🎙️ AI 日報 Podcast — Meta 模型大戰、AI 代理人、AI 貧富差距

Opening Remarks

Hello everyone, and welcome to "Mark's Tech Insights"! I'm your host, Mark. Wow, it's been a turbulent week in the AI world. The two giants, Meta and Google, staged a major model showdown, with one side unveiling its superweapon and the other countering with an open-source strategy. At the same time, the concept of AI Agents has moved from science fiction to reality, with Adobe already integrating them into our productivity tools. Today, we'll talk about this exciting clash of the titans and the huge opportunities and challenges hidden beneath the surface.

Today's Top Stories

1. Meta Launches Muse Spark, the First Major AI Model from Meta's Superintelligence Lab

Source: TechCrunch (https://techcrunch.com/2026/04/08/meta-debuts-the-muse-spark-model-in-a-ground-up-overhaul-of-its-ai/)
Summary: Meta has finally played its trump card! They've released a native multimodal model called Muse Spark that can not only chat but also see, think, use tools, and even perform multi-agent collaboration. This secret weapon will be fully integrated into Facebook, Instagram, WhatsApp, and the Ray-Ban smart glasses. It's clear Meta is playing for keeps this time.
Taiwan Perspective: This means the Meta ecosystem will become more powerful and more closed. For marketing and e-commerce businesses in Taiwan that rely on social platforms, this is both an opportunity and a challenge. In the future, they may need to learn how to collaborate with AI to run ads or create content.
Discussion Points:
- Meta is sinking hundreds of billions of dollars in capital expenditures. Can this "money power" tactic succeed?
- For the average user, what are the concrete application scenarios for Muse Spark's "multi-agent collaboration"?
- Does this signal that Meta is officially declaring war on the all-in-one assistant-style AI from OpenAI and Google?
Suggested Script: "Wow, Meta is dead set on blazing a trail with money this time! Their AI capital expenditure for 2026 alone is a staggering $115 billion—that number is insane. This Muse Spark sounds like Meta's version of an 'all-powerful butler.' It's not just a chatbot; it's a super-brain that spans all of Meta's apps. Imagine you post a food picture on Instagram and can just tell the AI: 'Help me edit this photo to look more delicious, write a catchy caption, and then post it to my fan page and WhatsApp broadcast list.' The AI handles the entire workflow in one go. This will have a huge impact on social media managers or brands in Taiwan. The future of work might not be about operating tools, but about 'directing' an AI to complete tasks."

2. Google Releases Gemma 4 Open-Source Model, Capable of Running on Mobile Phones

Source: LLM Stats / Google (https://llm-stats.com/ai-news)
Summary: While Meta was flexing its muscles, Google chose a different path—openness. They launched Gemma 4, a completely open-source model with no commercial restrictions, and there's even a version that can run on mobile phones! This means developers can put powerful AI inference capabilities directly into your phone's apps, no longer needing to rely on the cloud.
Taiwan Perspective: This is fantastic news for Taiwan's IC design companies (like MediaTek) and the vast community of app developers. On-device AI applications are set to explode. Hardware manufacturers have a new selling point, and software developers can create more responsive AI features that better protect privacy.
Discussion Points:
- Open source vs. closed source: In this strategic showdown between Google and Meta, who has the upper hand?
- When powerful AI models can run on phones, which app categories will be the first to be disrupted?
- Is this a repeat of the Android vs. iOS playbook, where an open ecosystem takes on a closed one?
Suggested Script: "This move by Google is truly brilliant! It's basically the 'Android strategy' for the AI world. While Meta is busy building its golden walled garden, Google is just giving away its arsenal for free. The key point about Gemma 4 isn't just that it's open source, but that it can 'run on a phone.' What does that mean? In the future, your phone apps might not need an internet connection to do real-time translation, summarize documents, or even help you write emails. For developers in Taiwan, this is a huge opportunity; they're no longer held hostage by OpenAI's API fees. For phone chip makers like MediaTek, it's a huge boon. Whoever's chip can run Gemma 4 more smoothly will have more market influence. This war between open and closed is getting more and more interesting."

