Binance will airdrop 3500 EPT tokens to eligible users

Beating AI Express reports that independent developer Wang Zan has open-sourced an Android AI assistant named Mika, which can directly control apps like WeChat and Xiaohongshu on mobile devices to help users organize chat logs and send messages. The developer demonstrated Mika’s capabilities: it can browse WeChat group chats in the background, summarize chat records, and even send messages to WeChat’s File Transfer Assistant. Throughout these operations, users can continue using other apps on their home screen without ceding screen control to the AI. Unlike the Doubao mobile assistant, which relies on system-level permissions provided by phone manufacturers, Mika uses Android wireless debugging to launch an auxiliary process with Shell permissions, creating an invisible virtual screen on standard mobile phones to run apps like WeChat. The AI reads buttons, text, and input fields via Accessibility Services to perform taps, swipes, and text input; for interfaces where controls are unrecognizable, it uses local OCR or vision models to operate based on visual content. To use Mika for the first time, users need to enable Developer Options, complete wireless debugging pairing, and grant Accessibility Service permissions. The model supports custom API configuration and login via ChatGPT accounts. It does not require a computer connection during runtime, and tools like Codex on PCs can assign tasks to it via MCP. The cross-app operation feature works on Android 12 and above. On Android 14, Mika can leverage the virtual screen’s independent focus capability, allowing it to operate in the background without seizing the main screen’s input focus.

According to on-chain analyst Ai Yi (@ai_9684xtpa), Binance’s SAFU wallet transferred 7,500 BTC worth approximately $580 million half an hour ago. The transfer is likely related to replenishing reserves for affected assets. Following the transfer, the address still holds 7,500 BTC, roughly half of its balance before the transaction. Subsequently, Binance Co-CEO He Yi announced that BTC, ETH, USDT, BNB, SOL, USDC, XRP, USD1, RLUSD, and ZEC have all resumed 100% full reserve backing.

Bitwise’s Q4 staking report shows that Ethereum’s staked ETH has reached 43.8 million tokens, accounting for 35.9% of its total supply, a new all-time high. Net staking inflows for Q3 stood at 1.99 million ETH, marking six consecutive quarters of net inflows. Solana’s active staked SOL has risen to 440.5 million tokens. However, staking yields for both networks have declined. Ethereum’s average annualized staking yield in Q3 fell to 2.63% from 2.81% in Q2, with 96.3% of that yield coming from ETH issuance. Solana’s gross staking yield dropped to around 6% from 6.38%, with nearly 90% still derived from token issuance. The above figures do not deduct service provider fees and are not equivalent to actual returns for clients. Coinbase noted that staking ETFs and digital asset treasuries are expanding participation channels. Kean Gilbert, head of institutional relations at the Lido Ecosystem Foundation, pointed out that with custodian support for staking, institutions can participate via their existing custodial relationships, and current discussions have shifted to implementation issues such as validator decentralization, code audits, and fund staking ratios. Liquidity remains a key constraint for institutions. Gilbert explained that funds typically need to meet redemptions within one to two days, while ETH un-staking can take weeks, leading institutions to hold some un-staked assets. He estimated that the average staking ratio for related products in Europe is only 50% to 60%. Liquid staking tokens can offer alternative exit channels, but actual sales still depend on market liquidity and prices.

According to Axios, a U.S. official has stated that if Republicans suffer a landslide defeat in the midterm elections, Donald Trump would be more inclined to order large-scale military operations. The official noted: “If Trump feels embarrassed or sidelined, he will take bold actions. That’s what everyone is telling the Iranians. Now is the best time for them to reach a deal.” However, a White House official clarified that the election results will not impact Trump’s decisions on Iran-related issues.

Beating AI News: U.S. company TBC, founded by two neurosurgeons, has released the video model Mischief 1. The new model is modified from Alibaba’s open-source Wan 2.2, with less than 0.1% additional parameters. TBC claims that with comparable image quality, its video generation speed is at least 5 times faster, and inference costs are reduced by around 80%. This technology stems from the team’s four-year research on living neurons. Researchers cultured brain cells on a chip equipped with 4096 electrodes, converted visual information into electrical stimuli, and recorded how neural activity spreads spatially and decays over time. Based on these patterns, they designed lightweight software modules, which were embedded into the video model’s diffusion Transformer structure, helping the AI maintain consistency between consecutive frames without training the entire model from scratch. Video models can be accelerated by reducing denoising steps, but too many steps lead to blurry, distorted images and difficulty maintaining frame consistency. TBC previously ran experiments using an AI model that generates real-time Minecraft visuals based on player inputs. After adding the bio-inspired modules, the model retained stable visuals while cutting denoising steps. Combined with caching and sampling optimizations, the overall system’s throughput increased by 4.4 times. Today, TBC is commercializing this research line. Mischief 1 has been integrated into fal and OpenRouter, and currently supports generating 3-second or 5-second videos from images. A 5-second 480p video costs just 2.5 cents to produce, while a 720p video costs 5 cents.

Beating AI Express News: On the evening of October 31, the DeepSeek Harness team published a technical paper titled "Exploring Super-agent Design: Towards More Autonomous Agent Systems". DeepSeek noted that advancements in model capabilities have spurred new application forms, with personal agents recently emerging in large numbers. However, mainstream agent designs today struggle to support this model at scale for multi-tasking and team collaboration. Agents still operate with a single session as their basic unit, and different sessions function independently. Under this design, the upper limit for single tasks depends on model capabilities, while the upper limit for multi-tasks depends on human attention. Through concepts including a two-layer architecture, event-driven design, asynchronous multi-channel interaction, separation of storage and computation, and an external watchdog, the DeepSeek team has overcome these limitations. The system handles tedious session management and context synchronization at the underlying layer, allowing humans to shift focus from specific execution details to higher-level decision-making and acceptance. DeepSeek refers to this agent form and its system architecture as a "Super-agent". The paper notes that internal Super-agent instances built using this framework are already running continuously, regularly participating in issue reproduction and fixes, code review responses, and the iteration of DeepSeek Harness itself. (DeepSeek Harness Team)