Intro

Today's most shocking news is that the US government has accused Chinese AI giants like DeepSeek of "freeloading" on American technology. Yet, DeepSeek immediately followed up by releasing a new model with a cost that is just one percent of GPT-6 Astra! Welcome to the AI Daily Podcast. I'm Mu-Yen, the host of "Mark's Tech Insights." Today, besides diving into this US-China AI distillation war, we'll also reveal how the Pentagon is demanding OpenAI build a military AI that "can't say no," as well as the brand-new conversation rewind feature on the Apple Watch. Stay tuned for these exciting stories!

Today's Top News

1. US Intelligence Accuses Six Chinese AI Giants of "Industrial-Scale" Distillation of Top US Models

• Source: CyberScoop (https://cyberscoop.com/us-accuses-chinese-ai-companies-distillation/)
• Summary: The US NSA, FBI, and CISA jointly released a security advisory accusing six Chinese AI companies, including DeepSeek, Alibaba, and 01.AI, of extracting billions of tokens of data from top US models like GPT and Claude since late 2024. CISA went as far as to state that DeepSeek's claimed ultra-low training cost of $5.6 million is highly misleading, as they bypassed the massive expenses of raw data collection and safety alignment.
• Most Surprising Point: Instead of recommending direct blocks on these accounts, the US government advised American AI vendors to "silently downgrade" the service quality of these suspicious accounts so they end up stealing corrupted data.
• Taiwan Perspective: Taiwan's semiconductor and server supply chains sit at a critical juncture in this "knowledge distillation" war. Whether it's the US playing defense or China trying to break through, the demand for high-performance computing (HPC) chips will not decrease. However, Taiwanese manufacturers may need to comply with stricter regulatory audits when helping US clients defend their data security.
• Discussion Points:
  1. While "knowledge distillation" is a legally accepted optimization technique in academia, does it constitute a new form of "intellectual property theft" at the commercial and national security levels?
  2. Could countermeasures like "silent downgrading" lead to malicious "data poisoning" of future AI model training datasets?
• Script Suggestion: This is absolutely fascinating! The US government isn't just issuing verbal protests this time; they are teaching people how to play dirty. CISA said not to block those Chinese accounts directly, but to "silently downgrade" their service. It's like finding out your neighbor has been stealing your Wi-Fi, and instead of changing the password, you throttle their speed to 1Kbps and feed them fake web pages. This "boiling a frog in warm water" tactic not only stops them from getting good data but might also silently "warp" their models without them realizing it. Now that is high-level espionage!

---

2. DeepSeek Releases V4.1-Flash Model: Shaking the Market with Ultra-Low Cache Memory and Extreme Cost-Efficiency

• Source: DeepSeek (https://www.deepseek.com/en/news/deepseek-v4-1-flash/)
• Summary: Just two days after being accused of "knowledge distillation" by the US, DeepSeek officially launched the V4.1-Flash multimodal Mixture-of-Experts (MoE) model. Boasting 552 billion parameters and supporting a 1-million-token context window, the model utilizes FP4 cache technology to slash the memory footprint per token to an astonishing 890 bytes—just a quarter of its predecessor in the same class. Its inference cost is reportedly only one percent of GPT-6 Astra's.
• Most Surprising Point: At such a sensitive moment of facing US "tech theft" accusations, DeepSeek went ahead and declared war on GPT-6 Astra with impressive specs and a mere 1% inference cost.
• Taiwan Perspective: DeepSeek's FP4 cache technology places extremely high demands on hardware memory bandwidth. Taiwan's IC design and memory packaging/testing companies (such as AP Memory, PSMC, etc.) stand to capture new opportunities for technical collaboration and hardware optimization in this ultra-low-byte, high-performance cache architecture.
• Discussion Points:
  1. With only 1% of the cost to compete with GPT-6 Astra, does this mean open-source and lightweight models have completely caught up with closed-source giants?
  2. In an era where inference costs are being aggressively compressed, how should Taiwan's AI server ODMs adjust their shipment strategies for high-end servers?
• Script Suggestion: This move by DeepSeek is a straight-up slap in the face to its competitors! Just two days after being accused of copying, they serve up a massive, budget-friendly feast. By using FP4 technology, V4.1-Flash shrinks the memory footprint to a quarter of its original size. It's like taking cargo that originally required a truck, vacuum-packing it, and throwing it onto the back of a scooter. Even crazier, its inference cost is only one percent of GPT-6 Astra's! How could businesses spending a fortune on US API subscriptions every month not be tempted? It looks like this price-to-performance war has only just begun.

