🎙️ AI 日報 Podcast — GPT-5.5、AI 淘金熱、類腦晶片

Opening Remarks

Hello everyone, and welcome to "Mark's Tech Insights"! I'm Mark. The AI world is absolutely buzzing today; it feels like all the money in the world is rushing into this space. We'll be talking about OpenAI's latest bombshell, GPT-5.5, looking at the insane valuations of AI coding tools, and finally, taking a step back for some sober reflection on AI's limitations and future hardware solutions.

Today's Top Stories

1. OpenAI Announces GPT-5.5, Revenue Soars Past $25 Billion as It Prepares for IPO

Source: CNBC (https://www.cnbc.com/2026/04/23/openai-announces-latest-artificial-intelligence-model.html)
Summary: OpenAI is keeping us all up at night again, launching GPT-5.5, which claims major advancements in coding, computer operation, and in-depth research. Even more astonishing, their annualized revenue has surpassed $25 billion, and they're reportedly considering an IPO.
Taiwan Perspective: For AI startups and developers in Taiwan, this means the barrier to entry for competing with foundational models is getting higher, but it also provides more powerful API tools to leverage.
Discussion Points:
- What exactly is GPT-5.5's "computer-use" capability? Is it an evolution of agents?
- What does $25 billion in annualized revenue even mean? What impact does this have on valuations in the AI industry?
- If OpenAI really goes public, what kind of changes will it bring to the entire AI ecosystem?
Suggested Script: Hey, did you see? OpenAI dropped another bombshell, just releasing GPT-5.5. They say it's even better at coding and research this time, but I think the most noteworthy part is the "computer operation" capability. This sounds a lot like the Agentic AI we've been talking about, where the AI doesn't just chat with you but can directly operate software and complete tasks for you. For us developers in Taiwan, this is both a pressure and an opportunity. The pressure is that OpenAI is evolving so fast it's hard to keep up; the opportunity is that their stronger APIs allow us to develop more killer apps. But then again, with $25 billion in annualized revenue and a potential IPO, they're basically a money-printing machine. Do you think this will inflate the AI bubble even more?

2. Chinese Startup DeepSeek Open-Sources V4 Model, Challenging Closed-Source Giants

Source: MIT Technology Review (https://www.technologyreview.com/2026/04/24/1136422/why-deepseeks-v4-matters/)
Summary: A Chinese company called DeepSeek has also released a V4 preview, and it's open-source! The most powerful version, V4-Pro, is said to have performance comparable to top-tier closed-source models and supports an ultra-long 1 million token context. They've also partnered with Huawei, using the Ascend 950 chip for computation.
Taiwan Perspective: A powerful open-source model from China, combined with a non-Nvidia hardware solution (Huawei's Ascend), is a variable that Taiwan's hardware supply chain and software ecosystem should watch closely.
Discussion Points:
- With open-source models catching up to closed-source ones in performance, what impact will this have on OpenAI and Google?
- What does DeepSeek V4's 1.6T parameters (49B activated) mean? Has the MoE architecture truly gone mainstream?
- How important is the role of Huawei's Ascend chip here? Does this signal a potential "de-Nvidization" of AI computing power?
Suggested Script: Right after talking about OpenAI, we have to look at the challenger from across the strait. This company, DeepSeek, is really coming in strong, open-sourcing its V4 model, and the benchmarks look amazing—they claim it can compete with GPT-4 level models. Two things are most interesting: first, it supports a 1 million token context, which can handle extremely long documents or codebases, opening up many use cases. Second, they're partnering with Huawei and using the Ascend 950 chip. For Taiwan, this sends a complex signal. On one hand, we see another powerful option for the open-source community. But on the other, an AI behemoth based on a non-US tech stack is taking shape, from chip to model. Could this be a warning sign for Taiwan's tech industry, which is heavily reliant on contract manufacturing and the US supply chain?

