$WATERWATER Preço

On August 15, the latest filing from the U.S. SEC revealed that Leopold Aschenbrenner's Situational Awareness LP submitted a 13F-HR on August 14, 2026, disclosing a snapshot of its public holdings before the liquidation crisis in July. This 13F information indicates that Situational Awareness reported a total declared value of approximately $20.24 billion in holdings. The positions are highly concentrated: SanDisk accounts for about $5.674 billion, approximately 28.0%; Micron accounts for about $5.574 billion, approximately 27.5%. Together, these two storage chain stocks exceed $11.2 billion, making up about 55.5% of the portfolio. In addition to storage chips, the portfolio also heavily bets on AI infrastructure and computing power chains: Bloom Energy approximately $1.899 billion, TSMC ADR approximately $1.265 billion, Nebius approximately $1.233 billion, CoreWeave approximately $0.745 billion, and Core Scientific approximately $0.666 billion. The filing also shows holdings in Applied Digital, IREN, Riot Platforms, CleanSpark, and other data center, power, and Bitcoin mining-related assets. This 'pre-collapse ledger' presents a highly concentrated bet on the 'AI computing power bottleneck': storage, wafer foundry, cloud computing, power, data center, and mining infrastructure are bundled in the same direction. At the end of July, the market reported that Situational Awareness was forced to sell most of its public stock positions due to the decline in AI-related stocks and leverage pressures, ultimately leading Citadel to take over the problematic stock portfolio, raising concerns in the market due to the fund's turmoil. According to this holding, Leopold compressed multiple high-volatility assets in the AI infrastructure chain into a single super crowded trade. When AI trading is favorable, it feels mythical; when the chip, computing power, and power chains retract together, leverage can quickly rewrite the myth into an accident. However, it is important to emphasize that the 13F only discloses U.S. listed securities and some options as of June 30, and does not reveal intraday trading, the full picture of short positions, financing structure, or sales after June 30.

On August 15, the AI community has been buzzing about the possibility of multiple versions of DeepSeek-V4-Pro. Some users discovered that when calling the deepseek-v4-pro API, changing the IP or recreating a session resulted in three different 'inference styles': one frequently starting with 'Let me', resembling the previous V4 Pro Preview; another often stating 'The user wants me', similar to V4 Flash; and a third heavily using 'we', referred to by some users as the more powerful 'God Version V4 Pro'. Since the performance tends to remain stable once a session enters a certain mode, the community speculated whether DeepSeek was hiding multiple models behind the API and distributing them through a routing mechanism. However, further analysis of the DeepSeek Harness source code provided another explanation: the differences may not stem from different model weights but rather from the Agent operating environment. The community noted that on August 10, the official DeepSeek Harness repository updated a key commit: 'fix(preset): align minimal agent with RL composition'. This update aimed to ensure that the Minimal Agent was consistent with the Agent environment used during reinforcement learning (RL) training. The official documentation indicates that the Minimal preset includes a minimal system prompt, a persistent Bash environment, specified editing tools, and a compaction policy used in RL training, while removing additional identity prompts, web prompts, and tool descriptions. This suggests that DSH Minimal may not be a 'stripped-down version' of the Standard version but rather simulates the real Agent environment encountered during model training. Community tests also support this view. The same DeepSeek V4 Pro performed differently across various Harness environments: DSH Standard: 91 points; DSH PTC: 92 points; DSH Minimal: 99/96 points. Subsequently, testers developed the 'Anchored Standard' plugin: the first request simulated the Minimal environment, only opening shell and read tools, and after completing the first tool call, restored the full Standard toolset, resulting in consecutive scores of 98/99 points. Testers believe that the key to the V4 Pro Agent's performance may not depend on how many tools it ultimately has, but rather on what the model first encounters: System Prompt + Tool Schema + Agent Scaffold. Therefore, the so-called 'three DeepSeek models' may actually be a combination of two layers of factors: on one hand, differences in API service environments, deployment configurations, or gray instances; on the other hand, whether the model enters an Agent environment close to the RL training distribution. However, this assertion has not yet been confirmed by DeepSeek officials. The official API documentation indicates that deepseek-v4-pro corresponds to the DeepSeek-V4-Pro-0813 official version and does not disclose a multi-model automatic routing mechanism. Currently, it appears that the different performances of DeepSeek-V4-Pro are more likely the result of the interplay between model weights, inference environments, and Agent frameworks, rather than the simple existence of three hidden models.

On August 15, according to Poloniex market data, PUMP has increased by 19.65% in the past 7 days and by 66.57% in the past 30 days, with its market cap now reaching $1.665 billion. This morning, the Pump.fun team and investors completed the monthly unlocking of tokens, releasing a total of 4.94 billion PUMP tokens, valued at approximately $13.6 million. The unlocked tokens were subsequently distributed to 125 wallet addresses.