| { |
| "schema": 1, |
| "title": "Papers, as Audio", |
| "generated": "2026-10-14T16:23:58Z", |
| "repo": "arjun10g/papers-audio", |
| "revision": "fdb148d61b5f45c4be015ad9b51ca12aa0d56b79", |
| "base": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/fdb148d61b5f45c4be015ad9b51ca12aa0d56b79/", |
| "lectures": [ |
| { |
| "id": "splines", |
| "n": 1, |
| "title": "Splines: Concepts and Applications", |
| "series": "Statistical Learning", |
| "audio": "audio/splines.mp3", |
| "bytes": 5695244, |
| "duration": 1423.75, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1635070041078-e363dbe005cb?w=1920&q=80", |
| "published": "2025-08-26", |
| "sha256": "6d130a7fd8182894c3ad32a284bc60a77ce855cda4cf353ccf5f3c49c39e565f" |
| }, |
| { |
| "id": "intro-ml", |
| "n": 2, |
| "title": "An Introduction to Machine Learning", |
| "series": "Statistical Learning", |
| "audio": "audio/intro-ml.mp3", |
| "bytes": 5967884, |
| "duration": 1491.91, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1677442135703-1787eea5ce01?w=1920&q=80", |
| "published": "2025-08-27", |
| "sha256": "9efbb908f39cd57c37b86b95334816f983df875de2ee9c0df0dde9fe972c795f" |
| }, |
| { |
| "id": "nonlinearity", |
| "n": 3, |
| "title": "An Introduction to Non-Linearity", |
| "series": "Statistical Learning", |
| "audio": "audio/nonlinearity.mp3", |
| "bytes": 10247852, |
| "duration": 2561.9, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1545987796-200677ee1011?w=1920&q=80", |
| "published": "2025-08-28", |
| "sha256": "464f7eae3ef24e9f71e27ce76127736df503b2c10f79d94261d6124f7887e778" |
| }, |
| { |
| "id": "longitudinal", |
| "n": 4, |
| "title": "Longitudinal Modelling", |
| "series": "Statistical Learning", |
| "audio": "audio/longitudinal.mp3", |
| "bytes": 12232172, |
| "duration": 3057.98, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1551288049-bebda4e38f71?w=1920&q=80", |
| "published": "2025-08-31", |
| "sha256": "fd3321f169aac2aca8f281b9a9f7a1f5d8add8f2c35c7d5d385c195a77c1e1db" |
| }, |
| { |
| "id": "spline-interpretation", |
| "n": 5, |
| "title": "Spline Interpretation", |
| "series": "Statistical Learning", |
| "audio": "audio/spline-interpretation.mp3", |
| "bytes": 5822348, |
| "duration": 1455.53, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1509228468518-180dd4864904?w=1920&q=80", |
| "published": "2025-09-02", |
| "sha256": "002761e38dc8d0474819351677c8cc8c5edc268fbf10aa61dfce1469b20f3e32" |
| }, |
| { |
| "id": "trees", |
| "n": 6, |
| "title": "Trees", |
| "series": "Statistical Learning", |
| "audio": "audio/trees.mp3", |
| "bytes": 10599116, |
| "duration": 2649.72, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1502082553048-f009c37129b9?w=1920&q=80", |
| "published": "2025-09-04", |
| "sha256": "541e3b26dbccac143d2f3ff7ed92a31d83161005c68f7d3386a60b29f000c428" |
| }, |
| { |
| "id": "svm", |
| "n": 7, |
| "title": "Support Vector Machines", |
| "series": "Statistical Learning", |
| "audio": "audio/svm.mp3", |
| "bytes": 5926412, |
| "duration": 1481.54, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1555949963-aa79dcee981c?w=1920&q=80", |
| "published": "2025-09-17", |
| "sha256": "e59e0cee0b9301e0b19087e710335d9bd96fc5c8892d9371aa876d1ab6b2d85e" |
| }, |
| { |
| "id": "neural-networks", |
