🎙️ AI 日報 Podcast，GPT-5.6 降臨、三星與 Anthropic 2奈米合作、歐盟 AI 法案生效

Intro

Hello everyone, and welcome to "Mark's Tech Insights"! I'm Mark. Today's episode is absolutely packed with information. We're going to talk about OpenAI's surprise release of the GPT-5.6 "Ragnarok" dual lineup—featuring both budget and flagship models. Also, Anthropic is reportedly bypassing TSMC to talk with Samsung about custom 2nm chips? Finally, are you ready for the EU AI Act, which officially goes into effect in August with eye-watering fines? Grab a seat, and let's dive right in!

Today's Top Stories

1. OpenAI Launches GPT-5.6 Three-Tier Lineup: Sol Focuses on Top-Tier Reasoning, Terra Cuts Price in Half

Source: ThursdAI (https://thursdai.news/releases/2026-07）
Summary: OpenAI officially launched the GPT-5.6 series of models in July 2026. This release adopts a tiered product strategy: "Sol" targets cutting-edge coding, scientific, and logical reasoning tasks, while "Terra" offers near-GPT-5.5-level performance at roughly half the price, clearly aiming to capture the high-value market currently held by Claude Sonnet and Gemini Flash.
Taiwan Perspective: This is a major boon for Taiwan's many budget-conscious startups and indie developers. Previously, wanting to use top-tier models meant watching USD API bills bleed you dry. Now, with a "bang-for-the-buck" option like Terra, development costs are cut in half.
Discussion Points:
- Will OpenAI's strategy of "tiered user harvesting... wait, I mean tiered pricing" become the standard playbook for large models in the future?
- How will Anthropic and Google respond to Terra's price war?
Script Suggestion: Have you ever heard the saying, "When buying a car, look at the specs; when buying a model, look at your wallet"? OpenAI's GPT-5.6 release is truly something else. Instead of incremental updates, they went straight for a tiered approach. The top-tier "Sol" is like a supercar, built specifically for incredibly difficult coding and scientific computing. But for us everyday developers, the budget-friendly "Terra" is where the real value is! It gives you near-flagship performance at half the price. This move directly targets the pain points of Claude and Gemini. Many AI application teams in Taiwan can finally stop arguing with their finance departments over API costs—just go ahead and use Terra!

2. Anthropic in Talks with Samsung to Develop Custom 2nm AI Chips

Source: TechCrunch (https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/）
Summary: Anthropic is currently in early-stage talks with Samsung to build its own custom AI chips using Samsung's latest 2nm foundry process. This move mirrors the in-house chip strategies of Google, Apple, and Meta to reduce reliance on NVIDIA, and represents a major milestone for Samsung's foundry business.
Taiwan Perspective: This news is quite sensitive for us in Taiwan! While everyone has their eyes on TSMC's advanced nodes, Anthropic is approaching Samsung for 2nm. Although this won't shake TSMC's foundations, it is definitely a wake-up call.
Discussion Points:
- Why did Anthropic choose to approach Samsung instead of partnering with TSMC this time?
- Can Samsung's 2nm yield and capacity actually satisfy Anthropic's massive computing ambitions?
Script Suggestion: As soon as this news broke, Taiwan's tech community was abuzz. We all know that NVIDIA currently has a stranglehold on AI chips, which is why all the tech giants want to design their own. But this time, Anthropic actually bypassed TSMC—our "sacred mountain protecting the nation"—to talk 2nm foundry work with Samsung! This is an absolute godsend for Samsung. However, speaking objectively as a Taiwanese, Samsung's advanced node yields have historically been as unpredictable as a mystery box. Will Anthropic's gamble successfully break NVIDIA's grip, or will they get choked by Samsung's yield rates? Let's grab some popcorn and see how this plays out.

3. Moonshot AI Launches Kimi K3: World's Largest 2.8-Trillion Parameter Open-Weights Model

Source: VentureBeat (https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems）
Summary: Chinese AI startup Moonshot AI has unveiled Kimi K3, a 2.8-trillion parameter open-weights model featuring a 1-million-token context window and native multimodal capabilities. In some third-party benchmarks, its performance even surpasses GPT-5.6, ranking second only to Claude Fable 5.
Taiwan Perspective: Even under strict US chip export restrictions, China is still able to stack existing compute to churn out monster-class models like this. This puts considerable pressure on Taiwan's positioning in cross-strait AI technology, especially regarding Traditional Chinese localized applications.
Discussion Points:
- How did China manage to train a 2.8-trillion parameter behemoth despite the compute blockade?
- How should Taiwanese enterprises evaluate security and privacy risks when using open-weights models from China?
Script Suggestion: We really shouldn't underestimate this! Despite sweeping US compute sanctions, China's Moonshot AI has managed to release Kimi K3 with 2.8 trillion parameters. This parameter count completely shatters the ceiling for open-source models. While enterprises in Taiwan are highly cautious about using Chinese models due to security and geopolitical concerns, we have to admit that their software optimization under extreme compute constraints is incredibly impressive. This is a reminder that the AI race isn't just about hardware; squeezing every drop of performance out of software and algorithms is the other key to winning!

