Posts

I developed vocaBull which is a flashcard web app a few years ago, and I use it almost every weekday. Introducing vocaBull github repo vocaBull (vb, for short) is deployed on Pythonanywhere.com, but the free plan for pythonanywhere gets slow frequently, forcing me to wait for a few seconds before I see a new flashcard. As vb is Flask and Jinja2-based, it needs to render an HTML page for every flashcard at the server. This has made it frustrating to use vb from Pythonanywhere. ...

Probably not. But I have to say that I hate subscriptions because they are monthly cost that are not predictable. We start subscribing services because we think they are priced reasonably. But, subscription fees are completely up to them. They can change fees anytime they want. It’s pretty common that some services are first cheap and later change pricing. Look at Spotify and Netflix. Look at Apple. Look at Google. ...

(I’ve forgotten to publish this…7/19/2026) Currently, I’m working on the 3rd full-stack project that involves Python+SQLAlchemy+Pydantic+FastAPI backend and JavaScript+Vue+Pinia+TailwindCSS+DaisyUI frontend. Before these projects, I knew Python+SQLAlchemy+Flask+Jinja2+BootStrap, ie, server-side only. I didn’t know that context switching is the hard part for full-stack. If I notice while working on a Vue file that I need to add a column to a database table, I need to change database ORM → Pydantic class → FastAPI → Pinia → Vue. ...

I like to talk with AIs and find anomalies in their answers. I think I have a basic understanding of how LLMs work, and so I want to know how those anomalies map to LLM architecture. How GenAI works - Transformer internals Recently, DeepSeek tends to answer in Chinese more often than before. More than half of the time, it even thinks in Chinese (we can observe its thinking/reasoning). As DeepSeek is developed and trained mostly in Chinese, I think this is understandable. ...

I know that AI doesn’t know much about itself, but I can’t help asking an AI about its new feature. Today, I asked DeepSeek if it has a new memory feature, and it confidently hallucinated that it has a new DeepSeek++ Chrome plugin. As it turned out, DeepSeek++ is actually a popular third party project. What’s hilarious is that humans never mix up a 3rd party feature as their own 1st party one. ...

Some AIs have memory. I’m sometimes surprised how much ChatGPT remembers about me. But, the AI memory is a trick. It’s just a context injection at the start of a new chat session. The real memory for AI (transformer) should be ingrained to matrices W1 and W2 in the MLP layer. How GenAI works - Transformer internals What’s fascinating about the transformer mechanism to me is that the vast amount of world knowledge is kept in the W1 and W2 matrices. ...

When you talk with AI (transformer) such as ChatGPT or Gemini, they are pretty much like a human. They are supportive and they answer all your questions. Sometimes I feel as if I’m having a real conversation with a real person. But at the same time, I know that what I’m feeling is an illusion. What AI is actually doing is generating most likely tokens based on the chat session and what they learned during training. ...

Your cognitive load is the bottleneck This past couple of months, I’ve been working on projects. AI has drastically accelerated my learning and development. Questions to AI can start shallow (eg, “what is X?”) and go as deep as I want. AI more or less understands what I want to make, and urges me to write code step-by-step. I instruct AI not to give code and to let me write code myself first. And more often than not, the code I’m expected to write includes new items for me. ...

Today, I was discussing with DeepSeek why AI is not a super wise brain even though it has vast amount of knowledge available on the entire Internet. It’s answer mostly made sense; that it’s just a probability prediction machine and it doesn’t learn from experience. Looking at the attention mechanism which is the core of DeepSeek, ChatGPT, Gemini, etc., I understand what it means. It only cares about context/patterns. It forgets everything in a new chat without “memory” feature. It even forgets its training data unless repeatedly appeared, though I am often marveled at how AI can retrieve niche info from its knowledge. ...

I’ve been learning JavaScript and Vue (and re-learning Python) with Claude as my coding partner. I will share some findings and tricks when learning from AIs. Note: I’m (sort of) intermediate for Python and a novice for JS/Vue. AIs are trained to help, not to teach If you use general chat AIs such as ChatGPT, Gemini, Claude or Grok, they tend to give your full working source files from the beginning. This is not good for your learning. ...