| "n": 8, |
| "title": "Neural Networks", |
| "series": "Statistical Learning", |
| "audio": "audio/neural-networks.mp3", |
| "bytes": 12943340, |
| "duration": 3235.78, |
| "doc": null, |
| "bg": "https://images.unsplash.com/photo-1620712943543-bcc4688e7485?w=1920&q=80", |
| "published": "2025-09-18", |
| "sha256": "61ad753a471fc50a7c6b0efcb254a50475b5bfbe5ce71f7cfe3b2f89f8cb87aa" |
| }, |
| { |
| "id": "rashomon", |
| "n": 9, |
| "title": "The Rashomon Effect", |
| "series": "Rashomon & Ensembles", |
| "audio": "audio/rashomon.mp3", |
| "bytes": 8536556, |
| "duration": 1067.04, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/rashomon.md", |
| "bg": "https://images.unsplash.com/photo-1478720568477-152d9b164e26?w=1920&q=80", |
| "published": "2026-03-31", |
| "sha256": "e8fda516cf88d88ac3556c6ab5bfe23dc338a206f6b042cb46b01799e3d532e5" |
| }, |
| { |
| "id": "rashomon-instability", |
| "n": 10, |
| "title": "Rashomon Analysis & Model Instability", |
| "series": "Rashomon & Ensembles", |
| "audio": "audio/rashomon-instability.mp3", |
| "bytes": 8578988, |
| "duration": 1072.34, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/rashomon-instability.md", |
| "bg": "https://images.unsplash.com/photo-1494232410401-ad00d5433cfa?w=1920&q=80", |
| "published": "2026-04-01", |
| "sha256": "1a34b934439c32384b383e99bfe6a67f90cb5457f64f7f1f942b93ffaf3745ad" |
| }, |
| { |
| "id": "boosting", |
| "n": 11, |
| "title": "Ensemble Learning & Gradient Boosting", |
| "series": "Rashomon & Ensembles", |
| "audio": "audio/boosting.mp3", |
| "bytes": 22136876, |
| "duration": 2767.08, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/boosting.md", |
| "bg": "https://images.unsplash.com/photo-1527474305487-b87b222841cc?w=1920&q=80", |
| "published": "2026-04-06", |
| "sha256": "398db81fda21a959f0f038d92a6e0f257af3d3f6f0afaf01bd63448ea5e298cd" |
| }, |
| { |
| "id": "decomposed-tree", |
| "n": 12, |
| "title": "Decomposed Random-Effects Tree", |
| "series": "Rashomon & Ensembles", |
| "audio": "audio/decomposed-tree.mp3", |
| "bytes": 17074604, |
| "duration": 2134.3, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/decomposed-tree.md", |
| "bg": "https://images.unsplash.com/photo-1518186285589-2f7649de83e0?w=1920&q=80", |
| "published": "2026-04-09", |
| "sha256": "5638289339a11d2c1e32c597aff8aeb936a1cb1f050eca3ff22c8440b9967471" |
| }, |
| { |
| "id": "transformers", |
| "n": 13, |
| "title": "Transformers & LLM Design Decisions", |
| "series": "Transformers & Embeddings", |
| "audio": "audio/transformers.mp3", |
| "bytes": 20098796, |
| "duration": 2512.32, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/transformers.md", |
| "bg": "https://images.unsplash.com/photo-1677442135703-1787eea5ce01?w=1920&q=80", |
| "published": "2026-04-17", |
| "sha256": "ff3b0bc7fe4aee4d3ac09417243c294fbcf90de908e3a7879278594369c40341" |
| }, |
| { |
| "id": "embedders", |
| "n": 14, |
| "title": "How Embedder Models Are Trained", |
| "series": "Transformers & Embeddings", |
| "audio": "audio/embedders.mp3", |
| "bytes": 9554924, |
| "duration": 1194.34, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/embedders.md", |
| "bg": "https://images.unsplash.com/photo-1639322537228-f710d846310a?w=1920&q=80", |
