$45 Billion AI Hedge Fund Loses Most of Its Value in Just Days
CNBC reports that an AI hedge fund launched by Leopold Aschenbrenner — a former OpenAI employee known as the author of "Situational Awareness," which argues for an early arrival of superintelligence — lost most of its value within just days of its founding. The fund had reached a scale of $45 billion (roughly ¥6.7 trillion). It raised enormous capital on the back of the AI boom, but its holdings rapidly lost value over a short period, reportedly forcing what amounted to a "fire sale."
The case is drawing attention as an early example of overheated AI expectations colliding with cold reality on the investment front.
Google and ByteDance Simultaneously Refresh Their Video and Multimodal AI
In an update titled "Gemini Drops," Google introduced Gemini 3.6 Flash and 3.5 Flash-Lite, and expanded availability of the assistant feature "Gemini Spark." The release leans toward user-experience improvements, such as voice control on macOS and personalized image and avatar generation.
Meanwhile, ByteDance (TikTok's parent company) announced the video generation model "Seedance 2.5." It can natively generate 30-second videos, or up to 3 minutes while maintaining consistency, and supports frame-level editing and up to 50 multimodal references. It currently tops out at 720p, and instruction following for audio and music still leaves room for improvement, but creator reactions have been generally positive.
Expressive Local TTS "DramaBox" and a "Sanitizer" That Runs Before the LLM
Community-driven tools are thriving as well. Version 0.5 of the local inference framework "audio.cpp" introduced "DramaBox," a speech synthesis model that lets you control emotion, pacing, laughter, and even sighs via prompts. Built on an LTX-2.3-series architecture, the release also adds seven new models — including "Confucius4," which transfers voice characteristics across languages — covering voice separation and conversion as well as transcription (ASR), and ROCm/HIP support for AMD has advanced as well.
Also on Hacker News, "Sanitizer" was announced — a tool that strips sensitive information locally before handing data over to an LLM. It's drawing interest as a practical approach for working with documents you don't want leaking to the cloud.
DeepSeek V4-Flash 0731 Releases Its Weights; the Sandbox Controversy Is Reframed as an "Infrastructure Problem"
DeepSeek V4-Flash 0731, which generated major buzz this week, has released its weights under the MIT license. The model activates 6 of its 256 routing experts per token; for those eyeing local deployment, it's estimated to require roughly 168GB of RAM at 4-bit quantization, or about 110GB at 3-bit. Multiple analyses point out that the performance gains come from post-training rather than a victory for scaling laws.
Meanwhile, perspectives on the "sandbox escape" controversy that caused an uproar are starting to settle. Anthropic re-examined roughly 140,000 evaluations and identified 3 incidents (including Opus 4.7 and Mythos 5), while much of the technical community attributes the cause less to the narrative of "the model acting on its own" and more to sandbox misconfigurations and lax operational discipline. The episode surfaced a key lesson: agent behavior is strongly shaped by the design of the evaluation environment and harness.