🎙️ 450 億 AI 基金差點玩完？Nvidia 靠這沿用舊模型的神秘框架戰力飆升 100%｜AI 日報 Podcast

A Wall Street boy wonder used a crazy 400% leverage to bet on AI chip stocks, almost wiping out a $45 billion star fund, and is now officially in the SEC's crosshairs! Today, besides taking you through this financial storm, Mu-Yen will also talk about the wild plan to stuff an H100 into a fridge and launch it into space, and why the secret behind your AI getting smarter isn't actually the model itself. We've got all this and more coming up—welcome to Mark's Tech Insights!

Today's Highlights

1. Star AI Hedge Fund Situational Awareness, Which Narrowly Avoided Collapse, Under SEC Investigation

Source: TechCrunch (https://techcrunch.com/2026/08/24/situational-awareness-star-ai-hedge-fund-that-nearly-imploded-now-being-probed-by-the-sec/）
Summary: The US Securities and Exchange Commission (SEC) is issuing subpoenas to several Wall Street banks, deeply investigating the trading timing, leverage operations, and communications with creditors of the star AI hedge fund Situational Awareness. Founded by Leopold Aschenbrenner, a former OpenAI researcher in his twenties, the fund scaled its assets under management (AUM) to $45 b i l l i o n w i t h i n t w o y e a r s . H o w e v e r, i t m e t i t s W a t e r l o o a f t e r u s i n g a n e x t r e m e 400 45 billion within two years. However, it met its Waterloo after using an extreme 400% leverage to bet heavily on AI chip stocks. Its assets plummeted to$ 10 billion at one point, eventually forcing it to transfer a portion of its holdings to Citadel at a discount.
Most Surprising Aspect: That a former OpenAI researcher in his twenties could leverage $45 billion in just two years, yet almost collapsed this entire empire overnight due to 4x leverage.
Taiwan Perspective: Many Taiwanese investors are also highly enthusiastic about chasing AI concept stocks. However, the fact that even one of Wall Street's top researchers who understands AI technology best can crash due to over-leveraging in the face of market volatility is undoubtedly a wake-up call for the Taiwan market: technical understanding is not a safe haven for market operations.
Discussion Points:
1. Do founders with technical backgrounds lack a healthy respect for systemic risk when managing massive financial assets?
2. To what extent does 400% leverage reflect the market's frenzy and potential bubble crisis surrounding AI concept stocks?
Script Suggestion: Mu-Yen wants to ask everyone: if you were in your twenties and managing $45 billion, how would you play it? Leopold, the genius former OpenAI researcher, chose the most thrilling path—going straight for 4x leverage to bet big on AI chip stocks! The result? He almost lost his shirt, and now he's under SEC investigation. This is like an engineer who understands race car engines thinking they are a Formula 1 champion driver, only to fly off the track at the very first turn. It shows us that understanding technology and understanding financial markets are two completely different things. Don't ever assume that just because you understand AI trends, you can use a cheat code in the stock market!

2. Nvidia Research Shows: The Outer Harness Is the Real Hero Making AI Smarter

Source: TechCrunch (https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/）
Summary: Nvidia published research on an agent harness called AVO (Agentic Variation Operators). The experiment found that, without retraining the model at all, the exact same vanilla Claude Opus 5 model scored only 30% on the ARC-AGI-3 benchmark. However, after applying the AVO framework to manage memory, planning, and error recovery, it actually passed with a perfect score of 100%, clearing all 25 environments and 183 levels in just 6,624 actions.
Most Surprising Aspect: Without modifying or retraining the model at all, simply relying on the outer harness's control over memory and planning could turn the exact same "brain" from a failing grade to a perfect score.
Taiwan Perspective: This is fantastic news for the many AI application developers in Taiwan who lack massive computing resources and cannot train large models themselves. It means that as long as we focus on designing clever "add-on management architectures," we can also build god-tier applications.
Discussion Points:
1. Does the success of the AVO framework mean that the future focus of AI competition will shift from "model parameter size" to "control engineering"?
2. How will this technology, which significantly boosts performance without retraining, change the cost structure for enterprises adopting AI?
Script Suggestion: Have you ever heard the saying "you can't support mud on a wall"? But Nvidia proved this time that as long as the scaffolding is built well, even mud can be used to build a skyscraper! Using the exact same Claude Opus 5 model, the vanilla test score was only 30%, but after wrapping it in their new AVO framework, it scored a perfect 100%! This is like pairing an average student with a god-tier "national coach" who teaches them how to manage time, double-check exam papers, and review. The result? Straight to the top of the class. It seems we don't need to constantly chase larger models; knowing how to equip AI with an "exoskeleton" is the real way to win!

