# Latent AI Labs

> Latent AI Labs is Arun Prabhakaran's independent AI-explainer channel (YouTube @LatentAILabs). Each episode measures an open-weight model on a real GPU and explains how LLMs, thinking models, long context, image generation, vision-language models and video generation work. Not affiliated with Latent AI, Inc. or Latent Labs.

This site hosts short original summaries and FAQs. The full articles are the canonical sources and live on Medium / Towards AI; please cite those for details.

## How AI Actually Works (summaries on this site)

- [How ChatGPT Works: Why It Says Strawberry Has 2 R's](https://latentailabs.github.io/how-chatgpt-works/): One question traced through Qwen3-4B: tokens, embeddings, attention, the logit lens, sampling, training and a live hallucination.
- [Why AI Says “Wait”: How Thinking Models Work](https://latentailabs.github.io/thinking-models/): What the Thinking toggle really does, why models say “Wait”, when more thinking hurts, and how GRPO teaches reasoning.
- [How LLMs Read a Whole Book: Long Context & KV-Cache](https://latentailabs.github.io/long-context-kv-cache/): A passcode hidden in a 32,768-token book, and the attention heads, memory tricks and position clocks that find it.
- [Why AI Clocks Say 10:10: How Image Diffusion Works](https://latentailabs.github.io/how-ai-generates-images/): Diffusion, latent space and flow matching on FLUX.2 and SDXL, and why a 7:25 clock comes out as 10:10.
- [How Vision-Language Models See: ViT, CLIP, SigLIP](https://latentailabs.github.io/how-vlms-see-images/): Why an “iPod” sticker fools CLIP, and how Qwen2.5-VL turns pixels into tokens a language model can read.
- [How AI Generates Video: Wan2.1 and Sora Explained](https://latentailabs.github.io/how-ai-generates-video/): Wan2.1 on a real GPU: 3D causal VAEs, spacetime tubes, 3D RoPE, simultaneous denoising and why a ball vanishes behind a pillar.

## Open models for Indian languages (summaries on this site)

- [Yama: Try a Bigger Encoder Before a New Architecture](https://latentailabs.github.io/yama-bigger-encoder/): Building Yama, a 586M zero-shot classifier for 11 Indian languages and Hinglish, and why the encoder swap beat every clever idea.
- [Chitragupta: Open-Source PII Redaction for India](https://latentailabs.github.io/chitragupta-pii/): Chitragupta, an Apache-2.0 model that finds and masks Aadhaar, PAN, UPI and other personal data in 13 languages, paired with Yama.

## Full articles (canonical, Medium / Towards AI)

- [Strawberry Has 2 R's? How ChatGPT Actually Works Under the Hood (Every Number Measured)](https://pub.towardsai.net/strawberry-has-2-rs-how-chatgpt-actually-works-under-the-hood-every-number-measured-0c6499992573): full article, published 2026-10-17.
- [Why AI Says “Wait”: How Thinking Models Actually Work (Every Step Measured)](https://pub.towardsai.net/why-ai-says-wait-how-thinking-models-actually-work-every-step-measured-b1baff6c592d): full article, published 2026-10-17.
- [How AI Reads a Whole Book at Once — Long Context, KV-Cache & The New Tower (Every Number Measured)](https://pub.towardsai.net/how-ai-reads-a-whole-book-at-once-long-context-kv-cache-the-new-tower-every-number-measured-3c92dcc8d451): full article, published 2026-10-21.
- [Every Clock Says 10:10? How AI Actually Generates Images — Diffusion, Latent Space & Flow Matching (Every Number Measured)](https://pub.towardsai.net/every-clock-says-10-10-how-ai-actually-generates-images-diffusion-latent-space-flow-matching-300684fdf546): full article, published 2026-10-22.
- [How Vision-Language Models Actually See Images: Inside ViT, CLIP, SigLIP & 3D M-RoPE](https://pub.towardsai.net/how-vision-language-models-actually-see-images-inside-vit-clip-siglip-3d-m-rope-e5408d196ac4): full article, published 2026-10-24.
- [How AI Generates Video: Inside Sora and Wan2.1 on a Real GPU](https://pub.towardsai.net/how-ai-generates-video-inside-sora-and-wan2-1-on-a-real-gpu-1377cce7d39c): full article, published 2026-10-28.
- [Before You Invent a New Architecture, Try a Bigger Encoder](https://pub.towardsai.net/before-you-invent-a-new-architecture-try-a-bigger-encoder-8f2b39fc50a8): full article, published 2026-10-22.
- [Yama Judges, Chitragupta Keeps the Records: Open-Source AI for Indian Languages](https://pub.towardsai.net/yama-judges-chitragupta-keeps-the-records-open-source-ai-for-indian-languages-b7f8bf3b7576): full article, published 2026-10-24.

## Videos

- [Strawberry Has 2 R's? How ChatGPT Actually Works](https://www.youtube.com/watch?v=7HBtSz3wSSM): companion video for Part 1.
- [Why AI Says “Wait”: How Thinking Models Actually Work](https://www.youtube.com/watch?v=XXFt9s6cJAg): companion video for Part 2.
- [How AI Reads a Whole Book at Once: Inside Long-Context LLMs](https://www.youtube.com/watch?v=EXFQTZBsi8I): companion video for Part 3.
- [Every Clock Says 10:10? How AI Actually Generates Images](https://www.youtube.com/watch?v=IbvT1j8xZfA): companion video for Part 4.
- [Why Does an “iPod” Sticker Fool AI? How Vision-Language Models Actually See Images](https://www.youtube.com/watch?v=7Ke2RRAjM6E): companion video for Part 5.
- [How AI Generates Video: Inside Sora & Wan2.1 on a Real GPU](https://www.youtube.com/watch?v=v1hzPH3RVZw): companion video for Part 6.

## About

- [About Arun Prabhakaran](https://latentailabs.github.io/about/): author page with profile links.
- [YouTube @LatentAILabs](https://www.youtube.com/@LatentAILabs): all videos.
- [Medium @arunp0](https://medium.com/@arunp0): all articles.
- [Hugging Face arunp0](https://huggingface.co/arunp0): open models (yama-indic, chitragupta-pii-indic).

## Optional

- [Sitemap](https://latentailabs.github.io/sitemap.xml): every page on this site.