| "published": "2026-04-21", |
| "sha256": "e966867b3f5ef2137771b4f9c2f478eadfaee89b97102fd5fdc3c383dbc153ad" |
| }, |
| { |
| "id": "recurrent-depth", |
| "n": 15, |
| "title": "Thinking in the Dark: Recurrent Depth Models", |
| "series": "Recurrent Depth & Ouro", |
| "audio": "audio/recurrent-depth.mp3", |
| "bytes": 22617836, |
| "duration": 2827.2, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/recurrent-depth.md", |
| "bg": "https://images.unsplash.com/photo-1526374965328-7f61d4dc18c5?w=1920&q=80", |
| "published": "2026-06-25", |
| "sha256": "8e784872809b61d1bc1b43dfae83be96a938c3b497a7ce72e838695d5bb85024" |
| }, |
| { |
| "id": "hardware-ouro", |
| "n": 16, |
| "title": "Making Ouro Fast: Hardware & Kernel Optimization", |
| "series": "Recurrent Depth & Ouro", |
| "audio": "audio/hardware-ouro.mp3", |
| "bytes": 15037868, |
| "duration": 1879.7, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/hardware-ouro.md", |
| "bg": "https://images.unsplash.com/photo-1591405351990-4726e331f141?w=1920&q=80", |
| "published": "2026-07-03", |
| "sha256": "71b34e853828d442f4873df79f13151459edb46dd91cea397cb0f2661c8b163d" |
| }, |
| { |
| "id": "model-comparison", |
| "n": 17, |
| "title": "Four Models, One GPU: A Hardware Comparison", |
| "series": "Recurrent Depth & Ouro", |
| "audio": "audio/model-comparison.mp3", |
| "bytes": 8733740, |
| "duration": 1091.69, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/model-comparison.md", |
| "bg": "https://images.unsplash.com/photo-1518770660439-4636190af475?w=1920&q=80", |
| "published": "2026-07-06", |
| "sha256": "8029d1f76844cb783e41d531af0cf1b5cd41607e3fad45b7eacbbc985ca269d4" |
| }, |
| { |
| "id": "ouro-implementation", |
| "n": 18, |
| "title": "Ouro in Practice: Implementation Plan", |
| "series": "Recurrent Depth & Ouro", |
| "audio": "audio/ouro-implementation.mp3", |
| "bytes": 28273772, |
| "duration": 3534.19, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/ouro-implementation.md", |
| "bg": "https://images.unsplash.com/photo-1515879218367-8466d910aaa4?w=1920&q=80", |
| "published": "2026-07-07", |
| "sha256": "fa8aa9dda5214af0c0021d0deb2f43d15be3c1d5e66c12bc6045c8dd0b6cab75" |
| }, |
| { |
| "id": "inference-serving", |
| "n": 19, |
| "title": "The Model Is No Longer Just the Model: Serving in August 2026", |
| "series": "Inference & Serving", |
| "audio": "audio/inference-serving.mp3", |
| "bytes": 26284460, |
| "duration": 1642.75, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/inference-serving.md", |
| "bg": "https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1920&q=80", |
| "published": "2026-08-25", |
| "sha256": "5a1899680325c2aeb423c7ec23d77080c2c9a0c48a1aeea3039b86149d330f6e" |
| }, |
| { |
| "id": "throughput", |
| "n": 20, |
| "title": "Same Model, Twice as Fast: Optimizing for Throughput", |
| "series": "Inference & Serving", |
| "audio": "audio/throughput.mp3", |
| "bytes": 24726956, |
| "duration": 1545.41, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/throughput.md", |
| "bg": "https://images.unsplash.com/photo-1504639725590-34d0984388bd?w=1920&q=80", |
| "published": "2026-08-25", |
| "sha256": "d9be5224448dfdb67f253859ed56a8739b3ea07ca9645a12b86d9733eb152590" |
| }, |
| { |
| "id": "expert-prefetch", |
| "n": 21, |