3. AI Coding Tools Explode: Cursor and Cognition Valuations Skyrocket

Source: CNBC (https://www.cnbc.com/2026/04/19/cursor-ai-2-billion-funding-round.html), Bloomberg (https://www.bloomberg.com/news/articles/2026-04-23/ai-coding-firm-cognition-in-funding-talks-at-25-billion-value)
Summary: The AI-assisted coding space is absolutely flooded with cash right now! The AI code editor Cursor is reportedly in talks for a $2 b i l l i o n f u n d i n g r o u n d a t a v a l u a t i o n t h a t c o u l d e x c e e d$ 50 billion. Another company, Cognition AI, which created the autonomous programmer Devin, is also rumored to be raising funds at a $25 billion valuation, several times its previous round.
Taiwan Perspective: Taiwan has many excellent software engineers. The rise of these AI agent tools could completely change the software development process and value chain. It's worth considering how to shift from "writing code" to "directing AI to write code."
Discussion Points:
- What are the differences in product positioning between Cursor and Cognition (Devin)? Why is the market willing to give them such high valuations?
- An AI coding tool valued at $50 billion—is that reasonable? Or is VC FOMO (Fear Of Missing Out) just too intense?
- When AI can code autonomously, what will be the value of human engineers?
Suggested Script: After talking about foundational models, let's look at the madness at the application layer, especially in the AI coding space. Did you know? An AI code editor called Cursor is rumored to be raising funds at a $50 b i l l i o n v a l u a t i o n!$ 50 billion! That's higher than the market cap of many large corporations. And then there's the recently famous AI engineer Devin; its developer, Cognition, has seen its valuation jump to $25 billion. This shows that the market is extremely bullish on the "AI Agent for Code" direction. For us engineers in Taiwan, this is truly a paradigm shift. In the future, it might not be about who writes the cleanest code, but who is better at writing prompts and who can better manage and direct a team of AI engineers. This also reminds us that maybe we should start cultivating "AI collaboration" skills, not just pure coding skills.

4. Google Getting Anxious? Sergey Brin Personally Forms a Strike Team to Catch Up to Anthropic's Agentic Coding Capabilities

Source: AI2Work (https://ai2.work/blog/google-s-brin-forms-deepmind-strike-team-to-close-anthropic-coding-gap)
Summary: Seeing all the buzz outside, Google is also making moves internally. Co-founder Sergey Brin has personally returned to the front lines, forming a strike team within DeepMind with a clear goal: to urgently catch up to Anthropic's capabilities in Agentic Coding. Internal documents state the objective is to make Google's model the "primary developer."
Taiwan Perspective: This shows that even a giant like Google is feeling immense pressure from the Agentic AI wave. Taiwanese companies and developers can see from this that agent technology is the next major battleground.
Discussion Points:
- Why did Google specifically name Anthropic, and not OpenAI or other companies?
- What does Sergey Brin's return to the front lines signify for Google's AI strategy?
- How is the goal of "making the model the primary developer" different from Cognition's Devin?
Suggested Script: We just talked about the insane valuations of Cursor and Cognition, and look, Google is immediately proving how anxious they are with their actions. Google's legendary co-founder, Sergey Brin, has reportedly had enough and personally formed a "DeepMind strike team." Their mission is super urgent: to catch up to Anthropic's ability in AI agent coding. The wording in the internal memo is intense, talking about "urgently closing the gap" and turning their own model into the "primary developer." This is interesting. They're not benchmarking against OpenAI, but Anthropic. This might mean that Anthropic has some specific agent capabilities that are a real thorn in Google's side. This once again confirms that the next battlefield in AI is definitely agents that can execute tasks autonomously.

5. Meta Isn't Holding Back! Annual AI Spending Could Reach $135 Billion

Source: CNBC (https://www.cnbc.com/2026/04/08/meta-debuts-first-major-ai-model-since-14-billion-deal-to-bring-in-alexandr-wang.html)
Summary: Meta is not to be outdone. After spending $14 b i l l i o n t o p o a c h S c a l e A I f o u n d e r A l e x a n d r W a n g, t h e y^{'} v e l a u n c h e d a n e w m o d e l, M u s e S p a r k . T o c a t c h u p w i t h G o o g l e a n d O p e n A I, M e t a^{'} s A I c a p i t a l e x p e n d i t u r e b u d g e t f o r t h i s y e a r i s a s h i g h a s$ 135 billion, showing incredible determination.
Taiwan Perspective: A large portion of that $135 billion in capital expenditure will go towards servers, chips, and data center construction, which represents a huge business opportunity for Taiwan's related supply chain (like TSMC, Quanta, and Wistron).
Discussion Points:
- Why did Meta spend a fortune to poach Alexandr Wang? What role does he play at Meta AI?
- What does $135 billion in capital expenditure even mean? What kind of AI empire does Meta plan to build with this money?
- How will the new Muse series of models differ from the original Llama series?
Suggested Script: Besides Google, another giant, Meta, is also going all out. They previously spent $14 b i l l i o n t o p o a c h S c a l e A I f o u n d e r A l e x a n d r W a n g, a n d n o w w e^{'} r e f i n a l l y s e e i n g t h e r e s u l t s w i t h t h e l a u n c h o f t h e n e w M u s e S p a r k m o d e l . B u t w h a t^{'} s e v e n m o r e t e r r i f y i n g i s t h e i r b u d g e t . G u e s s h o w m u c h M e t a^{'} s A I c a p i t a l e x p e n d i t u r e i s f o r t h i s y e a r ? I t c o u l d b e a s h i g h a s$ 135 billion! That number is insane; you could buy several large companies with that. Most of this money will be used to buy chips and build data centers. For us in Taiwan, this is a direct boon. Companies in the AI supply chain like TSMC, Quanta, and Wistron must be smiling to themselves when they see this news. This also shows that the AI war is no longer just a model competition, but an arms race based on deep pockets.

