Macro traders placed a $31.7 million short position on the Nasdaq ahead of the CPI.

Beating AI Express News: Tencent has dedicated its Qwen3.5-122B-A10B model to training a terminal agent, resulting in T1. The agent is designed to handle complex tasks in Linux terminals, with each task capable of calling tools consecutively for over 300 rounds at maximum. Its score on Terminal-Bench 2.1 rose from 43.8% for the base model to 64.0%, an increase of 20.2 percentage points. The bulk of this gain came from reinforcement learning: SFT alone lifted the score to 49.4%, while subsequent reinforcement learning added another 14.6 points. The team prepared around 15,000 terminal tasks, each paired with an automated test. Even if the agent does not complete the entire task, it can earn corresponding rewards for meeting partial requirements. Long tasks present another challenge: when the agent performs actual work and later uses that experience for training, the same content may be re-split into different tokens, and the Mixture of Experts (MoE) may switch to a different set of experts. T1 records both the tokens generated at the time and the selected experts, and replicates these during training, reducing this training-inference gap by approximately one-third.

According to official news, the community of STARLINK, a meme token in the X Layer ecosystem, opened its community donation charity fund at 9:00. STARLINK launched on September 9 on Ignix, X Layer’s general launch platform, with SPCX as the pricing token for pool pairing. Transaction taxes are allocated to the vault and distributed to holders in SPCX; as of press time, its market capitalization exceeds $7.5 million. Per its announcement, the community fund is managed via a fixed address and multi-sig wallet. The principal of tokens donated to the fund remains untouched—actual expenditures are funded solely by dividends generated from these holdings, which will be used to purchase Starlink satellite internet equipment and deploy it to areas without internet access, including disaster-hit regions and network deserts. All fund inflows and outflows are verifiable on the X Layer blockchain. The project team stated it will continuously disclose donation lists, every equipment procurement voucher, deployment locations, and reconciliation records of dividend income and deployed device counts. As of press time, the fund’s donation address has received a total of 350 million STARLINK, equivalent to approximately $2.84 million. Ignix is X Layer’s general launch platform, allowing creators to customize buy and sell taxes, select vault templates, and bind AI Agents for launches. BlockBeats reminds users that related token prices are highly volatile with significant uncertainty, so users should exercise caution when investing.

According to an official announcement, 4D Labs, a real-world embodied data layer protocol for physical AI, has secured investment from YZi Labs via Season 2 of the EASY Residency program. Relevant details show that 4D Labs is a spatial intelligence large model infrastructure company at the intersection of AI, 3D, and Web3, with its core mission being to solve the most fundamental bottleneck of the spatial computing era: large-scale, verifiable, and ownership-proven high-quality 3D data.

According to the latest compiled data on stablecoin current wealth management and earning products from HTX, Binance, OKX, and Bitget, major platforms still generally adopt a "high returns for small amounts, tiered reduction for excess amounts" mechanism. The table shows OKX’s yields are presented net of a 15% fee. For USDT: HTX’s 0-200 USDT tier offers an annualized yield of 10%, dropping to 1.95% for excess amounts; Bitget’s 0-300 USDT tier yields 7.33%, with excess amounts at 2.67%; Binance’s 0-800 USDT tier yields 6.51%, excess at 2.51%; OKX’s yield stands at 2.54%. For USDC: HTX’s 0-200 USDC tier has an annualized yield of 8%, excess at 2.75%; Binance’s 0-300 USDC tier yields 7.20%, excess at 2.20%; Bitget’s 0-300 USDC tier yields 6.66%, excess at 2.23%; OKX’s yield is 1.95%. For other stablecoins: HTX’s USDT VIP tier offers an annualized yield of 6%-9%, applicable for amounts between 50,000 and 100,000; Binance’s USDT VIP yield is 2.6%-2.65%; Bitget’s USDT VIP 0-300,000 tier yields 2.88%, with excess amounts at 1.88%. For USDE: HTX presents tiered yields of 5% and 3%, Binance offers 4.75%, and Bitget has a yield of 1.05%. HTX’s USDD yield is 4.00%; Binance’s U product’s 0-5,000 tier yields 8.70% annually, excess at 0.70%, while Bitget’s U product yields 1.50%. Overall, high returns remain concentrated in small amount tiers, with annualized yields generally declining as capital size increases. Users should also pay attention to tier limits, interest calculation rules, fees, and real-time availability when comparing products. The above data are page-displayed yields and do not constitute investment advice.

SlowMist tweeted that ether.fi suffered an attack, resulting in losses of approximately 15.45 ETH. According to analysis, the root cause lies in the lack of access control for the `solver` parameter passed by the caller in the `solve()` function of the AtomicQueue contract—there is no check for `solver == msg.sender`, nor any signature, registration, or authorization verification. The attacker first constructed a malicious AtomicRequest via the `updateAtomicRequest()` function, then forced a victim address to act as the solver. The AtomicQueue contract subsequently called `finishSolve` on the victim address, executing `want.transferFrom(solver, users[i], assetsToUser)`, and exploited the ERC-20 approval the victim had previously granted to steal funds.

Insight Beating AI News Flash: Y Combinator (YC)-incubated AI data firm Specific Labs has launched Real-SWE, a coding evaluation benchmark that tests coding agents on real private code from actual enterprises. The tasks are sourced directly from production environments, including tax calculation, bill migration, API billing, and customer data migration. Neither the code nor answers are publicly available, so agents must independently master the company’s business rules and code structure. Results show Fable 5.1 ranked first with a 38.8% pass rate; GPT-6 Astra followed at 33.8%, Gemini 3.8 Flash at 31.2%. GLM 5.3 took fourth place with 28.8%, outperforming Grok 4.6, Kimi K3, and GPT-5.6 Sol. Of the 10 currently public tasks, 6 have an overall pass rate below 15%, and one task saw all models fail. The standout performer was GLM 5.3. Real-SWE assesses agents’ ability to continuously read code, identify rules, and complete multi-step modifications in unfamiliar projects. Zhipu AI researcher Xiaopu Peng noted that his team has been training long-range tasks since GLM-5.1, and Real-SWE is more aligned with such capabilities than public SWE benchmarks, which partly explains GLM 5.3’s strong performance here.