| "title": "What We Learned About Predictive MoE Prefetching", |
| "series": "Inference & Serving", |
| "audio": "audio/expert-prefetch.mp3", |
| "bytes": 21785316, |
| "duration": 1815.41, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/expert-prefetch.md", |
| "bg": "https://images.unsplash.com/photo-1451187580459-43490279c0fa?w=1920&q=80", |
| "published": "2026-09-16", |
| "sha256": "e4c60b4eed6bf3ae52290d57447ed49e46367e30fb87723b278b8955f7c25243", |
| "chapters": [ |
| { |
| "title": "Introduction", |
| "start": 0 |
| }, |
| { |
| "title": "The question", |
| "start": 41.39 |
| }, |
| { |
| "title": "A short chronology", |
| "start": 123.95 |
| }, |
| { |
| "title": "Routing, and where the prediction error lives", |
| "start": 203.81 |
| }, |
| { |
| "title": "Native experts versus proposed candidates", |
| "start": 283.15 |
| }, |
| { |
| "title": "The timing pipeline", |
| "start": 352.65 |
| }, |
| { |
| "title": "The metrics, and why they are kept apart", |
| "start": 433.41 |
| }, |
| { |
| "title": "Predictor families", |
| "start": 512.68 |
| }, |
| { |
| "title": "Low-rank correction, step by step", |
| "start": 684.77 |
| }, |
| { |
| "title": "Model geometry and the selected configurations", |
| "start": 830.1 |
| }, |
| { |
| "title": "Offline results", |
| "start": 931.06 |
| }, |
| { |
| "title": "Hardware and transfer measurements", |
| "start": 1054.6 |
| }, |
| { |
| "title": "End-to-end results", |
| "start": 1186.63 |
| }, |
| { |
| "title": "Engineering repairs and reproducibility", |
| "start": 1308.92 |
| }, |
| { |
| "title": "What the results mean", |
| "start": 1459.7 |
| }, |
| { |
| "title": "What comes next", |
| "start": 1568.58 |
| }, |
| { |
| "title": "Glossary", |
| "start": 1629.7 |
| }, |
| { |
| "title": "Self-check questions", |
| "start": 1691.14 |
| } |
| ] |
| }, |
| { |
| "id": "compression-is-compute", |
| "n": 22, |
| "title": "Compression Is Part of the Compute", |
| "series": "Quantization & Kernels", |
| "audio": "audio/compression-is-compute.mp3", |
| "bytes": 35401847, |
| "duration": 2212.58, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/compression-is-compute.md", |
| "bg": "https://images.unsplash.com/photo-1597852074816-d933c7d2b988?w=1920&q=80", |
| "published": "2026-09-25", |
| "sha256": "595024d6a809f5a53d0e8cfae5ed589ce1ac4f47e13d30ffb38de16cc0c0b96e", |
| "chapters": [ |
| { |
| "title": "Introduction", |
| "start": 0 |
| }, |
| { |
| "title": "The obvious story", |
| "start": 24.51 |
| }, |
| { |
| "title": "Where the time goes", |
| "start": 133.99 |
| }, |
| { |
| "title": "What fewer bits buy on paper", |
| "start": 281.53 |
| }, |
| { |
| "title": "The hidden step: someone has to decode", |
| "start": 402.43 |
| }, |
| { |
| "title": "How the processor actually reads memory", |
| "start": 598.99 |
| }, |
| { |
| "title": "Registers and the shape the tensor core wants", |
| "start": 734.74 |
| }, |
| { |
| "title": "Non-uniform formats and the lookup problem", |
| "start": 851.52 |
| }, |
| { |
| "title": "FLUTE, piece by piece", |
| "start": 1010.33 |
| }, |
| { |
| "title": "The same move, elsewhere", |
| "start": 1298.07 |
| }, |
| { |
| "title": "When smaller is slower", |
| "start": 1442.77 |
| }, |
| { |
