BlackRock shifts to an overweight position in short-term US Treasuries for the first time in six years, and downgrades its rating on US equities.

Beating AI News Flash: OpenAI Product and Platform Lead Tibo Sottiaux announced Day 9 of the company’s 28-day product sprint, revealing an optimization to how ChatGPT and Codex desktop apps save local state data, which is expected to cut disk usage for this specific data by around 60%. It is worth noting that this reduction only applies to state data, not a 60% drop in the overall app storage footprint. Previously, many power users complained about Codex’s excessive disk usage. On Reddit, one user shared that their saved chat logs took up 98GB, with an additional 28GB used by the history database; another user noted Codex consumed roughly 50GB of space in just three weeks. Sottiaux confirmed the team received numerous complaints about insufficient disk space, prompting the rollout of this optimization.

According to Bitget market data, Hong Kong-listed Zhipu AI has been in a sustained downtrend, falling 1.83% intraday today to trade at HK$590, hitting a new low since March this year. On the news front, Anthropic released its new Claude Haiku 5.5 model on October 8, with market analysts noting that this model exerts "kill line"-like competitive pressure on domestic low-priced AI models. Additionally, on September 11 this year, Zhipu launched a roughly $2 billion share placement in Hong Kong and concurrently issued approximately $3 billion in convertible bonds. The company placed 21.97 million new shares at HK$714 per share; the current market price has fallen nearly 20% from the placement price.

Beating AI News (Dongcha) reports that independent technical author quanglee reviewed 51 public use cases of REA, finding that so-called "replication success" often only means restoring a single function. On October 12, he released his findings: 9 of the cases came from REA project contributors, all achieving their respective goals; among the other 42 independent attempts, 20 were fully completed, 16 partially completed, and 6 failed. The crux lies in the definition of "completion". Among the 20 independent successful cases, 9 targeted small games or mini-programs, 6 only restored one function of a larger software, and 5 did not even disclose the target name. One notable case is the puzzle game Untangle: after Agent analyzed its 167 KB program, it created a playable web version, with 52 out of 54 technical conclusions being fully accurate. However, developers later audited and found that 10 function and field names in the result did not come from REA’s analysis—likely because the model memorized the original’s open-source code. More complex projects require significant effort. In the reverse engineering of *Touhou Imperishable Night*, REA contributors matched 1106 out of 1107 functions to the original machine code, but the process took about two weeks and built on prior manual research. Another attempt to replicate voice input software Wispr Flow analyzed parts of its workflow (recording, transcription, text pasting) but only produced a 64-line Python script. The original’s core features—cloud-based models, context processing, real-time feedback—were not fully replicated. REA can help Agent call tools like Ghidra and IDA to analyze existing programs and rebuild partial functions. However, a commercial software’s capabilities are not limited to its installation package; server-side models, data, and complex business logic cannot be directly restored by decompiling the client.

Public security authorities in Bincheng, Shandong Province, China recently received a report from local resident Xie, who stated that 1,300 USDC stablecoins and NFT digital collectibles worth tens of thousands of yuan had vanished from his digital wallet. No obvious transfer alerts or phishing link traces were detected on the wallet. Police launched on-chain fund tracing by following the transfer records of the involved virtual currency, and actively collaborated with enterprises to analyze the suspect’s accounts. After extensive investigation, they identified the prime suspect Chen, a 30-year-old from Fengyang County, Chuzhou, Anhui Province. On October 9, technical police officers traveled to Fengyang County and arrested Chen. Upon being taken into custody, Chen initially denied the allegations. Police utilized the immutable and fully traceable nature of blockchain evidence to gradually break down his psychological defense, prompting Chen to eventually confess to the theft. Chen has been placed under criminal detention in accordance with the law, and his family has actively compensated the victim for the losses. This marks the first virtual currency theft case in Bincheng.

According to an official announcement, Binance will launch spot trading pairs for BBB/USDT, MDBB/USDT, PENGB/USDT, and UVXYB/USDT at 12:00 UTC on October 14, 2026, while simultaneously rolling out its spot algorithmic trading bot service. The rebalancing bot will be activated within 24 hours of the launch. Additionally, users can trade bStocks pairs with BTC, USDT, and other tokens on Binance's Spot Exchange with zero fees, and enjoy zero maker fees for the aforementioned trading pairs. This promotion is valid until 23:59 UTC on October 31, 2026. Withdrawals for BBB, MDBB, PENGB, and UVXYB will open at 13:00 UTC on October 14, 2026.

Bitcoin has pulled back from around $86,000 since early October, trading in a range of $81,000 to $84,000 in recent days, as the U.S. dollar strengthens and U.S. Treasury yields stay at high levels. Against this backdrop, Cloud, chief analyst at Huobi HTX, believes the range is more likely to break downward. The September ISM Services Prices Paid Index rose to 74.0, the highest since July 2022, while the University of Michigan’s 1-year inflation expectation for October rebounded to 4.7%. The concurrent rise in price sub-indexes and inflation expectations signals a high probability that this Thursday’s CPI will exceed market expectations. If that materializes, bets on a year-end interest rate hike will strengthen further, and the $81,000 lower bound will be tested first. Only if CPI comes in below expectations will a breakout above the $84,000 upper bound be sustainable. Recent sustained declines in chip stocks mainly stem from a repricing of AI capital expenditure timelines and hardware profit expectations within the sector, an industry-specific alpha event. In contrast, Bitcoin’s pullback has been far smaller, indicating its current pricing anchor leans more toward macro liquidity, with its connection to AI narratives weakening—their respective drivers are diverging. However, this divergence is temporary; once the U.S. dollar and long-term rates regain pricing dominance, their correlation will strengthen rapidly, with Thursday evening’s CPI release serving as the latest test window. Note: This article does not constitute investment advice, nor is it an offer, solicitation, or recommendation for any investment product.