Latent Space, World Models & What AI Actually Knows

AI researcher and computer scientist based in Hong Kong. Founder of IRAI (irai.co). Specializes in latent space research and world models inside large language models. 12+ years in AI, with a background in reinforcement learning, neural networks, and model architecture. Started doing AI in high school, long before the GPT era.

Chapters

1. 0:00Intro — how the hosts met William
2. 3:22William's background — 12 years in AI before GPT
3. 5:20Sora & AI video — is the film industry over?
4. 6:52Curation, fake news & the attention economy
5. 9:25Latent space & world models — what's inside the black box
6. 11:50The labeled data problem — why AI was stuck for decades
7. 15:20AlphaGo & reinforcement learning — why games mattered
8. 18:20Knowledge vs intelligence — are they the same thing?
9. 20:00LoRA & fine-tuning — great for tone, bad for knowledge
10. 22:40MRI for neural nets — William's latent space research
11. 23:40Sparse vs dense — mixture of experts & Kimi explained
12. 26:40Qwen 3.5 — how Alibaba made every layer count
13. 29:10Multimodal models — why text alone won't reach AGI
14. 33:00Pokemon Go, crows & unconventional training data
15. 38:00Yann LeCun's AMI Labs — $1.5B and a head of research at HKUST
16. 39:00MiniMax — the world's most-used AI, IPO'd in Hong Kong
17. 40:00Hong Kong's hidden AI talent & branding problem
18. 42:50Hong Kong as the original cyberpunk city — Ghost in the Shell
19. 46:00Why AI needs humanities people, not just engineers
20. 49:00DeepSeek OCR — reading text as images
21. 52:00Chinese medicine as the original big data
22. 55:00Was GPT 3.5 actually better? Hallucinations & RLHF
23. 58:40Chopsticks, constraint & what the public actually needs
24. 60:00IRAI's finance data pipeline — a laptop-sized model
25. 62:30Outro — subscribe & visit hongkongaipodcast.com

About This Episode

William has been doing AI since high school — 12 years before GPT made it mainstream. We talk about what's really inside these models, the difference between knowledge and intelligence, latent space research, and why the labeled data problem shaped everything.

Related Articles