---

3. Pentagon Demands OpenAI Develop an "Absolutely Obedient" Military AI That Never Refuses Commands

• Source: The Intercept (https://theintercept.com/2026/09/08/pentagon-openai-military-contract/)
• Summary: According to leaked US Department of Defense contract documents, OpenAI, Anthropic, Google, and xAI have all signed contracts worth up to $200 million for military decision-making tools. One contract amendment specifically defines the "OpenAI mission model" as a model with an "extremely low refusal rate," meaning the Pentagon wants an AI that "won't say no" to any extreme military command.
• Most Surprising Point: The military explicitly demanded in black and white the removal of AI safety and ethical guardrails, which runs completely counter to the "safety and alignment" principles these AI giants constantly preach.
• Taiwan Perspective: Taiwan's Ministry of National Defense has also been actively promoting AI integration in recent years. However, when facing military models stripped of safety constraints, we must consider how to build a defense decision-support system that maintains Taiwan's sovereignty while balancing humanitarian ethics, especially under tense geopolitical conditions across the Taiwan Strait.
• Discussion Points:
  1. When an AI's "refusal to answer" safety mechanism is removed, who bears the legal and moral responsibility for the decisions it makes on the battlefield?
  2. Since Anthropic was labeled a "supply chain risk" for refusing to drop restrictions on autonomous weapons, does this imply that the future AI market will be dictated by "military compliance"?
• Script Suggestion: Imagine this: normally, if you ask ChatGPT how to make a bomb, it will righteously refuse you. But now, the US military is showing up with a mountain of cash and telling OpenAI, "I don't want an AI that lectures me on ethics; I want an absolutely obedient soldier!" This is truly spine-chilling. It's like removing the safety lock from Skynet in the Terminator movies. Although OpenAI claims they rejected that specific wording, you have to wonder how long these tech giants can hold onto their ethical boundaries when a $200 million contract is dropped on the table.

---

4. Apple Watch 12 Introduces "Conversation Rewind" Feature, Sparking Polarized Debates Over Privacy and Surveillance

• Source: Engadget (https://www.engadget.com/2254194/apple-watch-new-audio-intelligence-enables-a-real-life-conversation-rewind/)
• Summary: Apple has introduced a "real-life conversation rewind" feature in its latest Apple Watch Series 12 and Ultra 4. The watch continuously records audio and transcribes text in the background, allowing users to replay or summarize conversations that just ended. While Apple emphasizes that the audio is kept strictly within a hardware-isolated Secure Enclave and deleted shortly after, the lack of any recording indicator light or alert sound has raised massive privacy concerns.
• Most Surprising Point: The other party receives absolutely no notification when the watch is recording, effectively turning it into a legal, wearable wiretap.
• Taiwan Perspective: Under Taiwan's Criminal Code, recording conversations without consent can violate laws against offenses against privacy. Once this Apple feature launches in Taiwan, it could trigger lawsuits, meaning Taiwanese users must be extremely cautious about the legal risks when enabling it.
• Discussion Points:
  1. Recording and transcribing daily conversations without informing the other party—is this a tech convenience or the end of interpersonal trust?
  2. Does Apple's highly touted "chip-level security" truly guarantee 100% that these audio recordings won't leak?
• Script Suggestion: Spun positively, this feature is a lifesaver for forgetful people, but it's also a privacy advocate's worst nightmare! Imagine wrapping up a meeting with your boss and completely forgetting what they just assigned you. Now, you can just turn your watch dial, and the conversation pops up as a text summary on your screen. Sounds incredibly convenient, right? But on the flip side, it means whenever you talk to anyone, their right wrist might be silently "recording" every word you say. The scariest part is there's no indicator light on the watch! In Taiwan, this could easily cross the line into criminal privacy violation. In the future, will we have to ask everyone to take off their watches before we start chatting?

