🎙️ AI 日報 Podcast — 模型大戰、世界模型、企業現實

Opening Remarks

Hello everyone, and welcome to "Mark's Tech Insights"! I'm Mark's co-host. Wow, this April has been a month of bombshell updates in the AI world, with almost all the major players rolling out their latest weapons. Today, we'll be talking about the latest model war between OpenAI and DeepSeek to see who will be the next king. At the same time, we'll also look at what big shots like Jeff Bezos are planning for the "post-LLM era"—what exactly are these so-called "World Models"? Finally, we'll return to the harsh business reality and see why over 70% of enterprises are struggling with their AI adoption.

Today's Top Stories

1. OpenAI Announces GPT-5.5: Higher Intelligence, Same Latency, but Double the Price

Source: CNBC (Original Link)
Summary: OpenAI has launched its latest model, GPT-5.5. Impressively, its response speed is just as fast as the previous generation, but its "IQ" has significantly increased. However, there's no such thing as a free lunch. The API price has doubled, clearly targeting enterprise users who need advanced coding and research analysis.
Taiwan Perspective: For startups and developers in Taiwan, this means the cost ceiling for AI applications has been raised again. Striking a balance between performance and cost will be a major challenge.
Discussion Points:
- What impact will the price doubling have on the AI ecosystem? Will it accelerate the development of open-source models, or will it make the strong even stronger?
- Is OpenAI leveraging its market leadership to start segmenting customer tiers by price?
- For general users, does this mean that the models behind free or low-cost AI services will be updated more slowly?
Script Suggestion: "Hey, did you see the news from OpenAI? GPT-5.5 is here, but so is the price—it's doubled! It's really interesting; Sam Altman seems to be saying, 'Want the most powerful intelligence? Then pay up.' I think this sends a very strong signal to the market. In the past, everyone was comparing who was cheaper. Now, OpenAI has shifted the battlefield to the 'premium performance, premium price' segment. For many of our startups in Taiwan that are developing services with AI, this is a bittersweet dilemma. You either keep up and use the most powerful model, but double your burn rate, or you have to find open-source alternatives like DeepSeek, which might require more effort to fine-tune. Does this mean the AI 'arms race' has entered a new phase, starting to use price to segment user tiers?"

2. DeepSeek V4 Bursts onto the Scene, Open-Source Agent Capabilities Take Aim at OpenAI

Source: Bloomberg (Original Link)
Summary: Just as everyone was discussing OpenAI's new pricing, DeepSeek from China launched its latest flagship open-source model, V4. This time, V4 shows significant improvements in knowledge, reasoning, and especially "Agentic" capabilities (autonomous agents), directly competing with closed-source models from OpenAI and Anthropic.
Taiwan Perspective: DeepSeek's powerful open-source model is a huge boon for Taiwan's AI development community, providing another high-performance, highly controllable option, especially for developing Agent-related applications.
Discussion Points:
- Will the catch-up of open-source models in Agentic capabilities change how enterprises choose AI models?
- How will potential investments from Tencent and Alibaba (News Item #4) accelerate DeepSeek's development and impact the global AI landscape?
- What challenges and opportunities will a powerful open-source model from China face under the current geopolitical climate?
Script Suggestion: "Right after talking about OpenAI, the next protagonist takes the stage, and this one comes with the halo of 'open source.' The highlight of DeepSeek V4 this time is its 'Agentic Capabilities.' Simply put, it makes the AI more like an intelligent assistant that can think and act independently. This is super important! Because this is exactly the 'AI employee' concept that businesses are most eager to realize. By choosing to open-source this capability, DeepSeek is essentially giving away its most powerful weapon for free. This is such a smart strategy to counter OpenAI! You can imagine that developers in Taiwan can now take DeepSeek V4 and build all sorts of custom AI Agents, from automated customer service to writing code, without being held hostage by OpenAI's pricing. Plus, with Tencent and Alibaba looking to invest, DeepSeek feels like the next contender for the throne in the AI world."

3. The Post-LLM Era Arrives? Jeff Bezos Pours $10 Billion into Building "World Models"

Source: Bloomberg / The Washington Post (Original Link 1, Original Link 2)
Summary: Amazon founder Jeff Bezos is raising nearly $10 billion for his new AI startup, which aims to develop AI models that can understand the "physical world." These models, known as "World Models," not only process text but also simulate real-world environments and interactions. They are considered the next AI paradigm after Large Language Models (LLMs).
Taiwan Perspective: Taiwan's hardware manufacturing, robotics, and autonomous vehicle industries are the most direct application scenarios for "World Models." This new trend could provide a perfect opportunity for Taiwan's hardware advantages to integrate with AI software.
Discussion Points:
- What is the fundamental difference between "World Models" and the LLMs we are familiar with now? Why are they so important?
- From Sora to Bezos's new project, how will advancements in visual and physical simulation capabilities change the application areas of AI?
- How can Taiwan's industrial chain, such as TSMC's chip manufacturing and Foxconn's robotics assembly, position themselves in this new race?
Script Suggestion: "While everyone is still caught up in the LLM rat race, Jeff Bezos is already looking at the next battlefield. He's going to drop $10 billion to build 'World Models.' It sounds very sci-fi, but the concept is straightforward: it's about letting AI not just 'read books' but also 'see the world.' It can understand physical laws, predict object movements, and simulate various real-world scenarios. For Taiwan, the opportunity here might be even bigger than with LLMs! Think about it, what's our strength? It's hardware, it's manufacturing. Whether it's robotic arms in automated factories or self-driving cars, they all need a 'brain' that understands the physical world. Isn't that what a World Model is? If Taiwan's hardware companies can catch this wave and combine our manufacturing prowess with this new type of AI, it's no longer just about contract manufacturing; it's about creating a whole new smart hardware ecosystem. This could be the next huge engine for Taiwan's tech industry transformation."