| "title": "When batching changes the answer", |
| "start": 1595.39 |
| }, |
| { |
| "title": "Hardware that speaks the format", |
| "start": 1724.72 |
| }, |
| { |
| "title": "A design rule for our own hardware", |
| "start": 1873.15 |
| }, |
| { |
| "title": "What to remember", |
| "start": 2016.26 |
| } |
| ] |
| }, |
| { |
| "id": "trellis-quantization", |
| "n": 23, |
| "title": "Trellis Quantization: Trading Memory for Cheap Compute", |
| "series": "Quantization & Kernels", |
| "audio": "audio/trellis-quantization.mp3", |
| "bytes": 46957195, |
| "duration": 2934.79, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/trellis-quantization.md", |
| "bg": "https://images.unsplash.com/photo-1555617981-dac3880eac6e?w=1920&q=80", |
| "published": "2026-09-25", |
| "sha256": "e8a07cb9f32e7ba9271ab7684b6b74d9c5c65d4dbcd61767b73c5feb6597159e", |
| "chapters": [ |
| { |
| "title": "Introduction", |
| "start": 0 |
| }, |
| { |
| "title": "Cold open", |
| "start": 35.98 |
| }, |
| { |
| "title": "Part one - the memory wall at batch one", |
| "start": 127.83 |
| }, |
| { |
| "title": "Part two - bytes are the expensive part", |
| "start": 293.01 |
| }, |
| { |
| "title": "Part three - the idle integer units", |
| "start": 401.46 |
| }, |
| { |
| "title": "Part four - the scalar grid and where it breaks", |
| "start": 566.66 |
| }, |
| { |
| "title": "Part five - vector quantization and the codebook explosion", |
| "start": 685.81 |
| }, |
| { |
| "title": "Part six - making weights look Gaussian", |
| "start": 828.61 |
| }, |
| { |
| "title": "Part seven - a trellis, borrowed from modems", |
| "start": 923.92 |
| }, |
| { |
| "title": "Part eight - the bitshift trellis", |
| "start": 1069.27 |
| }, |
| { |
| "title": "Part nine - a codebook made of instructions", |
| "start": 1227.38 |
| }, |
| { |
| "title": "Part ten - what QTIP bought", |
| "start": 1427.08 |
| }, |
| { |
| "title": "Part eleven - ExLlamaV3: turning a paper into a format", |
| "start": 1551.51 |
| }, |
| { |
| "title": "Part twelve - ik llama dot cpp: one idea, many machines", |
| "start": 1836.3 |
| }, |
| { |
| "title": "Part thirteen - why the same code runs differently everywhere", |
| "start": 2111.99 |
| }, |
| { |
| "title": "Part fourteen - prefill, the other half of the bill", |
| "start": 2291.53 |
| }, |
| { |
| "title": "Part fifteen - the lesson for custom hardware", |
| "start": 2392.0 |
| }, |
| { |
| "title": "Where the field is heading", |
| "start": 2608.99 |
| }, |
| { |
| "title": "Recap", |
| "start": 2675.19 |
| }, |
| { |
| "title": "Self-check questions", |
| "start": 2782.92 |
| } |
| ] |
| }, |
| { |
| "id": "llm-memory-hierarchy", |
| "n": 24, |
| "title": "Memory Hierarchy for LLMs", |
| "series": "Quantization & Kernels", |
| "audio": "audio/llm-memory-hierarchy.mp3", |
| "bytes": 24748142, |
| "duration": 1546.73, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/llm-memory-hierarchy.md", |
| "bg": "https://images.unsplash.com/photo-1591488320449-011701bb6704?w=1920&q=80", |
| "published": "2026-09-26", |
| "sha256": "461c88f9b09e0e1b8e4b2958e586d97ba72df9ecb28b66eef210986595de7e1b", |
| "chapters": [ |
| { |
| "title": "Opening: why LLM performance is really about moving data", |
| "start": 7.05 |
| }, |
| { |