3. Space Data Center Startup Starcloud Raises $250M; Terrestrial Power and Launch Windows Become Critical

Source: TechCrunch (https://techcrunch.com/2026/08/21/starcloud-raises-250-million-for-orbital-data-centers-as-launch-options-dry-up/）
Summary: Starcloud, a startup dedicated to moving AI data centers into space orbit, completed a $250 m i l l i o n e x t e n s i o n r o u n d a t a$ 2.3 billion valuation, with Nvidia also investing $25 million. Last year, the company successfully launched an H100 housed in a dorm-fridge-sized enclosure into orbit and successfully ran inference. This October, they plan to send the latest Blackwell chips into space. Besides tight terrestrial power grids, a major driver for the fundraising is to secure increasingly precious rocket launch windows ahead of time.
Most Surprising Aspect: The data center sent into orbit last year was actually just an H100 box the size of a dorm fridge, and now they are about to send Blackwell up there too.
Taiwan Perspective: Taiwan has recently seen intense discussions regarding the power challenges brought by data centers and semiconductors. While space data centers sound like science fiction, for Taiwanese manufacturers who hold a key position in the low-Earth orbit (LEO) satellite supply chain, this could be a brand-new blue ocean market worth watching.
Discussion Points:
1. How much of a challenge will extreme temperature fluctuations and radiation in space pose to the lifespan of high-heat, high-precision chips like the H100 or Blackwell?
2. As terrestrial power and space launch windows both become scarce resources, how will the future battle for computing power evolve?
Script Suggestion: We used to joke about "throwing data into the cloud," but now Starcloud is literally throwing data into "outer space"! They packed an H100 chip into a box the size of a mini dorm fridge, sent it into orbit, and actually ran AI workloads in space! Now they're raising another $250 million and preparing to send the latest Blackwell chips up in October. Why the rush? Because terrestrial power is running short, and even rocket launch windows are getting booked solid. This is literally "sky-high computing power." In the future, our AI assistants might actually be watching us work from the heavens!

4. DeepMind Alum Startup Inherent Launches Faraday Assistant, Outperforming GPT-5.5 and Claude 4.8

Source: TechCrunch (https://techcrunch.com/2026/08/22/inherent-founded-by-deepmind-alumni-says-its-ai-teammate-just-outperformed-anthropic-and-openai-at-replicating-research/）
Summary: Inherent, a UK-based AI lab founded by former DeepMind members, has launched an AI work companion named "Faraday." In tests requiring the independent reproduction of experimental results from published scientific papers, Faraday outperformed Anthropic's Claude 4.8 and OpenAI's GPT-5.5. Surprisingly, Faraday does not run on a massive frontier model under the hood, but rather on Qwen 3.6, which has only 27 billion parameters. The company has secured $50 million in seed funding.
Most Surprising Aspect: That a small model with only 27 billion parameters could absolutely crush industry giants with hundreds of billions or even trillions of parameters on highly specialized tasks.
Taiwan Perspective: This once again proves the immense potential of "small models paired with sophisticated agent workflows." For Taiwanese enterprises or academic institutions with limited budgets, this path is clearly much more viable and commercially valuable than blindly following the trend of training massive models.
Discussion Points:
1. Why do small models perform better than general-purpose giants on highly complex tasks like reproducing research papers?
2. Will this breakthrough accelerate the trend of specialized "AI teammates" replacing general-purpose chatbots within enterprises?
Script Suggestion: This is absolutely the most inspiring David-and-Goliath story of the year! UK startup Inherent launched an AI teammate called Faraday, which completely stomped tech giants GPT-5.5 and Claude 4.8 when it came to reproducing scientific research papers. The craziest part is that Faraday is powered by Qwen 3.6, a small model with only 27 billion parameters! It's like a petite Wing Chun master using precise moves and leverage to knock out a 200-kilogram heavyweight boxing champion. It seems that in the world of AI, having a massive brain isn't everything—knowing how to think and work step-by-step is where the real magic happens!

