About
Arun Prabhakaran
I make Latent AI Labs, a series of AI explainers built on measurement rather than metaphor.
For each episode I load an open-weight model, such as Qwen3-4B, FLUX.2, Qwen2.5-VL or Wan2.1, onto an NVIDIA RTX PRO 6000 Blackwell GPU, hook into its layers and record what actually happens: token IDs, attention weights, memory use, timings. The results become a long-form animated video on YouTube (@LatentAILabs) and a full written article on Medium, published in Towards AI. Claims in the articles cite the original research papers.
I also build small open models for Indian languages and publish them under Apache-2.0 on Hugging Face:
- Yama: a zero-shot text classifier for 11 Indian languages, Hinglish and English. Read the summary.
- Chitragupta: personal-data detection and redaction for Indian text. Read the summary.
Elsewhere
- Medium · @arunp0
- YouTube · @LatentAILabs
- Hugging Face · arunp0
- GitHub · arunp0
- GitHub · latentailabs
- LinkedIn · arunp0
Start reading
New here? Begin with How ChatGPT Works: Why It Says Strawberry Has 2 R's and follow the series in order.