PrismML compresses 9-billion-parameter model weights to 2GB: iPhones generate 34 tokens per second locally.

Tesla’s second-generation Roadster launch event showcased extreme stunts including inverted driving, airborne flips, and a 0-100 km/h acceleration of 0.81 seconds, with every attendee receiving a pair of earplugs. The inverted driving relies on a rear fan that pulls air from under the car to generate downforce equivalent to roughly twice the vehicle’s weight; the airborne stunts use the same cold gas thrusters as SpaceX, featuring 10 nozzles and about 250 kg of compressed gas that occupies the rear seats. A rough calculation of cold gas propulsion efficiency shows that hovering the 2-ton vehicle requires approximately 28 kg of gas per second, meaning the full 250 kg supply is only enough for 9 seconds of hovering, with actual usable time being just a few seconds. Media outlets described the feat not as flying, but as “coolly crashing”. While the performance was spectacular, the specifications are drastically misaligned with promises made nine years ago: its range has shrunk from roughly 1,000 km to around 480 km, delivery has been delayed from 2020 to 2028, and pricing remains unannounced. The thruster-equipped version will likely fail to meet traffic regulations, and U.S. aviation regulators have never certified such devices. The event also did not address the risks posed by the compressed gas tank (holding hundreds of atmospheres of pressure) behind the rear seats in collisions or rollovers. Musk’s consistent stance on the matter is: “If safety is your top priority, don’t buy the Roadster.” Separately, per BIT (bit.com) market data, Tesla’s U.S. stock is up 4.3% in pre-market trading, while SpaceX is up 1.9%. The Tesla Roadster is equipped with cold gas thrusters developed in collaboration with SpaceX.

According to Reuters, Gerald Podobnik, co-head of corporate banking at Deutsche Bank, stated that the lender will launch its first digital asset service in the coming weeks, kicking off with crypto custody in Germany before expanding the custody platform to Asian financial hubs next year. The service will be rolled out in phases: the first phase will cover custody of major cryptocurrencies such as Bitcoin and Ethereum, followed by adding acceptance services for mainstream stablecoins. The bank also plans to offer tokenized deposits and assist clients in issuing blockchain-based securities like digital bonds.

Beating AI Express News: NetEase UU Remote Control has released a third-party app integration protocol on GitHub, enabling developers to add PC remote control entry points to their iOS and Android applications. For example, an AI programming assistant can integrate a "Connect Work PC" button: when users click it for the first time, they are redirected to UU Remote to select and bind their Windows or Mac device. Subsequent clicks on the button will directly initiate remote control, eliminating the need to repeatedly search for devices. The remote control supports full-screen and picture-in-picture modes, allowing users to operate their PC via UU Remote’s floating window while using other mobile apps. For instance, users can check if Claude Code or Codex on a home Mac has completed tasks remotely through a third-party app when away. Developers can also obtain connection results from the picture-in-picture mode to determine if a device is offline or if the connection failed. It should be noted that the protocol being opened is an inter-app call protocol, not the source code of the remote control engine or interfaces for AI to automatically operate PCs. Third-party apps still need to call the installed UU Remote client, with actual remote control functionality provided by NetEase. Currently, controlled devices only support Windows and macOS, while the mobile client requires UU Remote version 4.42.0 or newer.

Ethereum founder Vitalik Buterin announced in a post that he has packaged zkAPI, Tor, additional obfuscated traffic, and local large language model prompt rewriting into an OpenAI interface-compatible daemon, and added a "paranoia mode" to the personal proxy framework pi. Each conversation is assigned an independent Tor circuit, requests are masked with random obfuscated traffic, zkAPI prevents identity leaks during payments, and prompts are rewritten by the local model. Vitalik noted that the proxy is also wrapped with the bwrap sandbox and torsocks, with all operations routed through Tor. It is now relatively usable but has significantly higher latency, especially when initiating new conversations. He added that Tor is not optimal for this use case, explaining a dedicated mix network would be more ideal, and cheaper gas fees for zkAPI deposits and withdrawals are set to launch on the mainnet soon.

Beating AI Express News: Tencent Hunyuan has upgraded Hy Memory to version 2.0, open-sourcing its memory framework, dedicated model, and MemHub app all at once. The 1.0 version launched in May this year focused on chat memory, capable of tracking changes to the same matter, identifying and resolving contradictions, and inferring unstated user habits from multiple conversations. When users adjust their preferences, the system retains old records while remembering the latest decisions. Version 2.0 adds working memory and code memory: it tracks task progress, adopted approaches, and final outcomes, and checks if past experiences apply to similar tasks. Code memory goes further, integrating with Git logs to track modification histories of files and functions—even if code is rolled back, it can retrieve prior changes and identify previously tested solutions. The supporting MemHub connects memory across different AI tools, allowing 10 tools including Cursor, Codex, and Claude Code to share a single local memory. With user authorization, MemHub imports historical conversations and continuously syncs new records. For example, if project specifications are set in Cursor and development continues in Codex, the specifications can be directly retrieved without re-explaining. Tencent also trained a dedicated model, Hy-Memory-30B-A3B, with 30 billion total parameters, activating only about 3 billion at a time. It is responsible for filtering, organizing, and updating memory, and can integrate with existing frameworks like Mem0 and MemOS. Tencent tested four memory systems, with an average score of 55.49, slightly higher than Opus 5’s 54.62.

According to TradingBeats monitoring, HYPE is trading at around $80.91. The $77–$78.12 range below the current price currently covers 10 buy orders totaling approximately 190,600 HYPE tokens, valued at around $14.8 million, which is 3.5% to 4.8% lower than the current price. Of this total, about $10.8 million is newly added today, accounting for roughly 73%. The buy orders are mainly concentrated in four tiers: - $77.44–$77.51: ~$8 million. Two buy orders of ~$4 million each, with related accounts currently holding no HYPE, belonging to long entry plans after a pullback; one order at $77.51 is newly added today (0x2c6f…). - $78.12: ~$4.689 million. Two addresses with no current HYPE holdings placed buy orders of ~$3 million and ~$1.689 million respectively, both added early today (0xafe4…, 0x6cb9…). - $77.149: ~$386,000. Corresponding to additional position-adding plans for existing longs (0x90db…). - $77.08–$77.717: ~$1.726 million. Two existing short positions arranged for covering; based on current positions, after execution, they will mainly be used to reduce short positions (0x6749…, 0xb0f0…). Based on current positions, approximately 88% of the buy orders correspond to new long entries or long position additions, while around 12% are for short covering.