L2BEAT has been renamed ETHBEAT, expanding its coverage to include Ethereum-related areas such as privacy, interoperability, and staking.

A protocol upgrade proposal has emerged on the Scroll DAO forum, aiming to reposition Scroll as a permissioned application-specific network focused on privacy AI, while migrating the SCR token to the Ethereum mainnet. The proposal states that if approved, the Scroll Foundation will lead the network upgrade and asset migration. The new network will support an integrated AI product stack including Compass, Compass API, CENO, USX and SCR, with a focus on privacy AI interactions, security credentials, agent workflows, payments, and computing power coordination. SCR’s total supply and overall tokenomics will remain unchanged, and the token will continue to fulfill governance functions. The governance scope is proposed to extend to core parameters of USX, and SCR may also act as collateral for certain high-risk commitments and validators.

Insight from Beating AI: U.S. AI startup Levanto Labs has launched Sage2, a model that combines information extraction with real-time decision-making. It can automatically extract details like amounts, dates, and user needs from chats, and continuously adjust its judgments based on new content. For example, Sage2 can monitor user sentiment for web-based AI customer service tools. If a user initially asks for a refund but grows increasingly frustrated, Sage2 will update its assessment accordingly. When the user’s dissatisfaction hits a pre-set threshold, the system automatically transfers the interaction to a human agent, no need for the user to request it explicitly. According to the company, Sage2 takes around 150 milliseconds to extract information from short text, with a judgment update as fast as 50 milliseconds. The system retains conversation context and reuses computational caches, eliminating the need to resubmit full chat logs each time. It is primarily targeted at use cases like customer service, business automation, and AI agents, making it suitable for handling large volumes of repetitive judgment tasks. Compared to calling general-purpose large language models, this specialized decision model is expected to reduce latency and costs. In the company’s in-house evaluation, Sage2 scored 77.8 in information extraction accuracy, slightly outperforming Cloudflare’s data extraction model Clairic, which scored 75.3.

Beating AI News Flash: Indian AI firm Sarvam has launched open-source decision-making model Prism, fine-tuned from Sarvam 105B. Prism adopts a Mixture of Experts (MoE) architecture, with a total of around 105 billion parameters, though only approximately 10 billion are activated per inference run. Thanks to inference optimization, the median response time for a single request has dropped from 664 milliseconds to 44 milliseconds, a speedup of roughly 15 times. Sarvam tested 25 decision models; Prism scored an average of 71.8 points across three benchmarks, making it the highest-scoring model with a response time under 100 milliseconds. The firm also ran tests on 2,000 multiple-choice questions, and Prism’s answers remained consistent regardless of option shuffling. The model’s weights are now released under the Apache 2.0 license, allowing enterprises to fine-tune and deploy it independently.

Hong Kong Special Administrative Region (HKSAR) government welcomes the Legislative Council’s passage today of the 2026 Inland Revenue (Amendment) Bill, which will implement the OECD’s Crypto-Asset Reporting Framework (CARF) and the amended Common Reporting Standard (CRS) in Hong Kong. The HKSAR government noted that CARF will take effect in Hong Kong next year, with relevant service providers required to collect specified information per the framework; the Inland Revenue Department (IRD) will automatically exchange tax information on crypto-asset transactions with partner jurisdictions starting from 2028. The amended CRS will also come into force in 2028, and the IRD will automatically exchange financial account information under the new standards from 2029. Hong Kong’s Secretary for Financial Services and the Treasury Christopher Hui Ching-yu said that implementing the latest international automatic information exchange standards will help consolidate Hong Kong’s reputation as an international financial and commercial center, and boost the confidence of crypto-asset service providers, financial institutions and digital asset users in operating and making long-term investments in the city. Hong Kong will conduct information exchange based on reciprocity, data confidentiality and security. The IRD will issue guidelines and set up a dedicated team to provide technical support and answer inquiries for the industry.

After U.S. Democrats regained control of Congress, congressional investigative powers are set to expand significantly. Robert Garcia, a Democrat on the House Oversight Committee, has publicly urged relevant parties to "get a lawyer"; Ron Wyden of the Senate Finance Committee has directly signaled the likelihood of issuing subpoenas. Multiple law firms have begun compiling potential investigation checklists for their clients, covering topics including allegations related to Trump’s business and family wealth accumulation, the Iran war, immigration enforcement, Dogecoin’s layoff impacts, and AI and the cost of living.

Beating AI News Flash: OpenAI has unveiled an unreleased internal model. In the BenchCAD 2.0 engineering design evaluation, the model posted an average score of 0.875, significantly outperforming GPT-6 Astra’s top inference tier score of 0.594, a relative improvement of around 47%. The model’s name and release date remain unannounced. BenchCAD 2.0 primarily tests AI’s capabilities to reconstruct 3D parts from blueprints, assemble mechanical structures, and identify circuit connections. Astra, which previously ranked first in public evaluations, has also made low-level errors: it misidentified 3.3V, 5V, and ground lines on a circuit board as the same network. Such errors could directly trigger hardware failures if applied to actual circuit design. OpenAI also showcased Astra’s robotics experiment: two employees designed a physical device that can interact with Codex via voice in less than a week, with Astra completing the casing, internal structure, circuit board, and software design. However, actual production still requires multiple rounds of manual review and adjustment; the contract manufacturer took nearly a month to produce the first batch of prototypes. OpenAI is also collaborating with racing team Chip Ganassi Racing to convert 3D scans of parts into editable CAD models and manufacturing drawings, and with Mazda to explore aerodynamic car design by adjusting front-end shapes and simulating airflow to reduce drag.