(Photo by The New York Public Library on Unsplash)

✳️ tl;dr

• On 2025-09-15, Anthropic released its third Economic Index (approaching from different dimensions), tracking Claude usage patterns across 150+ countries and all US states for the first time. ¹
• (Possibly the first comprehensive geographic distribution data of AI adoption in the model industry?!)
• Enterprise API customers show an automation rate of 77%, significantly higher than consumer users’ 50%, indicating that enterprises are actively shifting AI from collaborative tools to productivity replacement solutions.
• Directive automation jumped from 27% to 39% within 8 months, marking the first time automation (49.1%) surpassed augmentation (47%), reflecting growing user confidence driven by improved model capabilities.

• API usage shows only 3% price sensitivity (each 1% increase in cost index reduces usage by only 0.29%), with enterprises prioritizing capability and value over cost,
• The speculated reason is that hidden infrastructure costs far exceed model fees (every $1 in model fees requires an additional $5-10 to deploy and reach production-ready status).
• Ernest’s field observations align with this: those who complain about token costs typically lack sound organizational operational systems or workflows. Conversely, those who see the overall value created are bold in adopting AI.
• Approximately 5% of API traffic is dedicated to developing and evaluating AI systems, forming a recursive improvement loop of “AI developing AI,” which is speculated to accelerate capability advancement but also require stronger safety oversight.
• US interstate GDP elasticity (1.8) is significantly higher than cross-country (0.7), yet income has lower explanatory power, indicating that industry composition and economic structure are stronger adoption drivers.

• AUI = Anthropic AI Usage Index
• Washington DC has the highest AUI (3.82), primarily for document editing and information search; California (third) focuses on programming; New York (fourth) prefers financial tasks, with local economic structures directly mapping to AI usage patterns.
• Educational instruction tasks grew 40% (9% → 13%), scientific research grew 33% (6% → 8%), showing rapid adoption in knowledge-intensive fields, suggesting that high-skilled workers are leveraging AI to enhance professional capabilities.
• Business management tasks declined 40% (5% → 3%), financial operations tasks halved (6% → 3%), suggesting these fields may be undergoing automation or users are shifting to more specialized tools.
• Wealthy countries tend to use AI for augmentation, while poorer countries prefer automation, with each 1% increase in population-adjusted usage corresponding to approximately 3% reduction in automation after controlling for task mix.

• The research uses privacy-preserving classification methods combining the O*NET database (19,498 task descriptions) and Claude’s proprietary classification system for dual verification, ensuring data anonymization.
• However, its static nature and coarse-grained classification may fail to capture emerging tasks created by AI and programming work of varying complexity. ²³
• The true cost of enterprise AI includes data engineering, security compliance, continuous monitoring, and integration architecture, far exceeding surface-level API fees, which is speculated to explain why enterprises are price-insensitive. ⁴⁵
• Model capability improvements (Sonnet 3.6 → 4.x series) directly drive behavioral changes, with better output quality reducing iteration needs, suggesting that future more powerful models may further increase automation ratios and transform human-AI collaboration patterns.

^{1 4 5 2 6 7 3 8 9}

Read More

(Caption: Early morning at a coffee shop—croissant, coffee, and conversations about AI Agents and life. Taken at Anchorhead Coffee, Seattle. Image source: Ernest.)

✳️ tl;dr

• Back in late August, I traveled to North America thinking I’d catch the tail end of summer. Instead, everywhere I went—Bay Area, Seattle—everyone was talking about AI Agents and Agentic Workflows.
• I was fortunate to get hands-on with Amazon Bedrock AgentCore after the AWS New York Summit.
• Lucky for me, I had presented on Firecracker microVMs at COSCUP a few years back, so I already knew how powerful and secure this isolation technology is.
• Back then, I spun up an i3.metal EC2 bare metal instance and managed to launch 4,000 microVMs with Firecracker in under 90 seconds ¹, visualizing the entire boot process. Startup speed? Not a concern. (Though it depends on your use case—but I’d argue we don’t need to remind AI Agents that “haste makes waste” :p)
• Even Cloudflare Containers borrowed Firecracker’s open-source project to power their services ².

• Today (2025-10-13), AWS officially launched Amazon Bedrock AgentCore—an enterprise-grade agentic platform designed to help organizations move AI agents from prototype to production ³.
• The AgentCore SDK has been downloaded over 1 million times, with early adopters including Clearwater Analytics, Ericsson, Sony, Thomson Reuters, and other cross-industry enterprises.
• Built on microVM technology for enterprise-grade security isolation—each agent session runs in its own isolated virtual machine instance, preventing data leakage and cross-tenant attacks.