| "title": "Start with the complete memory journey", |
| "start": 55.04 |
| }, |
| { |
| "title": "VRAM, HBM, DRAM, and SRAM", |
| "start": 122.13 |
| }, |
| { |
| "title": "The GPU is made of streaming multiprocessors", |
| "start": 195.32 |
| }, |
| { |
| "title": "Registers - each thread's private working area", |
| "start": 259.59 |
| }, |
| { |
| "title": "Why we want many warps: latency hiding", |
| "start": 331.25 |
| }, |
| { |
| "title": "Shared memory - the block's cooperative scratchpad", |
| "start": 397.99 |
| }, |
| { |
| "title": "Coalescing and bank conflicts", |
| "start": 470.97 |
| }, |
| { |
| "title": "Tensor Cores - the arithmetic engines", |
| "start": 536.55 |
| }, |
| { |
| "title": "Tiling: how a GEMM uses the hierarchy", |
| "start": 604.45 |
| }, |
| { |
| "title": "Pipeline the loads instead of waiting", |
| "start": 672.56 |
| }, |
| { |
| "title": "Prefill versus decode", |
| "start": 727.64 |
| }, |
| { |
| "title": "Arithmetic intensity and the roofline idea", |
| "start": 794.83 |
| }, |
| { |
| "title": "Why quantization can speed up decode", |
| "start": 858.02 |
| }, |
| { |
| "title": "The KV cache is another bandwidth problem", |
| "start": 929.5 |
| }, |
| { |
| "title": "FlashAttention: same math, smarter movement", |
| "start": 984.09 |
| }, |
| { |
| "title": "Kernel fusion follows the same principle", |
| "start": 1049.32 |
| }, |
| { |
| "title": "Mixture of experts and prefetching", |
| "start": 1108.35 |
| }, |
| { |
| "title": "L two cache versus shared memory in prefetching", |
| "start": 1182.63 |
| }, |
| { |
| "title": "One complete decode-token walkthrough", |
| "start": 1244.33 |
| }, |
| { |
| "title": "Why a giant GPU can still struggle with one sequence", |
| "start": 1327.99 |
| }, |
| { |
| "title": "A practical framework for analyzing any LLM kernel", |
| "start": 1393.74 |
| }, |
| { |
| "title": "Final recap", |
| "start": 1446.74 |
| } |
| ] |
| }, |
| { |
| "id": "training-free-compression", |
| "n": 25, |
| "title": "Compression Without Retraining", |
| "series": "Quantization & Kernels", |
| "audio": "audio/training-free-compression.mp3", |
| "bytes": 34043634, |
| "duration": 2127.7, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/training-free-compression.md", |
| "bg": "https://images.unsplash.com/photo-1555617981-dac3880eac6e?w=1920&q=80", |
| "published": "2026-10-02", |
| "sha256": "cbe7eeff855fbf6d4342a6c19a9c0b9533393dc2d85a6e8b22e5798bd5d006dd", |
| "chapters": [ |
| { |
| "title": "Introduction", |
| "start": 0 |
| }, |
| { |
| "title": "Why training-free", |
| "start": 25.86 |
| }, |
| { |
| "title": "What training-free actually means", |
| "start": 95.71 |
| }, |
| { |
| "title": "The objective almost every method shares", |
| "start": 162.16 |
| }, |
| { |
| "title": "Weight-only quantization: four bits is solved", |
| "start": 255.08 |
| }, |
| { |
| "title": "Three bits and two bits: where the frontier is", |
| "start": 332.54 |
| }, |
| { |
| "title": "Quantizing activations too", |
| "start": 432.41 |
| }, |
| { |
| "title": "The hardware formats", |
| "start": 572.62 |
| }, |
| { |
| "title": "Compressing the attention cache", |
| "start": 647.21 |
| }, |
| { |
| "title": "Pruning individual weights", |
| "start": 697.46 |
