One of the more surprising results in recent language-model research is that you can get a model to reason without ever showing it how. You reward the right answer, and the reasoning grows in on it...

Train a self-supervised vision model long enough and something strange happens. On the headline benchmark it keeps improving. On another, quieter benchmark it starts getting worse — during the same...

Richard Hamming had a habit, recounted in Vivek’s essay How to Be Good at Research, that made him unpopular at lunch. He would ask colleagues what the important problems in their field were — and t...

Here is a quietly instructive failure. For years, methods that reconstruct a 3D face from a single photo got the geometry roughly right — the shape of the head, the pose, the rough placement of the...

Suppose you have a fixed pile of compute and you want the best language model it can buy. Do you spend it on a bigger model, or on more training data? For a couple of years the field had the wrong ...

How do you teach a vision model to understand images without labels? One of the cleanest answers came from a 2022 paper, Masked Autoencoders Are Scalable Vision Learners (He et al.). The recipe is ...

There is a line from Vivek’s essay How to Be Good at Research that is worth taping above a monitor: “a descending loss curve is not analysis, it’s reassurance.” The curve going down tells you the ...

In 2019 Rich Sutton wrote an essay of about a thousand words called The Bitter Lesson. Its claim is simple and a little deflating: across 70 years of AI research, the methods that won were the ones...

从平时coding中收集一些常用的代码片段，记录在这里，方便以后复用。 1. 项目代码入口 import argparse def parse_args(): parser = argparse.ArgumentParser() parser.add_argument('--model', type=str, default='yolov5s') parser.a...

文件传输 一台机器上的复制 $ cp -i file1 file2 将文档 file1复制成file2，复制后名称被改file2 $ cp -i file1 dir1 将文档 file1复制到dir1目录下，复制后名称仍未file1 $ cp -r dir1 dir2 将目录dir1下的所有文件复制到dir2目录下，复制结果目录被改名为dir2 ...