Video Learning Hub
Master multimodal AI concepts through comprehensive tutorials, guides, and best practices from our expert team.
Trusted by engineers at
Done watching? Point Mixpeek at your own video, image, audio, and document storage and make it searchable — free.
Search your own data, free
The $50K Licensing Deal That Dies in an Unindexed Archive
For a footage licensing business, search latency is the close rate: a buyer's request has a window, and every request unfulfilled inside it is revenue for whoever answers first. This video walks the failure case (40,000 unindexed hours, a three-day manual search against a six-hour deadline, the deal licensed elsewhere at 4:52) and the indexed alternative: every matching moment surfaces with timecodes in seconds, faces matched, transcripts searched. Explore: mixpeek.com/use-cases/media-archive-face-search
Late Interaction: How ColBERT and ColPali Search Documents as Images
Bi-encoders squash query and document into single vectors (fast, lossy); cross-encoders read them jointly (accurate, unscalable). Late interaction (ColBERT, 2020) sits between: every token keeps its own embedding and they interact only at scoring, where each query word takes its best-matching token (MaxSim). ColPali (2024) extends this to visual documents: PaliGemma-3B embeds raw page images as patch vectors (D=128), queries hunt patches, and the OCR pipeline disappears from indexing. ViDoRe: 81.3 avg nDCG@5 vs 67.0 (captioning) and 66.1 (OCR + BGE-M3). Explore: mixpeek.com/visual-document-retrieval
Thompson Sampling: Why Your A/B Test Wastes Half Its Traffic
A fixed 50/50 A/B test sends half of every cohort to the inferior variant until the test ends. Multi-armed bandits reframe the decision as allocation under uncertainty, and Thompson sampling (1933) solves it in one loop: keep a belief distribution per option, sample once from each every round, play the highest sample, update with the result. Uncertain options explore themselves, winners take over automatically, and regret is provably near-optimal (Agrawal & Goyal 2012). Netflix picks homescreen artwork with contextual bandits; ranking under uncertain relevance is the same math. Explore: mixpeek.com/glossary/relevance-ranking
Two-Tower Architecture: The Two Networks Behind Every Feed
You cannot run a big model a billion times per request, so recommenders split it: a user tower and an item tower embed into the same vector space, trained contrastively on clicks, meeting only at a dot product. Item vectors are precomputed offline into an ANN index; one query-tower pass returns hundreds of candidates from millions in milliseconds (YouTube's candidate-generation design), and a heavy reranker scores just that shortlist. The trade is fine-grained query-item interaction, which rerankers and late-interaction models recover downstream. Explore: mixpeek.com/diagrams/muvera-two-tier-retrieval
Meta's SAM 3: Type a Phrase, Segment Every Instance
SAM 1 and 2 returned one mask per click. SAM 3 (Meta, November 2025, open weights) introduces promptable concept segmentation: a short noun phrase or exemplar image makes one model detect, segment, and track every instance of that concept across images and video. A shared Perception Encoder feeds a DETR-based detector and a SAM 2 memory tracker, trained on 4M+ auto-annotated concepts. Results: 55.7 cgF1 on SA-Co Gold vs 24.5 for OWLv2, 75-80% of human performance, 30ms per image with 100+ objects; SAM 3.1 tracks 16 objects in one pass at 32 fps. Already powers Instagram Edits and Marketplace View in Room. Explore: ai.meta.com/blog/segment-anything-model-3
Google's MUVERA: The One-Vector Trick Behind Fast Multi-Vector Search
Multi-vector models like ColBERT keep one embedding per token, which wins on accuracy but breaks fast indexes: Chamfer scoring forces custom pipelines like PLAID. MUVERA's Fixed Dimensional Encodings randomly partition the embedding space, aggregate each region's tokens (queries sum, documents average), and concatenate the blocks into one vector whose dot product provably approximates the multi-vector score, the first single-vector proxy with theoretical guarantees. On BEIR: about 10% higher recall at 90% lower latency than PLAID, 2 to 5x fewer candidates, 32x memory compression via product quantization. Paper: arxiv.org/abs/2405.19504
Baidu's Unlimited-OCR: The Model Whose Memory Never Grows
Long-document OCR hits a memory wall: the KV cache grows with every token written, so pipelines chop PDFs into single pages and tables get cut at page breaks. Baidu's Unlimited-OCR (MIT, June 2026) replaces decoder attention with Reference Sliding Window Attention: each token attends to the document's visual tokens plus the last 128 outputs, making the cache a fixed-size queue. 40+ page PDFs parse in one forward pass, scoring 93.23 on OmniDocBench v1.5 vs DeepSeek-OCR's 87.01, about 12% faster, from a 3B MoE with 500M active. Paper: arxiv.org/abs/2606.23050
Mamba Explained: The LLM Architecture With No Attention (5x Faster, Million-Token Context)
Transformers pay a quadratic tax: every new token re-reads the entire sequence, and the cache grows forever. Mamba replaces attention with a selective state space model: one fixed-size state updated per token, with a learned keep-or-forget decision. The paper reports linear scaling, 5x Transformer inference throughput, a 3B model matching Transformers twice its size, and quality holding up to million-token sequences. Paper: arxiv.org/abs/2312.00752
LoRA Fine-Tuning Explained on a Real Model: Train 1% of Qwen2.5-1.5B
Fine-tuning nudges a pretrained model's weights toward your data, but classic fine-tuning updates all of them and carries optimizer state for each. LoRA freezes the base model and trains a low-rank pair beside each weight matrix: on Qwen2.5-1.5B a 1536x1536 attention matrix holds about 2.36M numbers while its rank-8 adapter holds 24,576, roughly 1% of the network overall, merged in afterward with zero added inference latency. Includes the operational rule: adapters teach behavior, retrieval carries knowledge. Paper: arxiv.org/abs/2106.09685