Changxin Memory Technologies' Q2 EBIT margin reached 82%, surpassing Samsung and SK Hynix.

Beating AI News Flash: Moonshot AI has fully rolled out Kimi K2.8 Preview to Kimi Code CLI. Users do not need to switch Model IDs, as the existing "kimi-for-coding" will automatically point to the new model. Kimi stated that K2.8 Preview’s overall performance is nearly on par with K3, while offering higher thinking efficiency. It also supports the same thinking intensity adjustment as K3 and expands the context window to 1 million tokens, compared to the previous K2.7 Code’s 256k context window which only allowed thinking mode activation. This update essentially brings two core capabilities from K3 to the default code model.

Dongcha Beating AI News Flash: OpenAI has launched Data Agent on ChatGPT Work. Employees can simply ask questions like "Why are sales dropping?" or "Which group of customers is most likely to churn?" and the tool will independently retrieve the company's data, analyze the reasons, and generate data dashboards that update in line with real-time data. It can connect to databases including Snowflake, BigQuery, Databricks, Redshift, MongoDB, and also access Google Drive and SharePoint. The tool adheres to the company's existing definitions and access permissions for metrics such as "active users" and "revenue", preventing inconsistent calculations of the same metric by different users. After analysis, it can send results to Slack or email, and carry out follow-up operations once approved by the user. OpenAI stated that nearly all of its product teams, as well as more than two-thirds of its sales, marketing, and other teams, are already using such data agents.

According to Slow Mist monitoring, Web3 IP copyright platform BeatSwap was hacked, resulting in a loss of 2,984,557 BTX, worth approximately $77,512. The attacker borrowed 6,000,000 BTX via a flash loan and dumped it into the Uniswap V3 pool to lower the sqrtPriceX96. Exploiting a vulnerability in the LiquidityVestingConvert contract — which only uses the V3 spot price and lacks Time-Weighted Average Price (TWAP) and deviation limits — the attacker called the deposit() function twice (for 10,000 and 2,000 USDT) to manipulate prices, mint positions, and drain all BTX from the liquidity pool.

According to on-chain analyst Ai Yi (handle @ai_9684xtpa), crypto KOL XXAntiWar recently completed three trades. The KOL spent a total of around $1.787 million to build a position in token "Niu Lai" at an average price below $0.086; if he has not sold the position, his estimated profit exceeds $750,000. He shorted 5,199.97 ZEC (valued at ~$6.62 million) at an average of $1,273, then closed the short position at an average of $1,145.4, earning a profit of $660,000. He also went long on 4.765 million PONS (worth ~$3.15 million) at an average price of $0.5984, with an unrealized profit of $308,000.

Beating AI Express News: OpenAI has released the GPT-Live-1 API, the real-time voice model currently powering ChatGPT. Capable of listening and speaking simultaneously, it directly understands speech to handle everyday conversations, and can also adapt to interruptions, pauses, and background noise. For complex reasoning or tool calls, GPT-Live-1 can offload tasks to GPT-6 Astra, Luna, or third-party models, then resume conversations with users after retrieving results. Developers can choose which models to integrate on the backend. Compared with the traditional "speech-to-text → large language model → text-to-speech" workflow, GPT-Live-1 cuts out intermediate model handoffs, making real-time conversations more coherent. GPT-Live-1 is priced at $0.05 per minute, with separate charges for backend models.

Beating AI Express (Insight): AI Agent startup NeoCognition has released ApprenticeBench, a benchmark specifically designed to test whether AI can learn a job independently like a new hire. The first version places an Agent at a simulated California construction company to handle accounts payable: it reads six months of historical bills, company manuals, and Odoo tutorials, then processes 100 bills over seven consecutive months, with rule changes and supervisor feedback occurring throughout. Results show Fable 5.1 achieved a 72% success rate, GPT-6 Astra hit 68%, while the best human test subject scored just 51%. The task includes invoice entry, error detection, cost code allocation, vendor outreach, approval routing, and learning internal company rules from historical records and feedback. Notably, the two models performed slightly better when directly operating the GUI than when calling APIs, largely eliminating the performance loss common in Computer Use scenarios. However, the "surpassing humans" finding is limited to this single simulated role, with only two human test samples. Cost is also higher: Fable 5.1 averages $18.23 per task, compared to ~$7.21 for humans. Humans grow faster over time, while Agents slow down as they accumulate more stored information.