🎙️ AI 日報 Podcast — Anthropic 神秘模型、OpenAI 上市角力、華為晶片新突破

Intro

(Intro music)

Mark: Hey everyone! Welcome to "Mark's Tech Insights." I'm your host, Mark. Today has been both exciting and dramatic in the world of AI! We saw Anthropic announce a god-tier model, Mythos, that's so powerful they're afraid to release it publicly. Meanwhile, its competitor OpenAI is internally divided over whether or not to go public. Plus, China's new breakthroughs in AI chips and models are definitely worth our attention. Ready? Let's dive right into the highlights!

Today's Top Stories

1. Anthropic Announces Ultra-Powerful Cybersecurity AI Model Mythos, but Decides "Not to Release" It for Now Due to Its Dangers

Source: TechCrunch (https://techcrunch.com/2026/04/07/anthropic-mythos-ai-model-preview-security/) & CNBC (https://www.cnbc.com/2026/04/07/anthropic-claude-mythos-ai-hackers-cyberattacks.html)
Summary: Anthropic has launched a cybersecurity initiative called "Project Glasswing," exclusively licensing its unreleased, ultra-powerful model, Mythos, to about 50 top organizations, including AWS, Apple, and Google. Mythos reportedly found thousands of zero-day vulnerabilities within weeks. Its capabilities are so potent that Anthropic fears it could be used by hackers to launch devastating cyberattacks. Therefore, they will not release it publicly until new safety mechanisms are in place.
Taiwan Perspective: This is both a wake-up call and an opportunity for Taiwan's cybersecurity industry. We need to start thinking about how to defend against "AI-level" cyberattacks. At the same time, if we can get access to tools like this, it would be a huge boost to Taiwan's overall digital defense resilience.
Discussion Points:
- The "double-edged sword" nature of AI is becoming increasingly apparent. How should developers and society establish regulations?
- Will this "Responsible Non-deployment" model become the new norm for the development of cutting-edge AI?
- Anthropic is committing hundreds of millions of dollars in compute credits and donations. Is this corporate responsibility or a clever PR strategy?
Script Suggestion: "Hey, this news story feels straight out of a movie! Anthropic has basically created a 'digital divine weapon' called Mythos. This weapon is so powerful it can easily find the most hidden vulnerabilities in a system. But they also realize that if it falls into the wrong hands, the consequences would be catastrophic. So they've chosen to 'seal it away' for now, giving access only to the 'chosen ones' like Google and Microsoft who have the power to wield it. This really brings up a core question: what do we do when AI's capabilities exceed our ability to control them? For Taiwan, we have many talented white-hat hackers and cybersecurity teams. This is like a whole new class of offensive and defensive weaponry suddenly appearing. Should we try to get our hands on it, or should we hurry up and figure out how to defend against it? It's something that really deserves our deep consideration."

2. Internal Disagreement at OpenAI: Sam Altman Pushes for IPO, CFO Believes the Time Isn't Right

Source: PYMNTS (https://www.pymnts.com/news/ipo/2026/openai-leaders-at-odds-over-ipo-plans/)
Summary: Reports of internal disagreement at OpenAI over a potential IPO have surfaced. CEO Sam Altman is reportedly pushing for a public offering in Q4 2026, targeting a valuation of up to $1 t r i l l i o n . H o w e v e r, C F O S a r a h F r i a r b e l i e v e s t h e c o m p a n y i s n^{'} t r e a d y, e s p e c i a l l y s i n c e i t^{'} s p r o j e c t e d t o b u r n t h r o u g h a n o t h e r$ 200 billion before turning a profit, making an IPO at this stage too risky.
Taiwan Perspective: This reflects the immense capital pressure faced by top AI companies. Taiwanese startups can learn from this tug-of-war between "technical vision" and "financial reality" at OpenAI when planning their fundraising and growth strategies.
Discussion Points:
- Is it really appropriate for a company still in a massive spending phase to attempt a trillion-dollar IPO?
- Is Sam Altman rushing the IPO to raise more capital for AGI research, or are there other pressures at play?
- With 900 million weekly active users and $25 b i l l i o n i n a n n u a l i z e d r e v e n u e, w h y d o e s a c o m p a n y l i k e t h i s s t i l l n e e d t o b u r n$ 200 billion? Just how staggering are the costs of AI?
Script Suggestion: "This is really interesting. On one hand, you have Anthropic talking about 'responsibility,' and on the other, OpenAI is worrying about 'money.' It feels like Sam Altman has his foot on the gas, wanting to use an IPO to raise massive funds from the public market to continue his AGI dream. But his CFO, Sarah Friar, is the one hitting the brakes, saying, 'Hold on, we haven't even prepared our financial statements, and we still need to burn another two hundred billion dollars!' This is a classic Silicon Valley showdown: 'growth hacking vs. practical operations.' A trillion-dollar valuation is an insane number, but considering they raised $122 billion in their latest funding round, it doesn't seem entirely impossible. This also shows us that the AI race, in the end, is really a battle of capital."

