I want to control the model, not just use it.
ENGINEERING · ADAPTABLE · LIKES A HARD PROBLEM
- Background
- Beijing, high school. Research with a university lab group, an algorithm internship, and lyric generation models for a music company.
- Community
- Founded and run Alchemy, the school AI club. 10+ active members, ongoing tech sharing and collaboration.
- Research
- Model training, agent systems, interpretability, multimodal, 3D generation and representation analysis.
- Off the keyboard
- A native coastal marine tank. Have kept electric rays, catsharks, sea bream, filefish and grouper.
Research
Four directions: understand what happens inside a model, let it see images and 3D worlds, then run it with as little compute as possible.
LLM Interpretability
Using steering vectors to study how abstract and high-dimensional directions separate, and how to explain the concepts they encode.
Multimodal and Dynamic Feature Expansion
Dynamic Feature Expansion (DFE) on a Qwen2.5-VL-7B backbone, studying how input resolution affects recognition of 3D relief and topological detail.
3D Generation and World Models
Procedurally generated infinite environments on 3D Gaussian Splatting, comparing streaming video world models against persistent 3DGS, toward an open-world game.
Inference Optimization
Large-model inference on a single RTX PRO 6000: offload performance patches, quantization sensitivity and vLLM serving. GLM-5.3-Flash from 17.8 to 36.7 tok/s.
Recurrent Transformers
Trained a 51M sequence-recurrent Transformer-XL variant and a 110M Universal Transformer with ACT. Depth versus sequence recurrence, and whether external memory can move the compute-quality Pareto frontier.
Open Source
freetoken-ox-boost
GLM-5.3-Flash adaptation and offload performance patches for FreeToken v0.1.2. 17.8 → 36.7 tok/s on a single RTX PRO 6000.
endless-labyrinth
A procedurally generated infinite environment built on 3D Gaussian Splatting. Long-term goal: an open-world game.
recurrent-transformer-0.1b
A 110M-parameter Universal Transformer with ACT, released on Hugging Face.
Tree_Of_Thought
An implementation of, and experiments on, the Tree of Thoughts reasoning paradigm.
agent_team
A minimal multi-agent team aimed at getting the basics working.
MiniLLM-0.1B
A 0.1B-parameter language model trained from scratch.
A student club. Our own GPU.Free for classmates.
Alchemy is the student AI club at SDSZ. We run our own RTX PRO 6000. Compute is limited, so the inference API is internal and invite-only. We also build campus platforms and take on AI integration work.
INTERVIEW · 数字生命卡兹克
Two high-schoolers, free AI for their classmates
Two 16-year-olds plugged their own compute into a student club so more classmates could use AI for free.
Alchemy FreeAPI
A free GPU inference API. Compute is limited, so access is internal and invite-only. Serves custom-weight inference.