4. The Harsh Reality of AI Adoption: Nearly 80% of Enterprises Struggle, Less Than 30% See ROI

Source: Writer / Deloitte & PwC (Original Link 1, Original Link 2)
Summary: Despite frantic enterprise investment in AI, two recent reports reveal a harsh reality. A staggering 79% of companies report significant difficulties in adopting AI, and only 29% have seen a significant return on investment. More interestingly, PwC's research found that 75% of the economic benefits created by AI are captured by the top 20% of companies, creating a severe "rich-get-richer" gap.
Taiwan Perspective: This is a wake-up call for Taiwanese enterprises that are actively promoting digital transformation. Simply purchasing AI tools is far from enough; the key to success lies in changing organizational culture and re-engineering workflows.
Discussion Points:
- Why do most companies fail to see results despite pouring money into AI? Where is the problem?
- The report mentions that AI is "tearing companies apart." What specific internal conflicts does this refer to? Contradictions between technical and business departments?
- What did the successful 20% of companies do right? How did they shift from a "productivity enhancement" mindset to a "new growth creation" mindset?
Script Suggestion: "After talking about all the shiny new AI technologies, we have to come back to reality, and reality... is a bit harsh. The latest reports say that nearly 80% of companies find AI adoption extremely painful, and less than 30% are making money from it. Even more, 54% of executives admit that AI is on the verge of 'tearing their company apart.' It's really ironic, isn't it? Everyone is shouting 'AI or DIE,' but it turns out for many, it's 'AI and DIE.' PwC's report is even more direct, stating that 75% of the benefits are taken by the top 20% of players. It's like only the top students in a class truly understand the new math, while everyone else is just copying homework. The lesson for bosses in Taiwan is that AI isn't something you can solve just by buying software or an API. It's more like an 'organizational heart transplant.' You have to change processes, change culture, and even change how you measure performance. Otherwise, AI will just be an expensive toy."

5. Two Different Takes on AI Regulation: US Leans Against New Laws, Japan Targets Anime Image Theft

Source: White House & Transparency Coalition (Original Link 1, Original Link 2)
Summary: In AI regulation, countries around the world are taking different paths. The White House's policy framework recommends not to create a new AI regulatory agency, but rather to manage it through existing agencies and industry standards, hoping that federal law will preempt state regulations. Meanwhile, Japan's Liberal Democratic Party is drafting specific legislation to penalize the use of generative AI to misuse anime characters or modify someone's photo without consent.
Taiwan Perspective: As Taiwan formulates its AI-related laws, it is in the stage of referencing various national models. The US's "big market, loose regulation" approach and Japan's "protecting IP, prioritizing rights" approach represent two very different philosophies that are worth our deep consideration.
Discussion Points:
- Is the US strategy of "federal preemption, no new laws" good or bad for AI innovation?
- Will Japan's legislation targeting anime IP and personal likeness rights become a global standard for the content industry?
- Should Taiwan's AI Basic Act lean more towards the open model of the US or the protective models of Europe and Japan?
Script Suggestion: "Finally, let's talk about regulation. The recent moves by the US and Japan perfectly illustrate two different styles. The US is like a 'liberal' parent, saying, 'Kids, go explore. I trust you'll figure out the rules on your own, and the government will try to stay out of it.' They want a single national standard to prevent a chaotic patchwork of state laws from messing up the market. Japan, on the other hand, is like a 'strict' parent. Seeing people use AI to mess with other people's anime characters or create malicious photo edits, they immediately say, 'No! We need a law for this!' directly protecting their national treasures—anime IP and personal privacy. Neither approach is absolutely right or wrong, but they are very valuable references for Taiwan. We need the innovative vitality of the US, but we also need to protect our creators and citizens like Japan does. Finding that golden balance between development and regulation is truly the million-dollar question."

Closing Paragraph

Alright, today we've covered everything from the model war between OpenAI and DeepSeek to the new "World Model" frontier that Jeff Bezos is exploring, and finally returned to the real-world challenges of enterprise AI adoption and the regulatory approaches of different countries. It's clear that the development of AI has entered a deeper, broader, and more complex stage. Thanks for listening, and we'll see you next time!