• AgentCore offers composable services supporting multiple frameworks: CrewAI, Google ADK, LangGraph, LlamaIndex, OpenAI Agents SDK, Strands Agents, and more.
• Works with models on Amazon Bedrock, as well as external models like OpenAI and Gemini.
• AgentCore Code Interpreter enables agents to safely generate and execute code in isolated environments.
• AgentCore Browser allows agents to interact with web applications at scale.
• AgentCore Gateway transforms existing APIs and AWS Lambda functions into agent-compatible tools.
• Gateway connects to existing MCP servers and integrates third-party tools like Slack, Jira, Asana, and Zendesk.
• AgentCore Identity enables agents to securely access and operate various tools using OAuth standards.
• AgentCore Memory helps build context-aware agents without managing complex memory infrastructure.

• Provides industry-leading security through microVM technology, with each agent session in its own isolated compute environment.
• AgentCore’s MCP server integrates with IDEs like Kiro or Cursor AI.
• Offers an industry-leading 8-hour runtime for long-running tasks.

• Now that it’s GA (Generally Available), no more waiting in queue—just spin it up and start playing!

• Deploying AI Agents requires integration with existing workflows, fine-tuning, and alignment with organizational goals.
• For those interested in Process Automation whiteboards, check out the extended reading ⁴.

• P.S. On my return trip through Tokyo in September, I actually got hit by the tail end of summer there—scorching hot… Major respect to the Japanese salarymen in full suits. Orz…

Read More

(Illustration: Preparing to unbox the Remarkable Paper Pro. Image source: Ernest.)

✳️ tl;dr

• In Ernest PKM ¹, I mentioned that I still maintain paper-based notes, along with using digital handwritten notes.
• Last year, I acquired a reMarkable Paper Pro, which is also handwriting-based.
• There’s always this feeling that when surrounded by heaps of mixed information, whenever I want to quiet down and clarify the complex cognition at hand,
• I usually grab a handwriting tool, put on noise-canceling headphones, sit quietly for a few minutes, and then start writing to output, categorize, and compare,
• I can always sort things out. Even if no clear structure emerges, at least some branches grow.
• But feelings are just feelings. I always want to find a causal explanation (ah, is this a bad habit? Anyway, it’s a habitual action - the root cause habits formed at TSMC are too deeply ingrained).

• Searching and searching, I found that Dr. Audrey van der Meer ² has long been focused on this field,
• Below are her research results published in 2024-01. ³⁴
• The study recorded brain electrical activity in 36 university students as they were handwriting with a digital pen and typing on a keyboard.
• Brain connectivity patterns during handwriting were far more complex and elaborate than during typing.
• Handwriting produced widespread theta/alpha frequency connectivity patterns, which are crucial for memory formation and learning.

Read More

(Photo taken during my visit to Amazon Spheres, with The Westin Building visible on the far left. Image source: Ernest.)

✳️ tl;dr

• I’ve passed by this building countless times while attending meetings or exploring coffee shops and restaurants in downtown Seattle, never realizing its significance.
• The Westin Building Exchange is a major telecommunications hub facility located in downtown Seattle, Washington. ¹
• The building was constructed in 1981 (around the same era as me XDD),
• Originally named The Westin Building, it served as the headquarters for the Seattle-based Westin Hotel chain.
• It also houses the Seattle Internet Exchange (SIX) and the Pacific Northwest Gigapop Pacific Wave Exchange.

• Since 2019 or earlier, heat generated by the building’s data center has been piped to Amazon’s Doppler building next door, providing heating for Doppler and several other Amazon buildings.
• Amazon estimates that over the system’s expected 25-year lifespan, it will save 80 million kilowatt-hours of electricity, equivalent to 65 million pounds of coal.

• The trigger: I came across The Verge’s report on Microsoft’s plan to migrate GitHub entirely to Azure, which made me curious about its infrastructure before the acquisition. ²
• GitHub was acquired by Microsoft in 2018. So I went back to check the situation in 2017.
• GitHub’s blog once described: “Those facilities don’t store customer data, rather they’re focused on internet and backbone connectivity as well as direct connect and private network interfaces to Amazon Web Services.” ³
• GitHub is now (2025) undergoing what may be its largest infrastructure migration ever, planning to move the entire platform from self-owned data centers to Azure within 24 months.
• The core driver of this migration is the explosive growth of AI workloads: GitHub Copilot generates millions of code suggestions daily, consuming massive computing capacity, and the existing data centers have reached their physical expansion limits. (There’s always an official narrative for the public—take it with a grain of salt) (There’s a kind of cold where grandma thinks you’re cold. There’s a kind of migration where ____ tells you to migrate?!)
• Migration strategy: Most work needs to be completed within 12 months (because new and old systems need to run in parallel for at least 6 months), GitHub has asked teams to delay feature development and prioritize infrastructure migration.

