Reading

Applied Digital announced on October 2 that it has brought an additional 75 megawatts of critical IT load online at its Polaris Forge 1 AI data center campus in Ellendale, North Dakota, raising fully operational capacity to 250 MW.

This delivery completes Building 2’s full 150 MW phase, consisting of three new 25 MW data halls. Prior to this, the campus had already delivered 100 MW from Building 1 and an initial 75 MW from Building 2.

CEO Wes Cummins stated that the milestone demonstrates the company’s ability to transform large-scale power into usable AI infrastructure. The campus is contracted to deliver 400 MW at full buildout for CoreWeave, meaning more than 60% of the contracted capacity is now ready for customer equipment.

It is important to note that “Ready for Service” status indicates infrastructure availability but does not disclose server installation, utilization rates, or recognized rental revenue. The remaining 150 MW of construction remains a key metric to watch.

The Hong Kong SAR Government has officially introduced a four-step "AI talent ladder" strategy to mitigate the impact of artificial intelligence on entry-level employment.

Kevin Choi Kit-ming, Permanent Secretary for Innovation, Technology and Industry, stated at the Global Youth Powerhouse Summit that the "entry-level shock" is real, as AI performs well in many routine tasks that once constituted graduates' first jobs, such as basic coding and first-line customer service.

The government views AI as a "catalyst, not a threat," aiming to enhance overall digital literacy at a "population scale" rather than shielding young people from the technology. The strategy is designed to guide emerging professionals from foundational literacy to industry leadership.

Several major South Korean banks have reported customer data leaks from peripheral business systems, and the Financial Services Commission has launched an IT inspection across the entire financial industry.

Shinhan Bank's loan-agent inquiry system was attacked for three consecutive days from September 28, with data on 25,729 customers possibly exposed. KB Kookmin Bank said its staff mobile support system lost data on 119 customers, Hana Bank's sales support system involved 89, and BNK Busan Bank found on October 1 that personal data of 11 outsourced developers had leaked. None of the systems were core banking systems, but all had externally facing entry points.

FSC Secretary-General Shin Jin-chang called on banks to strengthen vulnerability checks on externally exposed systems, tighten authentication and access control, and share threat intelligence for a joint response. Some in Korea's cybersecurity community suspect attackers may have used AI agents to automate the attacks; the attacker's identity and methods remain unconfirmed.

On October 2, Meta open-sourced the ESP32 firmware and Linux development kits for Muse Gadgets under the Apache 2.0 license, enabling developers to connect low-cost boards or Raspberry Pis to the Muse AI agent.

Previously confined to software chatbot interactions, this update allows Muse to control displays, sensors, and actuators, and even access local HTTP API devices via home network tunnels. Official documentation explicitly warns: community pairing lacks manufacturer verification, tokens are embedded in firmware, and tinkering may result in bricking or voided warranties.

Meta also launched the finished device Muse Home Link, a USB-C powered small gateway for connecting smart home devices like TVs and speakers. According to product lead Nat Friedman, the first batch of approximately 5,000 units is being distributed for free to US Muse subscribers, one per person, on a first-come-first-served basis.

The system prompt for OpenAI's active code model, GPT-6 Sol Codex, has reportedly been leaked, with approximately 294,000 characters now public on GitHub.

According to IT Home, user @elder_plinius claims to have extracted the complete instruction set. The leaked content includes a strict ban list for "AI slop" words (such as delve, leverage) and mandates the use of ripgrep over grep for text searching.

The document also details long-context management strategies, such as using notes tools to save progress when token budgets are exhausted, and requires status updates every 60 seconds during tool execution. OpenAI has not officially commented on the leak, so its authenticity remains unverified.

Getting machines to learn continuous actions by trial and error—how much to press the gas, how much to rotate a joint—is called reinforcement learning. The machine has to maintain an estimate throughout: how many points it can get by following the current way of acting. The 2018 paper TD3 found that this estimate tends to be biased upward, and the machine chases the inflated score; the fix is to maintain two independent scorers, trust only the lower one, and slow down changes in the way of acting.

Today's demos of robots learning to walk and grasp things are often still backed by this trial-and-error learning, so the problem of inflated scores is still there. TD3's "take the lower of two scorers" was not bypassed by later new methods; instead, it became a standard feature of mainstream algorithms. Any system that chooses actions by estimating scores should ask one question: has the overestimation been prevented.

If the actions are discrete—choose left or right, rather than how many degrees—then don't use it to judge; its validation is entirely on simulated robot tasks, so don't directly extrapolate the results to real robots or other tasks. And suppressing the scores has a cost: genuinely good opportunities are also suppressed along with it, and overestimation is merely replaced by underestimation.

《Addressing Function Approximation Error in Actor-Critic Methods》(2018)｜Next review 2027-09-20