AIЦена Sleepless AI

This study from the University of Science and Technology of China presents a real-world stress test for AI in scientific research. Researchers developed a machine-readable, modular laboratory for catalysis, consisting of 45 automated workstations. They evaluated 48 configurations combining 6 agent frameworks and 9 large language models (LLMs) across 32 expert-defined research tasks, conducting 4,608 tests. The key findings reveal a significant gap between AI's planning ability and reliable physical execution. Only 3.3% of all generated workflows (151 out of 4,608) could be executed without human correction. The best-performing agent/LLM combination achieved a 28.1% success rate. In a five-round closed-loop experiment, the AI could adjust local parameters based on results but failed at higher-level scientific replanning, such as redesigning analytical methods or correcting persistent critical omissions. Maintaining logical integrity for long-sequence workflows (over 30 steps) also proved challenging. The research distinguishes three often-conflated AI capabilities: generating plausible experimental plans, creating physically executable workflows, and performing evidence-driven strategic replanning. Results show the first does not guarantee the second, and parameter tuning is not equivalent to scientific reasoning. The platform serves as both a "testing ground" to quantify AI's scientific aptitude and a potential "training ground." By forming a "plan-execute-feedback-replan" loop with real-world feedback, it can expose AI shortcomings and generate data for iterative improvement, moving toward true end-to-end autonomous discovery.

Chinese home furnishing listed companies are increasingly turning to AI and semiconductor cross-border ventures to revive their fortunes amid declining core businesses. A prominent example is PVC flooring giant **Elegant Home Furnishing**, whose stock price skyrocketed with 10 consecutive trading limits after announcing a plan to acquire a majority stake in **Ou Kang Nuo**, a semiconductor storage testing equipment and services company. Despite Elegant Home's own financial struggles—including recent losses and tight cash flow—the acquisition, funded through asset sales and loans, has dramatically boosted its market value. The deal involves a cross-shareholding structure with Ou Kang Nuo's controlling shareholder. This trend is widespread. Over the past year, numerous home furnishing firms with stagnant main operations have seen their stock prices surge after announcing moves into hot sectors like AI, semiconductors, or computing power. Key cases include: * **Markor Home Furnishings** ("high-end home furnishing first share"), after years of heavy losses, acquired an AI server high-speed copper cable company and later introduced AI computing power investors. * **Zhenai Meijia** ("blanket king") became the first A-listed manufacturer controlled by an AI large model company following a takeover. * **Fasilon** (integrated ceiling) saw its stock soar after establishing an AI subsidiary. * **Jinlong Decoration** experienced multiple trading limits after being linked to "commercial aerospace" and "AI computing," despite clarifying these segments contribute less than 1% of its business. Statistics show over 20 A-share companies announced semiconductor crossovers in 2025, with home furnishing and building materials firms being the most active. This shift is largely driven by pressures from the real estate downturn, shrinking overseas demand, and trade tariffs, pushing companies to seek growth through trendy concepts rather than core business improvement. However, the article warns that such speculative frenzies often end when the hype fades, leaving stock prices to eventually reflect the companies' weak fundamentals, as seen in Jinlong Decoration's subsequent sharp price correction. The repeated attempts to "ride the trend" highlight a desperate struggle for survival in a challenging traditional industry.

A counterintuitive trend is emerging in China's embodied AI sector: numerous startups that are still actively raising capital themselves are now making strategic investments in other companies. An analysis of 29 such enterprises reveals a pattern of "fundraising while investing," where companies, often still in early funding rounds, rapidly deploy capital into the ecosystem. These 29 investment entities, primarily based in the Pearl River Delta and Yangtze River Delta regions, have executed over 125 investment deals. Notably, 17 of them (59%) are humanoid robot manufacturers, making them the most active investors. The trend shows that newer companies are investing faster; firms founded after 2023 made their first external investment in an average of just 22.8 months, with one company, Poke Robotics, doing so within 3.8 months of its founding. Zhiyuan Robot (智元机器人) stands out as the most active corporate venture capital (CVC) player, completing 37 investments in 23 months. A more complex, networked investment structure is also forming, where companies like Lingchu Intelligent (灵初智能) and Lingxin Qiaoshou (灵心巧手), which received investments from larger players, have themselves become active investors, extending the strategic reach of capital down the supply chain. This collective shift from a few large players to widespread participation signals a strategic move to use capital as a lever to accelerate industry consolidation. Companies are seeking to quickly integrate critical supply chain elements—such as dexterous hands, joint motors, and AI data platforms—into their ecosystems to gain a competitive edge in a rapidly evolving, highly contested market window. The strategy aims to compress the decade-long process of traditional supply chain integration into just a few years, though it carries risks related to technology bets and ongoing capital burn. As the sector's IPO pipeline grows, this "fundraise and invest" model may soon become a standard industry practice.

Michael Saylor, founder of Strategy with the largest crypto reserve, revealed using AI to generate $15 billion for Bitcoin purchases. On The Diary Of A CEO podcast, he explained AI helped Strategy develop a novel financial hybrid—preferred shares ($STRK)—between stocks and bonds. The proceeds from issuing these shares were used to buy Bitcoin. Currently, Strategy cannot raise funds via these preferred shares as they trade below their $100 face value. However, Saylor remains optimistic about AI's necessity for success, advising to use AI for unique, unprecedented tasks rather than trying to outsmart it. He also stated Bitcoin is the best long-term capital asset, requiring no specialized knowledge to invest in, and contrasted his view with Elon Musk's predictions of a future of abundance, asserting money and wealth will retain value. As of July 31, Strategy holds 842,138 Bitcoins, acquired for $63.51 billion since 2020, now valued at approximately $54.5 billion. The company reported a Q2 loss of $8.2 billion and began selling Bitcoin from its reserve in late May to maintain financial stability.

In July, amidst significant AI stock volatility, hedge fund performance varied widely. Situational Awareness, an AI-focused fund, sold a majority of its public equity holdings to Citadel, reporting a 67% portfolio value decline for the month. In contrast, Citadel's equity-focused fund rose 14.2%. The narrative of uniform 'quant' success was challenged; while some systematic funds, like Renaissance's RIEF (+9.2%), posted gains, others like Qube's Torus declined. Business Insider's sample showed mostly negative returns for multi-strategy funds, yet BarclaysHedge's broader multi-strategy index estimated a 0.60% gain, highlighting sample and methodology differences. Tech-focused hedge funds fell 3.99%, while convertible arbitrose strategies rose 1.46%. Monthly gains don't always reflect year-to-date performance; RIEF's strong July brought its YTD return to only 4.5%, while Graham Capital's Tactical Trend strategy held a 23.7% YTD gain. The transfer of assets from Situational Awareness to Citadel was a pressured deleveraging event, not a direct arbitrage, as reported metrics cover different fund scopes and valuations. The data underscores that within the 'quant' label, performance diverges based on specific strategies, models, and market exposures.