3. Major Breakthrough in Chinese AI Hardware! DeepSeek V4 Model Confirmed to Run on Huawei Chips

Source: NxCode (https://www.nxcode.io/resources/news/deepseek-v4-release-specs-benchmarks-2026)
Summary: The DeepSeek V4 model, released by Chinese AI company DeepSeek, has been confirmed to be trained and run on Huawei's Ascend 950PR chips. It's a 1-trillion-parameter Mixture of Experts (MoE) model, but it only activates 370 billion parameters per inference, making it highly efficient. This marks the first time China has successfully built a frontier AI model on its own domestic semiconductor infrastructure.
Taiwan Perspective: This is an extremely important signal for Taiwan's semiconductor supply chain. We used to think that high-end AI training chips were dominated by Nvidia, but now Huawei's Ascend chips have proven capable of supporting a trillion-parameter model. The technological decoupling driven by geopolitics is accelerating.
Discussion Points:
- Can the performance of the Ascend 950PR really compete with Nvidia's H100 or B200? Or is it compensating for hardware gaps through software and model architecture (MoE)?
- The training cost was only $5.2 million. How credible is this figure? If true, it would significantly lower the barrier to entry for training AI models.
- DeepSeek V4 is released under the Apache 2.0 open-source license. What impact will this have on the global open-source community and the AI ecosystem?
Script Suggestion: "This news, in my opinion, is the most significant story of the day! In the past, everyone was wondering if China could develop its own high-end AI chips under US sanctions. Now, DeepSeek V4 has provided the answer: yes. And it's not a small model; it's a genuine 1-trillion-parameter MoE large model, running directly on Huawei's Ascend chips. What does this mean? It means that China's AI development might be taking its first real step towards 'de-Nvidiasization.' For Taiwan, this is something we absolutely cannot take lightly. In the past, TSMC was the arsenal for all AI chips. Now, Huawei is starting to build its own arsenal. This war in AI hardware has expanded from a simple technological competition to a full-blown geopolitical struggle."

4. Microsoft Does It Themselves! Launches Three In-House AI Foundational Models

Source: TechCrunch (https://techcrunch.com/2026/04/02/microsoft-takes-on-ai-rivals-with-three-new-foundational-models/)
Summary: Microsoft has announced three AI models developed by its internal "MAI Superintelligence" team: MAI-Transcribe-1 (speech-to-text), MAI-Voice-1 (text-to-speech), and MAI-Image-2 (image generation). The goal is to reduce its reliance on OpenAI. MAI-Voice-1 is claimed to be able to generate 60 seconds of audio in just 1 second, an astonishing level of efficiency.
Taiwan Perspective: Tech giants are all "de-risking"—not putting all their eggs in one basket. Taiwanese companies should consider similar strategies when adopting AI to avoid being locked in by a single vendor.
Discussion Points:
- Does this move by Microsoft signal that the "honeymoon period" with OpenAI is over, shifting towards a more direct competitive relationship?
- After Mustafa Suleyman (co-founder of DeepMind) joined Microsoft, he delivered results so quickly. How important is his team to Microsoft's AI strategy?
- Will these in-house models be integrated into Azure and Office 365, or will they be offered as standalone products to compete with OpenAI's API?
Script Suggestion: "The relationship between Microsoft and OpenAI has always been like a complex dance for two. Microsoft invested a ton of money in OpenAI and deeply integrated their models into its own products. But now, Microsoft is going back to building its own kitchen and cooking its own meals. This is actually very logical, because you can't bet your entire future on a single 'partner,' especially when that partner is growing stronger and could potentially become your biggest competitor. The team led by Mustafa Suleyman is moving really fast. These three models directly target the speech and image generation spaces, clearly aiming to shore up Microsoft's own foundational capabilities. This is good news for developers—more choices, and potentially more competitive pricing."

5. A Rare Alliance! OpenAI, Anthropic, and Google Team Up to Combat Model Copying from China

Source: Bloomberg (https://www.bloomberg.com/news/articles/2026-04-06/openai-anthropic-google-unite-to-combat-model-copying-in-china)
Summary: Usually competitors, OpenAI, Anthropic, and Google are now collaborating through the "Frontier Model Forum." They are sharing intelligence on how Chinese rivals are using "distillation" techniques to copy or replicate the outputs of their top AI models. This is a rare instance of direct cooperation among AI labs on a national security-related issue.
Taiwan Perspective: The protection of intellectual property faces new challenges in the age of AI. Taiwan's own AI research and development could face similar risks of plagiarism. It's worth considering how to establish effective protection mechanisms.
Discussion Points:
- What is the technical principle behind this "distillation" attack? How difficult is it to defend against?
- Does the willingness of these fierce rivals to cooperate mean they believe the threat from China now outweighs their business competition with each other?
- Will this collaboration be limited to intelligence sharing, or will it evolve into joint technical countermeasures?
Script Suggestion: "As the saying goes, 'the enemy of my enemy is my friend.' This news story perfectly illustrates that. OpenAI, Google, and Anthropic are usually at each other's throats in the market—poaching talent, fighting for customers, and competing for compute. But now, they're sitting at the same table to jointly deal with 'model thieves' from China. This 'distillation' technique, simply put, involves using a smaller model to imitate the behavior of a larger one. By constantly querying the large model and learning from its answers, it essentially 'steals' the larger model's intelligence. For companies that have spent billions of dollars on training, this is, of course, unacceptable. This collaboration shows that the AI race has escalated to the 'national team' level. These American companies realize that if they don't unite, they risk being 'overtaken on the curve'."

Outro

Mark: Alright, that was a ton of information today! From Anthropic's reverence for the power of AI, to OpenAI's ambition in the capital markets, and the fierce battle between the US and China in AI hardware and models. We can clearly see that the development of AI is no longer just a purely technical issue. It involves business strategy, national security, and massive capital operations. This grand drama is only going to get more exciting. Thanks for listening, and we'll see you next time!

(Outro music)