exposed about 500,000 lines of TypeScript source code.

According to CoinDesk, the latest paper released by researchers from institutions such as Theta Labs, the Ethereum Foundation, and StarkWare shows that the resources required for potential quantum attacks on Bitcoin (BTC) and Ethereum (ETH) have decreased by more than 50% compared to the estimates previously published by Google's quantum artificial intelligence team.The quantum circuit design proposed in the study requires only about 1,151 logical qubits and uses approximately 1.3 million Toffoli gate operations to complete the relevant calculations, with a comprehensive score of about 1.5 billion, lower than the approximately 3 billion level previously published by Google.The research primarily optimized the point addition calculation in Shor's algorithm, which theoretically can use sufficiently powerful quantum computers to derive private keys from public keys. Jieyi Long, one of the main authors of the paper and CTO of Theta Labs, stated that quantum attacks are not imminent at the moment, but transitioning to quantum resistance will take years, and remediation cannot be done after an attack occurs.

Piper Sandler initiated coverage on Intel with a neutral rating and a target price of $110; gave Qualcomm a neutral rating with a target price of $190; and assigned an overweight rating to NVIDIA with a target price of $300.

According to "Du Jia," Alibaba Cloud is currently beta testing a new AI investment research tool called Qovest, positioned as an AI financial research assistant for professional individual investors. It is currently open for experience in batches to professional investors, with the first batch focusing on verifying A-share condition searches, Deep Research, and continuous tracking and research of asset accumulation.Qovest supports research on companies, industries, themes, and markets. After users pose questions, it can help find data, read financial reports and announcements, analyze research reports, generate reports, and consolidate research into the investment research space, continuously tracking subsequent changes in research subjects. The product standardizes common workflows for individual investors' fundamental research into modules, such as organizing the operational trends of a single company, horizontally comparing the business structures and financial differences of two targets, breaking down business models, and generating visual charts and research cards. The product is aimed at individual investors with independent research habits, but ordinary retail investors can also use it.

The research institution Sage Road Research released an executive summary of "The AI Trade," stating that since the beginning of this year, the Magnificent Seven has underperformed the Russell 3000 index by about 8 percentage points and the MSCI ACWI by nearly 9 percentage points. In July, the CBOE NDX volatility index relative to the VIX reached its highest point since the internet bubble, and after entering a correction, the Nasdaq index saw a 5% rebound over four days. As of the writing of the report, AI company stock prices have dropped 20% from their 52-week highs in June.Companies are struggling to achieve returns on investment amid soaring AI costs, with Uber, Amazon, Meta, and Walmart implementing restrictions on employee AI usage. Model homogenization limits pricing power, and Chinese open-source models have become a cheap alternative to OpenAI and Anthropic. AI capital expenditures have exceeded expectations, with the consensus for 2026 rising from $527 billion at the end of 2025 to about $800 billion by mid-year. Capital expenditures for hyperscale cloud providers in 2027 are expected to account for 3% of U.S. GDP, more than double the peak of 1.2% during the late 1990s telecom fiber construction. Allianz Research calculates that there is nearly a 46% growth gap between AI investment and sales, worse than the 32% during the 2001 telecom bubble.As of June, hyperscale cloud providers and related entities like Nvidia issued $225 billion in bonds, a year-on-year increase of 973.7%. The off-balance-sheet liabilities of tech giants have increased eightfold over four years to $1.65 trillion.

Research institution Shoal Research released a report stating that bots account for about 60% of HTML content HTTP requests on the Cloudflare network, while humans account for 38.2%. Cloudflare CEO Matthew Prince previously predicted that bot traffic would surpass human traffic by 2027, and later announced that the growth of agent traffic has already led to bots exceeding human traffic.The report states that AI applications keep users in chat interfaces and summarize content, rather than redirecting them back to the original site. AI crawlers have a much higher crawl-to-return ratio than Googlebot, with ClaudeBot averaging about 4,200 pages crawled per return visit over the past year. Pew research shows that AI summaries have reduced click-through rates from 15% to 8%.The robots.txt and content exchange traffic model are difficult to maintain. Litigation and private authorization are stopgap measures. Pay-per-request is divided into a closed market with keys and open protocols using HTTP 402 and stablecoins like x402 and MPP. Most websites still offer free access or block it.

According to on-chain analyst @ai_9684xtpa's monitoring, 40 minutes ago, two newly created wallets received airdrop contracts for 4,276 LAPTOPs from Hunter Biden's Substack subscribers, then sold them for approximately $404,000 and $243,000 respectively, totaling a profit of over $647,000.On-chain data shows that there is an ETH transfer association between the two addresses. Among them, the address 0x8DA...4A18d has transferred the proceeds from the sale in USDC to the public tagged address of Safe architect FloB (@FloB_Safe).

Dwarkesh Patel and Jerry Han published research on September 8, 2026, breaking down the contributions of data and model improvements in pre-training progress from 2019 to 2025. The study was conducted on a smaller scale, focusing on pre-training, corresponding to the publicly released model recipes and publicly available data corpora for each year, and was trained at a computational scale of up to 1e19 FLOPs.The results show that under the 1e19 FLOPs computational budget, the computational efficiency gain from data improvements is 12.0 times, while model improvements yield 3.7 times, with the data-side gain being approximately 3.24 times that of the model side. The gains from data and model improvements are largely independent and interact very little, with the additive effect of both explaining 88% of the variance in OLMES scores under a linear model.On the model side, the evolution from GPT-2 to OLMo-2 encompasses optimizers, positional encoding, normalization, activation functions, and initialization, among others. On the data side, the corpus evolved from approximately 9 billion tokens of OpenWebText in 2019 to larger and more finely filtered corpora like UltraFineWeb by 2025.

