Julia Turc

An educational video that traces the evolution of Transformer architecture from machine translation to the backbone of modern AI, explaining how attention replaced recurrence, how GPT's autoregressive training differs from diffusion-based generation, and how BERT's masked language modeling inspired the LLaDA diffusion LLM. It includes a step-by-step walkthrough of LLaDA's masked diffusion process for text generation.

A technical breakdown of Flow Matching, covering time-variant probability densities, the continuity equation, and how conditional velocity fields are used to train continuous normalizing flows. It connects the topic to concurrent research (Rectified Flow, Stochastic Interpolants) and addresses subtleties like optimal transport paths and conflicting velocity labels during training.

An educational video breaking down why GPUs vastly underperform their theoretical token throughput during LLM inference. It uses the roofline model to explain memory bandwidth bottlenecks, KV caching, speculative decoding, and how diffusion-based LLMs shift the workload from memory-bound to compute-bound.

A video walkthrough of methods researchers use to speed up diffusion-based language models, covering self-distillation through time, curriculum learning, confidence-based unmasking, guided diffusion (FlashDLM), approximate KV caching (dLLM-Cache, dKV-Cache), and block diffusion. It links these techniques to specific papers and models like LLaDA, Seed Diffusion, and Mercury, framing them against the standard autoregressive LLM paradigm.

An educational video walkthrough explaining how diffusion-based language models (like Gemini Diffusion, Mercury Coder, and Seed Diffusion) generate text by refining a full draft in parallel rather than predicting tokens sequentially, offering roughly 10x faster inference than autoregressive models like GPT or Claude. It works through the D3PM and LLaDA papers as case studies, covering Markov chain formulations and diffusion in embedding vs. token space.

An explainer video examining full duplex voice AI architecture — where a model listens and speaks simultaneously rather than turn-by-turn — using Kyutai's open-sourced Moshi as the reference implementation. It compares cascade voice pipelines (ASR + LLM + TTS) against end-to-end speech-to-speech models and includes commentary from Kyutai founder Neil Zeghidour on tool calling, RAG, and why full duplex systems aren't yet widespread.

An educational video breakdown of Mistral's Voxtral Realtime and Voxtral TTS models as case studies in modern real-time voice AI, covering streaming speech-to-text, the Voxtral audio codec, and text-to-speech architecture. It also traces the history of speech models (Whisper, Whisper Streaming, WaveNet) and explains core audio tokenization concepts like VQ, RVQ, FSQ, and semantic vs acoustic tokens.

An educational video tracing the architectural evolution from the Vision Transformer (ViT) to the Diffusion Transformer (DiT) to the Multimodal Diffusion Transformer (MMDiT), explaining how Transformers displaced CNNs as the standard architecture for vision tasks. It covers key mechanisms like adaLN, adaLN-Zero, cross-attention, and FiLM-style conditioning, referencing the original ViT, DiT, MMDiT, and FiLM papers.

Jonathan Ross is the founder of Groq, a hardware chip specifically designed for LLM inference, which entered a $20 billion strategic agreement with NVIDIA. Topics covered: • The "success disaster" at Google that led to the TPU and eventually Groq • LPU vs. GPU: Pareto curves, cost-per-token, and when each wins • Static scheduling • Mixture-of-experts models • Auto-regressive vs. diffusion models • How Groq and NVIDIA's Vera Rubin work together at inference time • Jevons Paradox: why cheaper AI w