Hui Kang Tong

AGI is when you have nothing to be angry about in all AI responses

What is the sum of base pay of all software engineers around the world

Maybe OSS coding tools should calculate the exact FLOPs and VRAM-seconds that you are using for the OSS models

Household appliances should be made with CLI support

All the non-official claude code docs showing up on search will someday cause security issues

Plan mode designers overestimate how much I read

Soon we will have AI complaining to us

@yujia_bao @thinkymachines @tinkerapi Thanks for addressing the bug I posted - github.com/thinking-machi…

There's been a lot of excitement around auto-research, but one underappreciated bottleneck: coding agents struggle to run LLM training jobs at scale. A small infrastructure mistake can have major consequences on the output. I recently joined @thinkymachines, and @tinkerapi solves exactly this. It standardizes the training process — training a 1T parameter model is as simple as training a 4B one. That makes auto-research with coding agents like Claude Code actually viable.

As models get more compute efficient quantization will be as evil as tokenization

you need the slop history to fix slop code

We previously thought about ...

You can start reference-free, and then be reference-based

TTS should be CLI

All the algo coding and system design interviews but the pages are still slow

You don't need long context for long context

Can you trust something is maintained after your colleague ships and after your colleague leaves

It also seems that if you get the agent a score, it will keep trying to increase the score without appreciation of which part of the score is meaningful