Anthropic researcher Jacob Coxon announced his resignation and is preparing to leave the AI industry. He has conducted pre-training research at OpenAI and Anthropic over the past three years. He pointed out that neither company has responsibly advanced AI and are competing to create superintelligence that can continue to enhance its own capabilities, "betting our lives on it."He switched from OpenAI to Anthropic this year because the latter places more emphasis on model safety. Coxon acknowledged that Anthropic's safety work is serious, but believes that as long as companies are racing to be the first to create AGI (artificial general intelligence that surpasses humans in a wide range of tasks), safety compromises are unavoidable, and self-discipline from a single company cannot solve the problem. His assessment of the two companies is: many people inside OpenAI still do not truly realize how great this risk is; Anthropic knows the risks but believes others are unreliable, so they must create it themselves first.Coxon stated that some AI practitioners are much more pessimistic privately than they express publicly, and are even worried that AI could destroy humanity before the end of this decade. He personally judges that in the most radical scenario, AI could go out of control by the end of next year, so such decisions should not be made solely by the engineers of a few companies.

Minara officially released the next-generation universal financial intelligence tool Minara Harness, which is now available for download by users worldwide. Minara Harness covers multiple asset markets, allowing for investment research 24/7 and operating automated trading within the scope authorized by users. Users can query financial markets using natural language, generate research reports, develop quantitative strategies, and handle daily tasks such as programming, web searching, and document processing. The platform supports users in selecting their own underlying large models.On the investment research side, Minara Harness can simultaneously call multiple professional AI Agents to analyze the same investment opportunity from perspectives such as fundamentals, valuation, market trends, and potential risks. Users can view the information sources and different judgments of each Agent and decide whether to execute trades based on the research results. Minara Harness also supports writing trading strategies in natural language, automatically generating code, and completing backtesting. Users can also subscribe to other traders' strategies, view strategy codes, historical performance, and real-time operating metrics. Minara also offers an optional decision review mechanism, which can verify past judgments against subsequent market results and assess which experiences can be applied to future research. At the same time, Minara announced the benchmark test results of Harness, which outperformed OpenClaw, Hermes, and Claude Code in multiple tests and demonstrated leading performance in tasks such as financial reporting, web research, and spreadsheets.

Tether CEO Paolo Ardoino revealed that Tether's P2P technology team is developing a censorship-resistant and fault-tolerant decentralized peer-to-peer search engine called Unstoppable Search Engine. The architecture supports horizontal scaling across thousands of distributed nodes and employs a fault-tolerant structure for data self-organization to maintain high availability and censorship resistance.Ardoino stated that the research and development is progressing smoothly, and the complete archived data of Wikipedia has been distributed and stored across 100 distributed nodes in the testing environment, achieving millisecond-level retrieval with excellent performance.

Galaxy Research posted on X that yesterday the daily transaction count for Bitcoin reached 893,391, marking the fourth highest single-day transaction count in Bitcoin's history, and this figure is above the 99th percentile of its historical range.

Binance Research's monthly market report shows that the total market capitalization of the cryptocurrency market increased by 17.6% to $2.7 trillion in August, with ETF inflows achieving the strongest performance since 2026.BTC rose by 24.8% within 7 days, ranking as the top weekly increase since 2020. In terms of user asset allocation, the proportion of cryptocurrency allocation among stockholders increased from 64% to 72%, while the allocation to stablecoins decreased by 22%; the proportion of TradFi perpetual contract trading volume dropped from 40% to 20%, with funds and trading activities clearly flowing back into cryptocurrency assets.

The Research Institute of Financial Technology at Renmin University published a compiled article on stablecoin research in the Journal of International Money and Finance. The related research is based on large language model (LLM) agents analyzing the de-pegging risks of stablecoins, which indicates that the de-pegging of stablecoins has significant nonlinear and threshold characteristics. The arbitrage mechanism can absorb a certain degree of informational pressure, but when the severity of the narrative reaches a "critical threshold," the system may quickly enter a state of sustained de-pegging.

Galaxy Research published that the attackers in the Coldcard "Wave 3" attack are still continuously transferring the stolen funds. During this phase, the attackers created 293 2-of-2 multi-signature wallets for each victim's assets. The first batch of funds was transferred across chains to Ethereum via THORChain; the latest round of transfers has begun entering the CoinJoin mixing process.Currently, the Wave 3 attackers are processing the largest amounts of stolen funds in order of the stolen amount, having sequentially transferred the funds from wallets ranked 1 to 11. The next 10 wallets that have not yet been transferred hold a total of 30.81 BTC, while wallets ranked 61 to 293 hold a total of 33.77 BTC. So far, the attackers have transferred about 45% of the stolen assets from this exploit, with funds flowing to Ethereum (via THORChain) or entering CoinJoin mixing transactions. Additionally, this fund transfer has revealed a previously unknown wallet: 58 addresses jointly spent in a 2-of-2 multi-signature format identical to that of Wave 3, and these were further transferred by the Wave 3 attackers to a jump address that funds CoinJoin.The on-chain analysis team currently marks this wallet as "cause = open," but believes it likely also belongs to Coldcard victims, which means the number of wallets involved in Wave 3 may increase to 294, raising the previously reported total amount stolen from the Coldcard vulnerability to approximately 1806 BTC. Currently, about 82% of the stolen BTC remains in addresses initially controlled by the attackers, while about 18% has been transferred, with the flow of funds indicating that it may be undergoing laundering processes.