OpenAI officially announced it will merge ChatGPT and Codex usage quotas in December: regular chat users will now have to compete for programming quota, except for the Instant plan.

Beating AI Flash News: Elon Musk disclosed that Memphis-based supercomputing cluster Colossus 2 plans to deploy around 1.2 million NVIDIA Blackwell GPUs by the end of the year. Two additional facilities will launch soon: Gun Range is set to roll out 220,000 GB300 GPUs by the end of November for model training; Spam, converted from a warehouse originally used to store goods like luncheon meat, will deploy around 200,000 GB300 GPUs exclusively for AI inference. Currently, Macrohard is equipped with 110,000 GB200 GPUs and 220,000 GB300 GPUs, while Macroharder and Minihard each have 220,000 GB300 GPUs. Adding the planned 420,000 new GPUs, the five clusters will total around 1.19 million GPUs. Training and inference workloads have distinct network requirements: Macroharder, Minihard, and Gun Range all use 800Gbps network cards, while Spam’s inference tasks do not need the same level of high-speed interconnection. Musk noted that its network configuration is simpler, requiring less than half the fiber optic cable of training facilities.

Beating AI News: Meta partners with Shanghai-based Fourier Intelligence to develop humanoid robot prototype Max in just three months. Built on Fourier’s hardware and technical accumulations from its first three generations of humanoid robots, Max saw Meta primarily adjust its body proportions, head design, and cameras, plus add components like both hands. Meta’s Vice President of Robotics Products spent nearly two months at Fourier’s Shanghai factory, completing an internal demonstration at the end of August. When Meta established its robotics team last year, it aimed to create an "Android of the robot world": developing its own AI software to enable robots from various manufacturers to use it. CTO Bosworth noted at the time that running and jumping robots are not the biggest challenge; the real bottleneck lies in fine manipulation—for example, picking up a cup of water without crushing it or spilling it, which requires AI to understand the physical world and master precise control over force and movement. The partnership with Fourier lets Meta avoid building the robot body from scratch, speeding up such experiments. Max stands about 1.5 meters tall, is equipped with NVIDIA Jetson Thor chips, and its five-fingered hands can grasp small bottles and toys. The two sides launched Max 2 in September, focusing on improving bilateral hand operation and stair-climbing capabilities. Meta’s long-term goal is to integrate its Muse AI agent with a physical form to enter homes and perform practical tasks.

Beating AI News (Dongcha) reports: xAI has updated its developer documentation, announcing that starting November 7, it will add invisible text watermarks by default to API outputs for the Grok 4.7, 4.6, 4.5, 4.3 models and three Grok 4.20 variants. The measure covers not only EU but also global and regional APIs, with even structured outputs and tool call parameters marked. It remains unclear if the watermark will also be enabled for Grok’s web version and X’s built-in chat feature. The watermark works by slightly adjusting word selection probabilities during text generation, leaving identifiable statistical patterns in the generated text. Direct copy-pasting will not remove the watermark, while shortening text or extensive rewriting will reduce detection reliability. The move stems from transparency requirements under the EU AI Act, whose relevant provisions took effect on August 2. While xAI has not signed the EU’s voluntary transparency guidelines, it is still required to comply with applicable legal obligations. Developer teams can disable watermarks per model, but must confirm eligibility for relevant legal exemptions and bear full responsibility for marking and disclosure. Google implemented SynthID text watermarks for Gemini as early as 2024, and Anthropic began rolling out watermarks globally for its new Claude models in August this year. OpenAI announced in October that it would gradually add watermarks to ChatGPT and Codex for EU regions, while allowing global API users to enable the feature voluntarily, with it remaining disabled by default.

BTCFi protocol Troves announced that its xWBTC/WBTC vault was hacked early this morning, with an estimated initial loss of around 8.6 BTC. Investigations show the attack was limited to this specific vault, with no other Troves vaults found affected. The hack exploited an outdated rebalancing contract that had been replaced but still retained permissions. The team has revoked the relevant contract’s permissions and suspended affected and potentially high-risk vaults for security reasons. It is currently working with security partners to trace the stolen funds and evaluate recovery plans. Additionally, Troves sent an on-chain message to the attacker offering a white hat bounty: a reward of 1 WBTC upon return of the remaining funds. The team will release a detailed post-mortem report and plans for affected depositors later.

Beating AI News Brief: OpenAI, Google, Microsoft, and Anthropic have reached an agreement: going forward, websites only need to add two rules to their `robots.txt` file — `User-Agent: __usage_ai-training` and `Disallow: /` — to notify all four companies simultaneously not to use their content for AI training. Previously, websites had to set rules separately for crawlers like GPTBot and ClaudeBot. This rule set is exclusively for AI training, does not block crawlers from accessing websites entirely, and does not necessarily prevent AI from citing website content in its responses. Proposed by an Anthropic engineer, the framework is currently a technical draft from the Internet Engineering Task Force (IETF). The companies will not implement the rules uniformly: Microsoft allows websites to block AI training specifically while keeping their content eligible for citation in AI search results; Google plans to tie this rule to its existing Google-Extended setting, which may also restrict Gemini from citing website content. Anthropic has already begun adopting the rule, Microsoft plans to roll out its implementation in February 2027, Google aims to support it in 2027, and OpenAI will launch its version in the coming months. However, these rules cannot reverse data already used for prior training, nor can they force AI companies to comply.

Exclusive from Dongcha Beating AI Express: Y Combinator-backed startup Subvocal announced it has secured $1.3 million in total funding and is developing a pair of "silent typing" glasses. Users can input text into a computer by making subtle speaking mouth movements, no vocalization or keyboard contact required—ideal for writing emails, taking notes, or issuing commands to AI agents. The team claims its prototype can transcribe over 200 words per minute continuously. The glasses capture tiny movements of facial articulatory muscles via low-power radio frequency signals, which are then decoded by an AI model. Users still move their mouths, but do not need to produce sound. Similar technology is not novel: MIT launched AlterEgo in 2018, which used facial electrodes to recognize silent speech, though it was tested with a limited vocabulary of around 20 words at the time. Subvocal aims to advance this to contactless, unlimited-vocabulary continuous input. That said, the product is still a long way from launch. The current demo prototype relies on an external circuit board, differing significantly from the planned glasses form factor. Additionally, the 200+ words per minute speed has not been verified by independent public testing. Subvocal has opened pre-orders in the U.S., with an early bird total price of $499 and a refundable $199 deposit required to reserve a spot. The company expects to start shipping in Q1 2027, pending FCC approval.