4. EU AI Act to Be Fully Enforced Starting August 2, 2026

Source: OriginBrief (https://www.originbrief.app/en/reports/ai-regulation-policy/2026-07-13/weekly）
Summary: The world's first comprehensive AI regulation, the EU AI Act, will officially go into effect on August 2, 2026. The act imposes binding risk-management requirements on high-risk AI systems and general-purpose AI models. Non-compliant companies face massive fines of up to 3% to 7% of their global annual turnover.
Taiwan Perspective: As a core hub of the global tech supply chain, many of Taiwan's hardware and software vendors export products to Europe. Once this act takes effect, Taiwanese manufacturers must immediately conduct compliance reviews, or risk accidentally crossing the red line into massive international fines.
Discussion Points:
- Will the EU's strict regulations stifle local AI innovation, ultimately benefiting the US and China?
- Will the Taiwanese government follow in the EU's footsteps and introduce similarly stringent AI regulation bills?
Script Suggestion: Software developers and business owners, the leash is tightening! The EU AI Act is officially coming into force on August 2. This isn't like clicking "I agree" on a website cookie banner and moving on—fines can go up to 7% of your global annual revenue! Many teams in Taiwan building AI apps or selling software to Europe absolutely need to start compliance testing now. People used to think of AI as the Wild West where you just move fast and break things, but now "compliance" has become a survival threshold as critical as ESG. Don't play games with your wallet!

5. Gartner Survey: 72% of Enterprises Have Deployed Agentic AI, but 60% Face a Governance Gap

Source: Agentic AI Institute (https://agenticaiinstitute.org/agentic-ai-enterprise-adoption-2026-governance-gap/）
Summary: According to Gartner's latest survey, while 72% of enterprises have deployed Agentic AI in production environments, only 10% have successfully scaled it beyond a single function. A staggering 60% of enterprises have governance gaps, and Gartner predicts that by 2027, over 40% of AI agent projects will be abandoned due to unclear ROI and regulatory failures.
Taiwan Perspective: When adopting AI, Taiwanese enterprises often fall into the cycle of "beautiful POCs, but disastrous production deployments." This report hits the nail on the head: cool technology alone isn't enough; management and error-prevention mechanisms are the real keys to success.
Discussion Points:
- Why is scaling Agentic AI across departments so difficult?
- How can enterprises establish effective AI governance mechanisms so that AI agents don't turn into "runaway horses"?
Script Suggestion: Everyone is talking about Agentic AI right now—letting AI send emails, schedule meetings, and even integrate systems for you. Gartner's survey shows that over 70% of companies have jumped on the bandwagon and deployed these AI agents, but surprisingly, only 10% have seen real cross-departmental impact. Why? Because everyone focuses on how cool the tech is, without thinking: what if this AI agent suddenly sends an angry email to a client or messes up the accounting? Who takes the blame? Taiwanese enterprises often run into this "all hype, no substance" situation. Without proper process governance, this wave of AI investment might just end up down the drain.

6. Boston Dynamics Partners with Google to Put Gemini 1.6 Visual Brain into Spot Robot Dog

Source: AIapps (https://aiapps.com/blog/july-ai-mega-update-major-breakthroughs-launches/）
Summary: Boston Dynamics has partnered with Google Cloud and DeepMind to integrate Gemini Robotics-ER 1.6 into the Spot robot dog and the Orbit AI visual inspection platform. This collaboration aims to enhance autonomous navigation, visual inspection, and real-time decision-making in industrial environments.
Taiwan Perspective: Taiwan possesses a robust smart manufacturing and industrial robotics supply chain. When robot dogs equipped with "multimodal brains" enter factories, Taiwan's semiconductor fabs and traditional manufacturing plants will be able to achieve truly unmanned inspection and maintenance.
Discussion Points:
- How will putting LLMs with semantic understanding into physical robots revolutionize the efficiency of industrial inspections?
- What role does Taiwan's hardware supply chain play in this wave of "Embodied AI"?
Script Suggestion: Finally, let's talk about Spot, the Boston Dynamics robot dog we all know so well. It has finally been fitted with Google Gemini's multimodal brain! Previously, robot dog inspections could only follow rigid code, getting stuck whenever they hit an obstacle. Now, with Gemini's visual understanding, if it spots a puddle on the floor during a semiconductor fab inspection, it can figure out on its own whether it's an abnormal leak or just a freshly mopped floor, and report it immediately. As a hub for smart manufacturing, if these brainy robots become mainstream in Taiwan, night patrols in factories could be fully handed over to robot dogs. Sci-fi movie scenes are literally coming to life right before our eyes!

Outro

Alright, that's all the major tech news we have for you today on "Mark's Tech Insights." From the GPT-5.6 price war and the undercurrents between Samsung and Anthropic, to the enforcement of the EU AI Act, the pace of the AI era is truly breathtaking—but it's also too exciting to look away. I hope today's episode got you fully charged! If you enjoyed the show, don't forget to subscribe and share it with your friends. See you next time, bye!