| }, |
| { |
| "title": "Narrower and smaller matrices", |
| "start": 784.57 |
| }, |
| { |
| "title": "Removing layers and experts", |
| "start": 925.06 |
| }, |
| { |
| "title": "The verdict at equal memory", |
| "start": 996.5 |
| }, |
| { |
| "title": "A uniform head-to-head study", |
| "start": 1044.9 |
| }, |
| { |
| "title": "Where it actually gets faster", |
| "start": 1484.68 |
| }, |
| { |
| "title": "Improving it statistically", |
| "start": 1539.51 |
| }, |
| { |
| "title": "Improving it for efficiency", |
| "start": 1895.43 |
| }, |
| { |
| "title": "What to remember", |
| "start": 1928.61 |
| }, |
| { |
| "title": "Self-check questions", |
| "start": 1984.24 |
| } |
| ] |
| }, |
| { |
| "id": "checkpoint-to-silicon", |
| "n": 26, |
| "title": "From Checkpoint to Silicon", |
| "series": "Inference & Serving", |
| "audio": "audio/checkpoint-to-silicon.mp3", |
| "bytes": 50924697, |
| "duration": 3182.76, |
| "doc": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/main/lectures/checkpoint-to-silicon.md", |
| "bg": "https://images.unsplash.com/photo-1518770660439-4636190af475?w=1920&q=80", |
| "published": "2026-10-13", |
| "sha256": "112e0692df8a3d8441227c182ec360ef540546c0f5974573b7340e7849e9687c", |
| "chapters": [ |
| { |
| "title": "Introduction", |
| "start": 0 |
| }, |
| { |
| "title": "What Training Actually Produces", |
| "start": 53.34 |
| }, |
| { |
| "title": "Tokens, and One Pass Through the Model", |
| "start": 349.4 |
| }, |
| { |
| "title": "Prefill, Decode, and the Key-Value Cache", |
| "start": 507.97 |
| }, |
| { |
| "title": "Inside the Inference Runtime", |
| "start": 628.47 |
| }, |
| { |
| "title": "From Source Code to Machine Code", |
| "start": 825.78 |
| }, |
| { |
| "title": "RISC-V and Its Vector Instructions", |
| "start": 969.77 |
| }, |
| { |
| "title": "The Memory Wall", |
| "start": 1251.92 |
| }, |
| { |
| "title": "How a GPU Works", |
| "start": 1497.54 |
| }, |
| { |
| "title": "CPUs and GPUs Working Together", |
| "start": 1732.21 |
| }, |
| { |
| "title": "Memory, Long Prompts, and FlashAttention", |
| "start": 1877.35 |
| }, |
| { |
| "title": "Serving Many Users at Once", |
| "start": 2103.39 |
| }, |
| { |
| "title": "Paged Caches and Speculative Decoding", |
| "start": 2301.12 |
| }, |
| { |
| "title": "Scaling Across Several GPUs", |
| "start": 2425.4 |
| }, |
| { |
| "title": "A Real Deployment", |
| "start": 2610.6 |
| }, |
| { |
| "title": "Measuring and Diagnosing Performance", |
| "start": 2755.46 |
| }, |
| { |
| "title": "The Whole Journey", |
| "start": 2917.85 |
| } |
| ] |
| }, |
| { |
| "id": "locked-a610335b", |
| "title": "Locked lecture", |
| "series": "Locked", |
| "audio": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/65c32825f126898ecbb32543ae1f301b4576511d/locked/a610335b.mp3.enc", |
| "bytes": 22283656, |
| "duration": 1392.7, |
| "doc": null, |
| "bg": null, |
| "published": "2026-09-27", |
| "sha256": "1e802a8a67fb8798926f9df89e7ce04ae5c47598e27a4d50766d3de0e2d1f02a", |
| "locked": { |
| "v": 1, |
| "kdf": { |
| "name": "PBKDF2", |
| "hash": "SHA-256", |
| "iterations": 600000, |
| "salt": "0W2/yf1qjt9vZ0RgZKvoAQ==" |
| }, |
| "meta": { |
| "iv": "TWkFeARzJZuJWtqM", |
| "ct": "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" |