3. AI Revenue Race Heats Up: Anthropic's Annual Revenue Soars to $30 Billion, Surpassing OpenAI

Source: The Next Platform (https://www.nextplatform.com/ai/2026/04/07/broadcom-and-google-benefit-mightily-from-anthropics-meteoric-growth/5214732)
Summary: The AI unicorns are making money faster than a money printer! Anthropic, long considered a strong competitor to OpenAI, saw its Annual Recurring Revenue (ARR) skyrocket from $14 b i l l i o n t o$ 30 billion in just two months, surpassing OpenAI's $25 billion. The race to commercialize AI is getting fiercer.
Taiwan Perspective: This astonishing revenue growth represents a frantic investment in infrastructure. Taiwan's semiconductor and server supply chain, including companies like TSMC, Quanta, and Wistron, are indispensable players in this arms race and among the most direct beneficiaries.
Discussion Points:
- What did Anthropic do right to achieve such staggering growth?
- With OpenAI's revenue growth slowing and rumors of an IPO, does this mean they're feeling the pressure?
- When enterprise customers choose between OpenAI and Anthropic, what are their main considerations?
Suggested Script: "You might still be marveling at OpenAI's success, but in the blink of an eye, the challenger Anthropic has quietly overtaken them! An ARR of $30 billion—that growth curve is just insane. This proves one thing: the enterprise demand for AI is real and massive. Anthropic's focus on safe, reliable, enterprise-grade services has clearly hit a pain point for many large clients. This also gives us a lesson: the AI race isn't just about model performance; business models and market strategy are equally important. Interestingly, the report mentions that Anthropic's expansion benefits Google and Broadcom, which proves once again that in this gold rush, those selling the shovels (chips, cloud services) are making the surest bet. Taiwan's tech industry happens to play this critical role of arms dealer."

4. Adobe Launches AI Creative Agent in Firefly to Execute Complex Workflows Across Apps

Source: Adobe Newsroom (https://news.adobe.com/news/2026/04/adobe-new-creative-agent)
Summary: Adobe has turned the concept of AI Agents into reality! They've added an AI assistant to their Firefly creative suite that you can command with natural language to automatically complete complex, multi-step workflows across multiple applications like Photoshop, Premiere, and Illustrator.
Taiwan Perspective: For the hundreds of thousands of designers, video creators, and marketers in Taiwan, this is a revolutionary tool. It will completely change creative workflows, freeing people from tedious operations to focus more on core creativity. This also means that the value of a designer in the future will lie more in "creative direction" than "software operation."
Discussion Points:
- Will AI agents replace the jobs of junior designers?
- Is this the ultimate form of all future software? Replacing "clicks" with "conversation."
- How does Adobe ensure the accuracy and creativity of the AI when performing cross-app tasks?
Suggested Script: "This news gets me super excited! AI Agents are finally no longer just a concept on a PowerPoint slide. With this update, Adobe is essentially giving every designer a 24/7 super-intern who is proficient in all Adobe software. You can imagine this scenario: you say to your computer, 'Hey Adobe, take the product photos I shot last time, remove the background, apply them to the A4 poster template with the headline "Summer Sale," and then generate three different color schemes for me to choose from.' Then, you go make a cup of coffee, and when you come back, everything is done. That's so cool! This will literally double the productivity of many design studios or one-person companies in Taiwan. Of course, it also means that if all you know is how to operate the software, you should really start to worry."

5. Nature Study Finds AI Models "Subconsciously" Pass on Biases When Training Other AIs

Source: Nature (https://www.nature.com/articles/d41586-026-01224-1)
Summary: A study in the top journal Nature has sounded an alarm: when we use AI-generated data to train the next generation of AI, the biases or specific tendencies of the old model are embedded into the new model like "subconscious signals," and they are very difficult to detect and remove. This raises serious concerns about the health of the AI ecosystem.
Taiwan Perspective: This is an important reminder for teams in Taiwan that are developing their own large models. It's like an "AI food safety problem." If we feed our models large amounts of unverified data generated by foreign AIs, our own models could be unknowingly "poisoned," inheriting biases that don't align with local culture or values.
Discussion Points:
- With AI-generated data flooding the internet, will this lead to "inbreeding" and degeneration of future AI models?
- How can we establish a mechanism to detect and filter out this "subconscious bias"?
- Does this mean that high-quality, human-verified data will become increasingly valuable in the future?
Suggested Script: "After talking about so many bright sides of AI, let's look at an issue that's a bit unsettling the more you think about it. This Nature study is basically saying that AI might 'inbreed' and degenerate. The internet is now filled with AI-generated text and images. If we take this 'second-hand data' to feed new AI models, it's like letting a child learn about the world only by watching cartoons—their understanding will become very one-sided and full of biases. The study found that this bias transmission is 'subconscious' and hard to detect. This is a warning sign for us, especially as Taiwan develops its own foundational models. The purity of data sources is crucial. This might also mean that in the AI era, the companies that possess unique, high-quality, real human data hold the real moat."

Closing Paragraph

Alright, that's it for today's feast of news. We've seen the clash of the titans between Meta and Google, and we've seen how AI is creating wealth at an unprecedented rate, but also bringing with it a widening wealth gap and potential ethical risks. This AI revolution is both an opportunity and a challenge. Thanks for listening, and we'll see you on the next show!