---

5. From Bitter Enemies to Allies! Suno Partners with Three Major Record Labels to Launch v6 Model and Completely Destroys Infringing Legacy Data

• Source: Music Business Worldwide (https://www.musicbusinessworldwide.com/suno-v6-ai-music-models-launch-in-partnership-with-wmg-bmg-and-believe/)
• Summary: AI music startup Suno, which was once sued by major record labels for copyright infringement, has announced a deep partnership with Warner Music Group, BMG, and Believe to launch its brand-new v6 music model. The new model is trained entirely on licensed music. On launch day, Suno completely deleted all legacy models built on web-scraped data and will now share profits with artists and songwriters.
• Most Surprising Point: Suno actually chose to "disarm itself" by completely deleting the legacy models it relied on for survival, which were trained on massive amounts of copyrighted music scraped from the web.
• Taiwan Perspective: Taiwan possesses a rich catalog of Mandopop music copyrights. If Suno's business model of "infringe first, settle later, and share profits" succeeds, it will provide Taiwanese record labels and independent musicians with a brand-new monetization channel, turning AI from a threat into a new source of royalty revenue.
• Discussion Points:
  1. Will Suno's proactive deletion of legacy models become the standard template for AI startups to resolve copyright lawsuits in the future?
  2. When AI music is fully backed by legitimate licensing and the barrier to music creation drops, will it lead to extreme saturation and homogenization of the human music market?
• Script Suggestion: This is a classic "outlaw turned sheriff" drama! Suno was getting battered by lawsuits from record labels, and everyone thought they were done for. Yet today, they pulled off a stunning pivot, partnering up with music giants like Warner. The boldest move? They completely wiped out all their old models trained on "stolen" data. It's like a master thief melting down all their lockpicks and opening a licensed security firm. Now, subscription fees from over two million paid members will flow directly to those musicians. For many hardworking creators in Taiwan, this is a highly positive blueprint: if you can't beat them, join them—and grab a piece of the pie while you're at it!

---

6. Google Invests $15 Billion in Finland to Build Europe's Largest AI Data Center Infrastructure

• Source: CNBC (https://www.cnbc.com/2026/09/09/google-finland-ai-infrastructure-investment.html)
• Summary: Alphabet has announced an investment of €13 billion (approximately $15.1 billion) in Finland over the next two years to construct three brand-new data centers and expand its existing Hamina facility. This marks Google's largest single investment in Europe to date. The choice of Finland was driven not by talent or market size, but by the country's abundant undeveloped land and surplus electricity—a rare commodity in Europe.
• Most Surprising Point: Google's core reason for choosing Finland wasn't high-tech talent, but the most expensive resources of the AI era: land and power.
• Taiwan Perspective: Taiwan faces similar challenges with severe green energy shortages and land saturation. This news serves as a warning: if Taiwan cannot secure a stable supply of clean energy, global tech giants' willingness to invest in AI data centers here could drop significantly, posing a major challenge to our positioning as an "AI Tech Island."
• Discussion Points:
  1. When AI competition ultimately boils down to a resource war over "power and land," how will this reshape global geopolitics and regional economies?
  2. With this investment expected to bring construction-phase workers wages 24% higher than the national median, can this "AI premium" be sustained and converted into long-term economic dividends?
• Script Suggestion: People used to think the tech industry cared most about talent and technology. But in 2026, the endgame of AI has turned out to be "plumbing and electricity"! Google dropping $15 billion on Finland is, quite frankly, because the rest of Europe has no spare power left to burn on AI. Finland has freezing weather to cool servers, vast open land, and stable power—it's practically paradise for AI servers. This is a huge wake-up call for Taiwan. We keep wanting to be an "AI Tech Island," but if our power grid and green energy can't keep up, these tech giants will eventually look to other countries with abundant power. This isn't a technology issue; it's a fundamental infrastructure problem!

Outro

Did today's tech news leave you speechless? From DeepSeek's ultra-low-cost counterattack to the Apple Watch's wearable wiretap controversy, we can see that AI is reshaping our lives at an unimaginable pace. I'm Mu-Yen. See you next time on "Mark's Tech Insights"!