After one month apart, we sit down to talk about how life—and our relationship—has changed.

From returning to Taiwan and taking my parents on a trip to navigating round after round of job interviews, this month has been full. On the other side, long distance has meant doing chores, eating alone, and facing moments of loneliness.

In this episode, we look back on the past month and talk about what comes next: staying connected, figuring out work, and possibly starting a new chapter in Taipei. Plus, a few playful moments and the kind of support only couples understand.

Powered by Firstory Hosting

In this episode, I talk through two AI model design ideas:

First, how GPT Live handles voice interaction: why older cascaded systems moved from speech to text to LLM output to text-to-speech, why audio-native models can respond faster, and why turn detection can still accidentally interrupt when silence, pauses, breathing, or background noise are misread.

Second, I explain mixture-of-experts models: how transformer blocks use attention and feed-forward networks, how MoE splits the feed-forward layer into experts, how a gate routes tokens to selected experts, and why this can reduce inference cost while letting different experts specialize.

Leave a comment and share your thoughts:

Comment Link

Powered by Firstory Hosting

In this first episode, we talk about two pieces of the hardware foundation behind modern AI: Taiwan's electronics industry and GPU memory offloading.

The first part explains what the electronics industry is, how it differs from the broader tech industry, and why semiconductors, server manufacturing, foundries, and PCB suppliers matter so much to global technology. We also discuss Taiwan's role in chip manufacturing, AI servers, and NVIDIA-related supply chains.

The second part explains GPU memory in plain language: CPU RAM, SSD storage, VRAM, HBM, and PCIe. From there, we discuss why large AI models often cannot fit entirely into GPU memory, and how offloading moves model weights between CPU RAM and GPU VRAM layer by layer during inference or training.

Leave a comment and share your thoughts:

https://open.firstory.me/user/cmrcw7e6804iz01uo38ln04mg/comments

Powered by Firstory Hosting