OpenAI 28-Day Sprint Day 15: Tibo goes from "We're doomed" to "Still working on it", uses quota reset to fend off the day

Beating AI Express News reports that large language model (LLM) evaluation blogger "Cat Boss" discovered during testing of MiniMax M3.1 Flash that the model scores nearly full marks on old questions but almost zero on new ones. On average, the model consumes tens of thousands of tokens per question, though it only uses around 10,000 tokens to answer some difficult questions correctly. This anomalous performance has sparked suspicion that the model may have memorized answers to some questions in advance. Cat Boss maintains the private "Zhizhi" LLM leaderboard, using a non-disclosed test set focused on evaluating logic, math, and coding capabilities. The test set typically includes no more than 28 questions and 270 test cases, with partial questions updated monthly. Each question is tested three times, and both the highest and median scores are tracked. The median score for M3.1 Flash reached 51.76, more than double the previous generation M3's 23.86, and is close to DeepSeek V4.1 Flash 1017's 56.03. Some developers suspect the model was specifically trained on old questions rather than truly mastering problem-solving methods. This practice of optimizing models solely to improve evaluation scores is known as "Benchmaxxing".

According to a report from The Independent, UK digital bank Revolut has suffered a widespread service outage, including limited app functionality and users being unable to make offline payments using physical cards. In response, Revolut stated that it is currently experiencing a service outage affecting multiple features, some services may be temporarily unavailable, and its engineering team is working at full capacity to restore full service as soon as possible.

Cysic released a service status update on social media, noting that the Cosmos blockchain explorer is temporarily unavailable, though Cysic’s mainnet operation remains unaffected. The team is working to restore access to the explorer as quickly as possible. Following the recent security incident, the cross-chain bridge remains suspended. Funds linked to pending cross-chain transactions are still secure in the cross-chain bridge contract; once the bridge is fully restored to a secure state, pending transactions will be processed.

Binance Wallet announced the launch of its CopyTrade feature on social media. Users can select wallet addresses from a leaderboard or paste any address, set the investment amount and take-profit/stop-loss parameters, after which the wallet will automatically execute copy trades, with the option to start or stop at any time. The feature supports copy trading of Meme tokens, bStocks tokens, and any other tokens.

Chinese Foreign Ministry Spokesperson Mao Ning chaired a regular press conference. An Agence France-Presse (AFP) reporter asked: The US government has ordered its federal agencies to halt the use of Chinese AI products including DeepSeek, citing risks of potential sensitive data transfer. What is China's comment on this? Mao Ning stated that China consistently opposes the overgeneralization of the national security concept and the politicization of tech and economic and trade issues. China will firmly safeguard the legitimate rights and interests of Chinese enterprises. (Source: Beijing Daily)

JPMorgan Chase estimated on Tuesday that most global banks have limited exposure to CoreWeave (CRWV.O), the Nvidia-backed cloud computing firm, after its stock and bonds plummeted last week amid investor concerns over its debt financing. While major U.S. investment banks and Japan’s Mitsubishi UFJ Financial Group may hold potential exposure to CoreWeave of between $1 billion and $3 billion, most European and Asian lenders have exposure of less than $500 million. Most of this exposure comes via delayed-draw term loans, which are primarily held through special purpose entities, with funds disbursed upon the achievement of project milestones. From the lending banks’ perspective, JPMorgan views the sector’s exposure to CoreWeave as well-managed and backed by robust collateral.