6. Is AI Still Far Off? Nature Study Finds Human Scientists Still Outperform AI Agents in Complex Scientific Research

Source: Nature (https://www.nature.com/articles/d41586-026-01199-z)
Summary: Amidst the AI frenzy, the prestigious journal Nature published a study that throws some cold water on the hype. The research found that when it comes to complex scientific tasks requiring long-term planning and iterative experimentation, human scientists still far outperform the most powerful current AI agents. This serves as a reminder to lower expectations for the emergence of fully autonomous AI scientists in the near term.
Taiwan Perspective: This is good news for Taiwan's academic and R&D sectors, as it indicates that human creativity, critical thinking, and experimental design skills will remain core values for the foreseeable future.
Discussion Points:
- How was this study designed? What "shortcomings" of AI did it test?
- Why are "long-term planning" and "experimental iteration" so difficult for current AI?
- Will this study affect the sky-high valuations of the AI agent companies mentioned earlier?
Suggested Script: After hearing so much about how amazing AI is and how high the valuations are, it's time for a reality check. The top scientific journal Nature recently published a study with a very direct conclusion: on truly complex scientific research tasks, human scientists still crush the most powerful AI agents of today. The study says that AI still doesn't perform well on open-ended problems that require long-term planning, continuous trial-and-error, and course correction. I think this point is super important. It tells us that AI right now is more like a super-powered intern. It can help you quickly look up information, write code, and analyze data, but that "from 0 to 1" insight, that "taste" which determines the entire research direction, still relies on humans. So, everyone can relax a bit; your job isn't going to be snatched by an AI scientist just yet.

7. A Solution for AI's Power Consumption? Cambridge University Develops Brain-Like Chip That Could Save 70% of Energy

Source: ScienceDaily (https://www.sciencedaily.com/releases/2026/04/260422044633.htm)
Summary: AI is powerful, but it's also a power-hungry monster. It's said that AI alone consumes over 10% of the electricity in the United States. Researchers at the University of Cambridge have developed a new nanoelectronic memristor that mimics the way the human brain processes information, claiming it can reduce the power consumption of AI systems by a massive 70%.
Taiwan Perspective: Taiwan is a semiconductor powerhouse. This new type of chip architecture is a future technological direction that TSMC and IC design companies must pay attention to. If they can master it, it will be their moat for the next generation.
Discussion Points:
- What is the principle behind the memristor? How does it differ from traditional CMOS chips?
- AI power consumption is already at 10% of the US total? What kind of challenge does this pose to energy infrastructure?
- How far is this new chip from moving from the lab to commercial mass production?
Suggested Script: Finally, let's talk about a piece of news I find super critical, and it's related to hardware. We all know AI is power-hungry. There are reports that AI's electricity usage has already surpassed 10% of the entire US total, which is terrifying. Cambridge University recently came up with a "brain-like chip" that uses a new component called a "memristor" to mimic the way our brains process information. They say this thing can save 70% of the power! If this can actually be commercialized, the impact will be enormous. For Taiwan, as the semiconductor kingdom, we absolutely cannot miss this kind of revolutionary breakthrough in underlying hardware. This isn't just a matter of saving on electricity bills; it could redefine the architecture of AI computing power. It's a critical point where our IC design and manufacturing industries could either leap ahead or be overtaken.

Closing Paragraph

Alright, today's news was incredibly dense, from the model arms race and VCs throwing money around like crazy, to the reality check from scientific research and the dawn of future hardware. All I can say is, you can't let your guard down for a single day in AI development. Thanks for listening to "Mark's Tech Insights." See you next time!