| } |
| }, |
| "n": 27 |
| }, |
| { |
| "id": "locked-73a37eb8", |
| "title": "Locked lecture", |
| "series": "Locked", |
| "audio": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/453b9094882e90bfb0e510621439d913c027a814/locked/73a37eb8.mp3.enc", |
| "bytes": 25748111, |
| "duration": 1609.22, |
| "doc": null, |
| "bg": null, |
| "published": "2026-09-30", |
| "sha256": "a9eb291f12e6d6d0133e24d6b2f7a0147cbc2a68596b346319be115e1177f1e6", |
| "locked": { |
| "v": 1, |
| "kdf": { |
| "name": "PBKDF2", |
| "hash": "SHA-256", |
| "iterations": 600000, |
| "salt": "8Al5baxnewp6MkhLy1HzGQ==" |
| }, |
| "meta": { |
| "iv": "cQAOvjno8yF5g0Sq", |
| "ct": "GJdX9ChiPlx1VnT4Or3aiad/8uFWw4s27q9SjmtfqzhTummwsRaRtwLg78eL2xyMrvv4WwGBwYrZpYSkP1xg2nBjvX1NdElRs8MOLHXK772n84MHrlrp05AGygiG+t3Ng9XujezpUuUAgv+WzuDePUHAp2BsgJTpxdhAkTBJBn6z+Ce8hI3IN3zKBQP5OhZh5d+CkBbvGsTImtJQUyie7xfvRxlg6SCRRAgnEy0PoOz3NrAgUO8d+DvWXiRW5y21bSv3oo/NbDpNjVbp887sZ0i9wYqcrttpJMfOR43iAiz+5x7Scburf96cTjBmaJrHY8+Sb4dvb/pSUIIOaSQCLhEM+9+7KNnvae4sMXgV0kP/XgmNLoNlJwr+W1SY0hwdIShzbj9FPCCltOUiy3NTogqVqp3s9ju9RstgERibGTGZJdG0KuDEAPsP0QCGbxlpuhxYWsQ1BfkbSgJqa3kkZAGtUj6XkC/YOuMIrcM1imbOCzNBRiOwMu7MwEjgbni7GoaJax+3Q9XAMtarSwzdwiIoiLiXNsBqRd9THUXv4G5qaKDht1pZzbyvcYPJbPbFknIG8duVW8M/mZ1ych5CxokC/1WWqw97+f43oDP6K8ef049sn0dPwRZhFAXcRF8KG8A9Dp/U8osnPs4jSCvLhR7r+bG0oDti2hS8NXncwBMrh7/dncC+Y/hR2ECYhja5mIcghrBqdZHeiR6VCqm6S6fls6SUiBWyQvECG3s723qzRGP6lDcUkv97nGaNLQCLgv4NTVg+oPPdgfntnnz2hZcNFC9/CxE+ukQWhpWtxp2N3Z5K+DiFT+dqqudzitRZekQC2Doh+fJizXvU2r6shFc6+RIxxRE0k6qcLJgIW5qpL8ANgvKa0ee192aJ4cQaExMpeocPMZ9xEg7i1dhTHMtw5pUSQQHg4zEmYasEMGCk+qzVhIhcSYveiSAvqd4bjTSWnqwrtQALh5UgNMmQ/OIQepkhr7PU5gzllQo+hjtiw5GH8v4RiMJ6v+DcrFCOL3KmNToN+zTv3LN8a8l4t+nNHJ6LLS/GnyDCc2SiU/7YI0yKtaerWlYPiRIFvfjZ17q3+iU9W7GTbNa6iExy8Qn7+Nkgzak9tvlOAzdKJ8UP1sSh86Cr41SX6Yp/V26k7eE3S2L16bKRRZSPfQtuIhjged5b4ypOY7NE8X8HvOCvqwe8QK9cFQPGXJQbKb2HrLV3X6XgG1UvXB4a5iQGzBgV7zDzciD06dEy652XNz1HQDVT27vHNc+AlheLuTTBg4OBpnIMip/aQRaDXnm/fbEAZS9V6DLZL8oKk9m34OFtiRFZtVDn5fDiO7u+ExIcUcw4WZAuyvEjfoetJ4d49pqVjfKapbLPnwgN0n38aqfYo9CVRrzpRToQIyxS4544WMsdP/665JvV4LkqqrGy8+aF0C/t2gykY98fsYAtiY20tDj71/nPa9/wepyHmjz3IjNPesmXH74u33t+/3Q+ZsqmJ8RtHMyqSXVqFBgDx4wC1vzMpKbjDh/SIV620vC6SMGRjWI8eh7KHF/IN4HVOnd3QqSy40njm2iO9vnEg+j3O+qkLl4afZ+vVUfMBuFCLefHZjNYrMCJaF9w0baD3Jl5I60vWykPp0inAwLzS4tUojLWiyHmoB1g0DQEp94xoWwqufO/tZXkSICqTyHxNMJfDVFsyB00hrnQUaTHDHp9s/HV4MRtlRgrXolL0SnpIlCy2Rz28VXn+8csr8oc/86UJZU5cQNpAwErbyRP/HTAAqbnaM3b+cPKG5QCh7wePRgOwttYCoiD1XBuK4QFR10=" |
| } |
| }, |
| "n": 28 |
| }, |
| { |
| "id": "locked-612177a9", |
| "title": "Locked lecture", |
| "series": "Locked", |
| "audio": "https://huggingface.co/datasets/arjun10g/papers-audio/resolve/9581c80954915c038821ba8f8a7726f784afbca3/locked/612177a9.mp3.enc", |
| "bytes": 21072182, |
| "duration": 1316.98, |
| "doc": null, |
| "bg": null, |
| "published": "2026-10-01", |
| "sha256": "e61bcc56279d18c7faa8e60ff56690d34556e8747c3698c2072b4e8061c9b5d7", |
| "locked": { |
| "v": 1, |
| "kdf": { |
| "name": "PBKDF2", |
| "hash": "SHA-256", |
| "iterations": 600000, |
| "salt": "qn9IzSCeT5i76BN7xgDINQ==" |
| }, |
| "meta": { |
| "iv": "xA2jSEnkySF9hycj", |
| "ct": "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" |
| } |
| }, |
| "n": 29 |
| } |
| ] |
| } |
|
|
