Today is the last day for me in netease, I don’t much as much feeling as that in baidu, on the one hand, I spend much more time in baidu, almose five days, one the other hand, that’s my first job, ...

copilot-typed github copilot, cursor, codex-GPT, winsurf, trae are representative products of copilot-typed tools, what they can do is to help users type more code automatically when necessary, or ...

Gemini-cli Gemini-cli is command line tool supported by Gemini-2.5-pro model, it’s a similar product with Claude-code by Anthropic but free, what’s more, Google open source all code of Gemini-cli ...

Personal experience care the changing of loss/reward and test dataste performance, ensure they change with same trend, otherwise, reward hacking / invalid loss function appear adjust learning-...

How much do LLM memorize? key definition unintended memorization: memorize a specific dataset generalization (intended memorization): contains about the true data-generation pr...

ALL thing are certain in traditional computer technology, some programer say that there are beauty of certainty in traditional computer technology compared with current LLMs. For the influence or ...

Main idea Key point it to understand the below pictures Iteration steps for each input, generator G outputs for each output, calculate logits_prob for each token in current, old, referenc...

LLM Application The purpose of Appcalition is to try our best to meed the demand of users. Depth of technology is not very important. There are many unexplainable points in the field of LLM, as a...

LLM Inference It’s a great milestone when I finish a mature and iterable project, even it’s still in the first stage. There’re plenty AI infra repositoreis and projects in github, but most of the...

This is the first week to work as a LLM inference engineer, the real work content is what I want in the past. I don’t have much experience in this field, But it’s still acceptable. To do what I l...