Julia Turc is an AI researcher and educator who co-founded Storia AI. Her work spans compact language models, tools for understanding code, and explanations of the mechanisms behind modern AI systems.[1][2][4]
Start with diffusion.fyi →
X Career
Turc previously worked at Google Research and co-founded Storia AI. Y Combinator lists her as its CTO, although her current personal site focuses on AI education and does not confirm that title. She studied computer science at Oxford; the university lists her 2014 project on recurrent neural networks for machine translation among its distinguished MSc projects. She now publishes explanations of AI research alongside her work on software. Her papers also appear under the name Iulia Turc.[1][2][3][8][4]
Contributions
Making smaller language models useful
Turc is first-listed author of Well-Read Students Learn Better, a 2019 study of compact language models. The team tested pretraining a small model before transferring task knowledge from a larger model through distillation. Pretraining remained useful even for small architectures, and combining it with distillation improved the reported results. The study matters for deployments where the largest model is too expensive or slow: it examines how to retain useful capabilities within a smaller model.[4]
Explaining efficient NLP
She co-authored the 2020 High Performance Natural Language Processing tutorial. It examines ways to make models more practical to run, including distillation, quantization, pruning and more efficient architectures. These techniques reduce different costs, so a model's parameter count alone does not describe its deployment tradeoffs. The tutorial connects model design to hardware use and gives readers a research-oriented foundation for evaluating efficiency claims.[5]
Understanding a codebase with Storia
Storia describes its goal as helping developers understand code and the context around it. Its Sage project lets a reader ask questions about an existing repository rather than manually searching every file. Turc is a co-founder of the company behind that work; this does not establish sole authorship of its software. The public Sage repository was archived in February 2025, so it is useful as a reference implementation rather than evidence of an actively maintained service.[2][6]
Teaching diffusion through interaction
Turc built diffusion.fyi to make diffusion models easier to explore. Its lessons and playgrounds let readers examine how noise is added and removed, how training works, and how different sampling choices change the process. The site complements her videos with interactive examples and links to foundational papers. A newcomer can start with the fundamentals before reading the equations in a research paper.[3][7]
Tools, writing & talks
Apps
Start here
diffusion.fyidiffusion.fyiInteractive lessons and playgrounds for understanding noise, model training and reverse diffusion.
Begin with the Fundamentals tutorial and change the playground controls to see how noise and reverse diffusion work. The interactive examples provide a foundation for her more technical videos and paper explanations.
Tools
github.comWriting, research & talks
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.
Catalog workOriginal sourceAn 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.
Catalog workOriginal source- Voxtral Realtime and Voxtral TTS Explainedyoutube.com
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.
Catalog workOriginal source Groq founder Jonathan Ross on the success disaster at Google that produced the TPU, how LPUs and GPUs trade off on cost per token, static scheduling, and the $20 billion strategic agreement with NVIDIA.
Catalog workOriginal sourceA 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.
Catalog workOriginal sourceA 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.
Catalog workOriginal source- Vision Transformer, Diffusion Transformer, and Multimodal Diffusion Transformer Explainedyoutube.com
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.
Catalog workOriginal source 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.
Catalog workOriginal sourceAn 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.
Catalog workOriginal source- Well-Read Students Learn Betterarxiv.org
Research on combining pretraining and distillation to improve compact language models.
Co-authored paper[4] - High Performance Natural Language Processingaclanthology.org
A tutorial covering model compression, efficient architectures and practical hardware use.
Co-authored tutorial[5]