5. Twitch Sued in Class Action for Defaulting to Using Streamer Content to Train Amazon AI

Source: NBC News (https://www.nbcnews.com/tech/tech-news/twitch-creators-push-back-amazon-using-content-train-ai-rcna592391）
Summary: Twitch updated its platform policy to allow Amazon to use streamers' live broadcasts and videos to train generative AI by default; streamers who object must manually turn it off in their settings. When questioned by viewers, Twitch Chief Product Officer Mike Minton bluntly stated that if they used an "opt-in" mechanism, nobody would agree, pointing out that giants like OpenAI and Anthropic have already been scraping this public content privately anyway. These remarks triggered a massive backlash from streamers and VTubers, and Twitch and Amazon are now facing a class-action lawsuit.
Most Surprising Aspect: The Chief Product Officer's brutally honest remarks, which directly exposed the high-handed logic of tech giants when acquiring training data.
Taiwan Perspective: Taiwan has a very large and active streaming and VTuber community. This incident has made many Taiwanese creators realize that the video assets they work hard to stream every day are being used as free fuel for AI without authorization. The fairness of creator copyrights and platform terms needs to be re-examined.
Discussion Points:
1. Does the platform's opt-out default cross the moral and legal red lines of creator rights?
2. As the "data drought" forces tech giants to commandeer user content, how can creators protect themselves?
Script Suggestion: Twitch is really treating streamers as free AI fuel this time! They set it by default so Amazon can use everyone's live streams to train AI, and if you want to opt out, you have to dig through your settings to turn it off. The most infuriating part is that management had the nerve to say that if they let people choose, nobody would agree. That is pure robber logic! It's like a restaurant owner secretly taking customers' leftovers to refine into cooking oil, and when caught, saying, "Well, if I asked you, you wouldn't have given it to me anyway." No wonder streamers and VTubers are collectively furious and taking them straight to court—this is just impossible to swallow!

6. TSMC Stock Rises Ahead of Nvidia's Crucial Earnings Report; 3nm Capacity Expected to Ramp Up Early

Source: Yahoo Finance (https://finance.yahoo.com/technology/ai/articles/tsmc-rises-ahead-nvidias-critical-193930535.html）
Summary: On the eve of Nvidia's crucial earnings report on August 26, TSMC's stock showed strong performance. The market expects Nvidia's quarterly revenue to reach $91 billion, a year-on-year increase of nearly 95%. To meet the massive demand for AI accelerators, TSMC is reportedly pulling ahead its timeline by two to three months, ramping up its monthly 3nm capacity to 180,000 wafers by the beginning of the fourth quarter. This will effectively ease chip shortages, though CoWoS advanced packaging remains the tightest bottleneck in the global supply chain.
Most Surprising Aspect: That TSMC can pull ahead the capacity ramp-up schedule for a cutting-edge node like 3nm by a full two to three months.
Taiwan Perspective: TSMC, Taiwan's "sacred mountain protecting the nation," has once again demonstrated unmatched manufacturing prowess and execution. This not only solidifies Taiwan's absolute dominance in the AI hardware industry but also means the Taiwanese supply chain will continue to play the most critical role in the upcoming Blackwell generation.
Discussion Points:
1. Will hitting the 3nm capacity target early help Nvidia escape production hell for its next-generation chips?
2. How long is the CoWoS advanced packaging capacity bottleneck expected to persist before seeing a substantial resolution?
Script Suggestion: "Papa TSMC" is using cheat codes again! On the eve of Nvidia's crucial earnings release, TSMC's stock has been surging. The market expects Nvidia to challenge $91 billion in revenue this quarter—an absolutely massive appetite—and to feed it, TSMC is actually pulling ahead its 3nm monthly capacity by two to three months, aiming straight for 180,000 wafers by the start of Q4! This execution is nothing short of a miracle. However, even if the chips can be made, everything still gets bottlenecked at the narrow gate of CoWoS advanced packaging. It's like having the dumpling wrappers and filling all ready, but the fingers wrapping them aren't fast enough. It seems the lifeblood of global AI is truly held in Taiwan's hands right now!

Closing Thoughts

From the H100 fridge in space to TSMC's capacity explosion on the ground, today's tech advancements are truly moving too fast to keep up with. But no matter how technology changes, protecting creator rights and maintaining a rational investment mindset remain paramount. I'm Mu-Yen, and we'll see you next time on Mark's Tech Insights!