---
title: "Machine Learning"
description: "622 posts about Machine Learning, summarised, each linking to the original."
---

# Machine Learning
> 622 posts about Machine Learning, summarised, each linking to the original.

## Articles

### [Together AI at ICML 2026: frontier research across the full stack](https://yomu.fyi/post/together-ai-at-icml-2026-frontier-research-across-the-full-stack.md)
- Company: [Together AI](https://yomu.fyi/company/together-ai.md)
- