Opening

OpenAI's models are helping AMD modify GPU circuit designs, and some of those changes have already been submitted for manufacturing. Today, November 18, 2026, AI is doing more than writing code: it's helping make chips. Citibank has also updated its terms for AI agents accessing customer accounts. We'll look at who might have to pay if something goes wrong, as well as how Snapchat is letting creators make AR filters through conversation.

Today's Top Stories

1. OpenAI Models Help Advance Chip Design

  • Source: OpenAI (Original article)
  • Summary: OpenAI says its models helped AMD optimize circuit designs for its flagship GPU, improving some metrics by up to 14% over existing optimization baselines. AMD has included some of the changes in designs submitted for manufacturing, so the AI's suggestions have made it beyond the draft stage.
  • Most surprising point: Some of the circuit designs refined with AI assistance have made it into a version submitted for manufacturing.
  • Taiwan perspective: Taiwan has a comprehensive chip design and manufacturing industry. If AI can help engineers explore more design options, it could change parts of the R&D process. Improvements in individual metrics don't necessarily mean the entire chip is 14% faster, though; the actual impact will depend on specific tests.
  • Discussion points:
    • If AI is a design partner that proposes candidate solutions, how should engineers validate its suggestions?
    • How might greater circuit design efficiency change the skills chip engineers in Taiwan need?

2. More Banks Say Customers Using AI Agents May Have to Bear Losses Themselves

  • Source: TechCrunch (Original article)
  • Summary: Citibank has updated its terms to state that customers who authorize third-party AI agents to access their accounts are responsible for related activity. TechCrunch reports that customers could bear losses even if an agent tool leaks their credentials or is hacked.
  • Most surprising point: Letting an AI assistant manage a bank account may be convenient, but customers could also be left with the bill if something goes wrong.
  • Taiwan perspective: Developers in Taiwan building financial AI assistants need to make sure users understand the permissions they grant, alongside making the tools easy to use. Terms and their applicability may vary by bank, so they're worth reviewing carefully before use.
  • Discussion points:
    • How much account access does an AI agent need to get its job done?
    • When a third-party tool fails, how should responsibility be divided among the bank, the tool provider, and the customer?

3. Snapchat Lets Developers Build AR Filters With AI Assistants

  • Source: TechCrunch (Original article)
  • Summary: Snapchat has launched Easy Lens CLI, which lets creators use natural language to instruct AI assistants such as Claude Code and Cursor to create, test, and publish AR filters. Snapchat says a game can go from an idea to a published Lens in a single day.
  • Most surprising point: Snapchat says a game idea could become a live Lens in just one day.
  • Taiwan perspective: For brands and content creators in Taiwan, tools like this could make it easier for small teams to experiment with interactive campaigns. As the barrier to creation falls, how engaging and intuitive an experience is may matter even more in keeping people's attention.
  • Discussion points:
    • As natural language lowers the barrier to creation, what testing and design skills will creators still need?
    • How can brands in Taiwan use AR interactions to make campaigns more engaging than simply applying a filter?

4. Microsoft Ignite Announces On-Device and Cloud Options for AI Workloads

  • Source: Microsoft (Original article)
  • Summary: At Ignite on November 17, 2026, Microsoft announced that AI workloads on Windows can run on-device or in the cloud, depending on their needs. The first applications include GitHub Copilot for software development and Copilot for everyday tasks.
  • Most surprising point: AI workloads on the same Windows device will be able to run locally or in the cloud, depending on their needs.
  • Taiwan perspective: Taiwan is home to many companies in the PC and hardware supply chain, making the development of on-device AI worth watching. Businesses evaluating it will also need to understand where different workloads run and how their data is handled.
  • Discussion points:
    • Which workloads are suited to on-device processing, and which might need cloud resources?
    • How should developers help users understand where their data is processed?

5. Microsoft and Organizations Across Industries Launch AI Transparency Initiative

  • Source: Microsoft (Original article)
  • Summary: On November 17, 2026, Microsoft announced the AI Transparency Initiative with more than 40 organizations, including Adobe and Mastercard. The initiative aims to establish a common approach to governance across models and providers, making it easier for businesses to trace the actions of AI models and agents.
  • Most surprising point: More than 40 organizations are working together on AI transparency because businesses want to know not only what AI did, but also how to investigate when something goes wrong.
  • Taiwan perspective: As Taiwanese businesses adopt AI tools from different providers, they may need a management approach that makes those tools understandable and their actions traceable. Whether a common approach to governance can reduce the burden of managing multiple systems will depend on how it's put into practice.
  • Discussion points:
    • Which AI agent actions should businesses record to help with later investigations?
    • How can a common governance approach accommodate tools from different providers and the needs of different businesses?

6. Google Cloud Publishes AI Training Performance Results for TPUs and GPUs

  • Source: Google (Original article)
  • Summary: On November 16, 2026, Google Cloud published its MLPerf Training v6.1 results, including a run in which 8,192 Ironwood TPUs reached the target quality for a pretraining benchmark in 1.70 minutes. That time covers training from a specified reference checkpoint to the benchmark target, not training a complete model from scratch.
  • Most surprising point: The 1.70-minute result is eye-catching, but the clock starts at a specified checkpoint, not at the beginning of model training.
  • Taiwan perspective: As data center and hardware companies in Taiwan track AI computing performance, they also need to understand the starting points and targets of benchmarks. Just like comparing race times, you first need to know which part of the course each runner covered.
  • Discussion points:
    • When reviewing AI training benchmark results, what test conditions should you look at besides time?
    • What considerations might arise when designing a data center where large numbers of compute chips work together?

Closing

Today we covered everything from OpenAI helping AMD design chips to responsibility for AI agents using bank accounts. We also saw AI tools making their way into creative work, personal computers, and enterprise management. Technology is moving fast, and it's just as important to take time to understand what it does and who's responsible when things go wrong. This is "Mark's Tech Insights." I'm Mu Yan, and I'll see you next time!

Author

Mark Ku

10 年以上的軟體工程師,做過北美電商與 AI SaaS 訂閱收費系統。Read More

Found this useful?

The author's free tools, daily podcasts and newsletter are all here.

Mark Ku · This article is licensed under CC BY 4.0. Credit the author and link back to the original when reusing it.

Comments

Subscribe to Newsletter

Subscribe to get new posts delivered instantly — never miss a tech share.

By submitting, you agree to receive emails. You can anytime.

Popular Posts

View all
Mark Ku
··633

Oracle Cloud Always Free Tier: Linux Host and Static IP for a $0 Cloud Solution

Oracle Cloud Always Free Tier: Linux Host and Static IP for a $0 Cloud Solution
Mark Ku
··464

Say Goodbye to Postman's Fee Trap! A Hands-on Guide to Bruno, the Open-Source Git-Native API Testing Powerhouse.

Say Goodbye to Postman's Fee Trap! A Hands-on Guide to Bruno, the Open-Source Git-Native API Testing Powerhouse.
Mark Ku
··249

A Free, Open-Source, Notion-like Knowledge Base — A Complete Guide to Deploying and Backing Up Outline Wiki

A Free, Open-Source, Notion-like Knowledge Base — A Complete Guide to Deploying and Backing Up Outline Wiki
Mark Ku
··205

Building an Efficient API Management Platform: Deploying Kong Gateway from Scratch - Part 1

Building an Efficient API Management Platform: Deploying Kong Gateway from Scratch - Part 1
Mark Ku
··202

Setting Up Samba on Ubuntu to Share Folders with Windows 11

Setting Up Samba on Ubuntu to Share Folders with Windows 11
Mark Ku
··198

Training Your Own AI Voice: Hardware Requirements, Open-Source Model Comparison, and LoRA Fine-Tuning

Training Your Own AI Voice: Hardware Requirements, Open-Source Model Comparison, and LoRA Fine-Tuning