I came across this article from Benedict’s Newsletter arguing that AI would replace knowledge workers very soon: https://shumer.dev/something-big-is-happening. I rarely spend time reading through this kind of article because I tend to see them as essentially promoting anxiety for some kind of self-promotion. However, Benedict says in his newsletter, “You should read it, not to […]

Vibe coding/engineering cultivates high-level thinking and a tendency to resist focusing on concrete details. It favors task completion over internalized understanding. The sense of FOMO motivates people to think about what’s next instead of asking why. With the purchased computing power, tokens, or GPUs, we have to keep the machine busy. We need to plan […]

Some random (and mildly fun) thoughts while helping with a recent conference: Recent major AI conferences (NeurIPS, ICLR, CVPR) have all suffered from the issues of overwhelming submission volumes, and the reviewer pool inevitably expanded. The Organizers have then found themselves battling against LLM-generated reviews. ACs were asked to identify such irresponsible reviewers, authors were […]

Design patterns emerge from practice. I was semi-vibe-coding a specific planner for myself and noticed the agent had trouble responding to my message while being asked to output a plan at the same time. Splitting it into two agents readily improved quality. You can even use different models for each task to save cost and […]

LLMs evolve so fast. I feel it's worth recording my current thoughts on how this technology changes my day to day experience in writing code.

There is a lot happening in tech, and AI is apparently replacing developers in batches, according to all kinds of "authoritatives". The reason the singularity is behind schedule, and we still have jobs, must because AI is using Emacs. LOL.

My 2nd grader didn't do well in his weekly English spelling quiz recently. Out of 12 words, he got about five of them right. I was the one to blame, of course, for some very legitimate reason :p

💾

Surrendering to FOMO on important AI "breakthroughs", I subscribed to Arxiv mailing list. Like all of the subscribers, I can't read all the submissions in the daily email. To be completely honest, I simply archive the email from my inbox everyday without even opening the email at all. There is a slight chance that I try to browse through titles and abstracts one by one. But I wouldn't be able to read more than 10 abstracts. There are simply too many to read through.

Spent some time on Saturday to play with the Gradio widgets and added utilities to quickly generate masks for inpainting.

I am having fun playing with all those image generation models and think it would be more interesting if we can make it into a 3D photo.

OpenAI recently updated its ChatGPT API with this function call capability. In short, it lets the user to tell the GPT model that there are one or more available function calls that ChatGPT can leverage. My first reaction to this news is to re-evaluate how GPT works on the sequence addition question.

几年前看Andrej Karpathy写的RNN教程的时候，就有些工作在用RNN做代码生成。我还记得刚看到这种工作的时候觉得这也太离谱了。果然，作为普通人，不出意外地又一次错误判断了事物的发展。才几年的时间，大家的普遍认知已经转变成，“代码生成是对于LLM来说相对比较简单的任务” ¯\_(ツ)_/¯

This demo shows how to run a pre-trained neural networks in web-browser. The user would first download the pre-trained style transfer model to local by opening up the webpage. Then everything will get processed locally without accessing any remote resource. The user can then open-up a picture from their hard drive and click "run" to let the style transferring neural networks to do its job.