Ask about Kefei
written answers · site notes
Hi, I'm the assistant on Kefei Yan's site. Ask about his projects, research, or what he's looking for next.
About Local AI
Checking whether this browser can run it…
- What it adds
- Local AI turns the site's own notes into an answer instead of just listing them. It can compare two projects on the same criteria, explain a paper in plain language, summarise, use bullet points, translate between English and Chinese, and relate what is recorded here to a role you have in mind. With it off, the assistant still returns the answers Kefei wrote and searches the site's notes, but it will not rewrite or combine them.
- How it works
- A question is matched against this site's notes with BM25 retrieval. The matching notes are put in front of Qwen3 1.7B, which runs through WebLLM and WebGPU in a worker inside this browser tab. Every reply shows which notes it came from, and the generated text is checked against those notes for links, numbers, year ranges and invented pronouns before it is shown.
- The download
- The model is about 985 MB, fetched once from Hugging Face and kept in the browser's cache. It stays until you remove it or the browser clears its storage, and a later visit may download it again.
- Where what you type goes
- Questions and generated text are processed on this device and are not sent to any inference service. Downloading the model and the page's scripts still connects to their hosts.
- What it needs
- A desktop browser with working WebGPU and a real GPU, roughly 2.3 GB of free video memory, and data saver switched off. Phones and tablets are not offered this version. The page checks before offering it and says why when it cannot.
- What it will not do
- Qwen3 1.7B is a small model and can still be wrong or leave things out. The link and number check is not fact checking, and it does not read prose, so a plausible wrong sentence can pass. Anything that matters is worth confirming against the full profile.