I am a technologist who builds complex systems and products across both mature companies and startups. I am currently the Founder & CEO of FuSearch and a co-founder of NumberTheoryRobot.
Zhanliang Liu 👋🏻
Founder & CEO @ FuSearch · Co-founder @ NumberTheoryRobot
Field notes from the early days of the intelligence revolution.
Start here
- ChatGPT and the Dawn of a New EraOn November 30, 2022, OpenAI released ChatGPT. In hindsight, that day may mark the beginning of the Fourth Industrial Revolution.
- What Comes After AGI? DeepMind's “From AGI to ASI”A short introduction to a new DeepMind report by Shane Legg and colleagues, which asks how AI might keep developing once it reaches human-level intelligence.
Notes
See all notesPaper notes
See all paper notes- Two Brains for a Self-Evolving AgentSiru Ouyang et al. · Google Cloud AI Research · UIUC · MIT · 2026
Latest posts
See all posts- Serving Qwen3.8-27B on One DGX SparkHow DFlash2, native rank-1 LoRA, cache isolation, and strict validation turned one Qwen3.8-27B process into two independently addressable model modes—without slowing the clean base path.
- A Brief History of Artificial IntelligenceA timeline of AI from the 1956 Dartmouth workshop to ChatGPT: symbolic systems, the rises and falls of neural networks, deep learning, and the LLM era.
- UIE - Universal Information ExtractionUIE: a unified text-to-structure framework that models diverse information-extraction tasks and generates targeted structures on demand.
- Vector Search EngineVector search engines or vector databases are a core piece of infrastracture that fuels every big deep learning deployment in industry.
- Language Model Is All You Need — Explores NLU as MRC QAAmazon Alexa AI's paper "Language Model Is All You Need" reframes natural-language understanding as a question-answering problem.
Projects
See all projects- Qwen3.8 on DGX Spark LabA reproducible Qwen3.8-27B serving lab for one NVIDIA DGX Spark: pinned ARM64 vLLM, DFlash2, native rank-1 LoRA, API gates, cache-isolation validation, and sanitized benchmarks.
- pullword: Unsupervised Word DiscoveryDiscovering words in unsegmented Chinese text with no training corpus and no prespecified vocabulary — for mining online and domain-specific text.
- FMR: functional meaning representationFMR is a strongly typed, open-domain formal language for representing meaning, paired with a CFG-based parser that maps natural language onto it.
- gocc: Golang version OpenCC 繁簡轉換A Go port of OpenCC for converting between Traditional and Simplified Chinese, including regional idioms and phrase-level substitution.
Recent talks
See all talks- 人工智能的发展简史
- 从AI到IA，知识图谱信息抽取在智能审单业务中的应用实践AI开发者大会2020
- 自然语言处理与行业应用世界人工智能融合发展大会
Let's Connect
If you want to get in touch about something or just to say hi, reach out on social media or send me an email.
- github /
- linkedin /
- twitter /
- weibo /
- liang@zliu.org