• In 2024, GitHub experienced 119 incidents, including 26 major outages, with an average resolution time of about 106 minutes. ⁴
• Curious to see how it performs after the move.
• For those of you tech managers using GitHub, I’m curious—what’s your take? Will this trigger any preparations on your end?

^{5 6}

Read More

(Image source: Arduino UNO Q.)

✳️ tl;dr

• Qualcomm acquired Arduino, marking the third strategic acquisition following Edge Impulse and Foundries.io, demonstrating Qualcomm’s determination (or strategic positioning?) to build a fullstack edge AI platform ^{1 2}
• Arduino boasts over 33 million active community members, a developer foundation that competitors like Raspberry Pi don’t have.

• The newly launched Arduino UNO Q adopts a “dual-brain architecture”: Qualcomm Dragonwing QRB2210 quad-core processor (Quad-core Arm Cortex-A53 @ 2GHz) paired with STM32U585 real-time microcontroller (Arm Cortex-M33 @ 160MHz), simultaneously handling high-performance computing and real-time control.
• Wireless connectivity = dual-band Wi-Fi 5 + Bluetooth 5.1
• Edge Impulse integration provides AutoML capabilities, enabling even non-AI experts to deploy machine learning models.
• The TinyML market is projected to reach $200 billion by 2030, and Qualcomm is positioning itself to capture this high-growth segment?!
• UNO Q is priced at $44 (2GB RAM/16GB eMMC) to $59 (4GB/32GB), benchmarking against Raspberry Pi 5 at about $7 less. (But RPi 5 now offers 8GB RAM and 16GB RAM options, worth tracking going forward.)

• Software is licensed under GPL 3.0 or Mozilla Public License, hardware design under CC BY-SA 4.0, meaning you can legally create and sell derivative versions

^{3 4 5}

Read More

(Photo by Rikin Katyal on Unsplash)

✳️ tl;dr

• AWS introduces ECS Managed Instances, achieving the optimal balance between operational simplicity and flexibility by offloading infrastructure management to AWS while maintaining full EC2 control ¹
• Runs on Bottlerocket OS, which maintains only ~100 package definitions compared to general-purpose operating systems with 50,000 packages, significantly reducing attack surface and management complexity ²
• Protects the root filesystem with dm-verity and SELinux enforcing mode, making it difficult to persist attacks even after container escape, with automatic system restart upon tampering detection ³⁴

• The container orchestration market is projected to grow from $10.8 billion in 2025 to $76.5 billion by 2034, with a CAGR of 24.16%, demonstrating strong demand for managed container services ⁵
• Research shows that heterogeneous task allocation strategies can reduce container orchestration costs by 23% to 32%, with ECS Managed Instances’ automatic workload consolidation being key to achieving this goal ⁶

• Bottlerocket’s atomic update model reduces the time to patch critical vulnerabilities from days or weeks to hours, potentially cutting update-related downtime by 80% compared to traditional systems ⁷⁴
• The service is currently available in six AWS regions, including US East (North Virginia), US West (Oregon), Europe (Ireland), Africa (Cape Town), Asia Pacific (Singapore), and Asia Pacific (Tokyo), with plans to expand to more regions to support global deployment needs
• Supports deployment through AWS Management Console, CLI, CDK, and CloudFormation, seamlessly integrating with existing DevOps toolchains to lower adoption barriers
• In 2022, Ernest shared “Running Laravel/PHP on AWS” at AWS Builders Day Taiwan, comparing various Amazon ECS Launch Types. Looks like it’s time to update those slides. ⁸

⁹¹⁰¹¹

Read More

(Photo by Brett Jordan on Unsplash)

✳️ tl;dr

• Claude Code introduces Checkpoints feature, enabling real-time progress saving and rollback to previous states, solving pain points in long-term development ¹²
• Native VS Code extension released, supporting real-time inline diffs display of Claude Code’s code changes
• Claude Sonnet 4.5 achieves 77.2% on SWE-bench Verified (full 500 problems), using simple bash and file editing tools ¹³

• High-compute configuration with parallel testing reaches 82.0%, using rejection sampling and internal scoring models to select best candidates ¹
• API adds memory tools and automatic context management, allowing agents to maintain context across long-running tasks ¹⁴
• Pricing remains the same as Claude Sonnet 4: $3 per million input tokens, $15 per million output tokens, with prompt caching saving up to 90% costs ⁵

• Cursor CEO states Claude Sonnet 4.5 demonstrates state-of-the-art programming performance on long-horizon tasks, making it the top choice for developers solving complex problems ¹

• Achieves 61.4% on OSWorld benchmark, leading all competitors; Sonnet 4 was only 42.2% four months ago
• Implements AI Safety Level 3 (ASL-3) protection framework, CBRN threat detection false positive rate reduced tenfold compared to initial description
• Makes progress in prompt injection attack defense, one of the most serious security risks facing agentic AI systems
• First inclusion of “Interpretability” technical assessment in System Card, improving model transparency and credibility ⁶

⁷

Read More

(Photo by Jonny Gios on Unsplash)

✳️ tl;dr

• New toy Claude for Chrome ^{1 2}
• Signed up and joined the waitlist a long time ago
• Received the invitation email yesterday
• Found some time to play with it today
• Need to continue working on my tasks
• Wait, shouldn’t I ask it to help me with my tasks?!

• Looks like I’m in the second batch of Max plan users getting Claude for Chrome research preview access
• Tried a few tasks with Claude for Chrome during the holiday
• It requests permission for a specific domain in the URL first, and you can approve step by step

• If I were to read the documentation and operate manually by myself, especially when unfamiliar (or doing it for the first time), it would take about 2~3 minutes (reading while operating on dual screens) to complete the task. But it requires mental processing and finding things on the screen (or using cmd+f to find action links or buttons to proceed).
• In comparison, Claude for Chrome takes about 10~15 minutes, but it did complete the tasks (for example, helping me set up GTM, helping me dig through Quickbooks to find the transaction history of a specific Account). This is because it needs to operate step by step - each step requires a screenshot, interpreting the screen and deciding where to click next, some screens require typing text into input fields, and even after clicking a text box it takes another screenshot to confirm the focus is on the text box (I almost impulsively wanted to help it click several times, but held back to avoid disrupting the flow).

• I’m currently using it with Edge (because I need vertical tabs and multiple isolated profiles), and it seems I encountered situations where Claude for Chrome changed my original tab group names, closed other tabs in the original tab group without asking me, and Claude for Chrome opened new tabs (e.g., GA4) in the tab group and started operating on its own, while I was still on the GTM tab on my screen.
• From an experience perspective, of course at this stage it’s not efficient.
• But from a results perspective, Claude for Chrome did complete the tasks, which means the reasoning for breaking down steps and the knowledge base are sufficient. We just need to wait for next-generation models or possible future local infrastructure to reduce the round-trip time, and efficiency will directly improve.

• Are you also thinking about how to introduce AI Agents or Agentic Workflows into your team or workflow? Feel free to chat.

Read More

Problem Description

Recently, while playing with Kiro (Agentic IDE) and experimenting with multiple-role agents and hooks, I encountered a quite frustrating problem: within the same CLI session, Kiro couldn't retrieve the return values from the second command.

The problem scenario was as follows:

• When I asked Kiro to execute the first command, everything worked perfectly - the command executed successfully and could retrieve the output correctly.
• However, when executing the second command within the same shell session, Kiro would be unable to get the return value from that command.
• So Kiro would just hang there waiting for the response, unable to get the vibe going.

This problem significantly impacted my workflow:

• Frequent session restarts:
  • After executing the first command, I had to manually close the CLI session each time. If I need to intervene this frequently as a human, that’s not the AI agent I want.
• Manual output copying:
  • Or I needed to manually copy the return values from the second command to Kiro Chat.
• Reduced work efficiency:
  • The originally smooth agentic coding workflow became fragmented. While I can relax when it’s time to relax, for work tasks, I want to improve efficiency.

Problem scenario video record (some key information in the video has been regenerated, please don’t panic):

Read More

(Photo by Felix Eka Putra Kuntjoro on Unsplash)

Background

I just thought of switching the post-processing for MacWhisper Dictation from the original google/gemma-3-12b to using openai/gpt-oss-20b in LM Studio, but I kept encountering an issue where the gpt-oss model would return the reasoning process as part of the dictation output results. Here’s the problematic output:

We need to correct punctuation: use full-width. 
Input: "嗨,我們明天去兒童樂園玩好嗎?" 
We replace comma with ，, question mark with ？. 
Also add period at end? 
The sentence ends with question mark already. 
So output: "嗨，我們明天去兒童樂園玩好嗎？"
嗨，我們明天去兒童樂園玩好嗎？

Read More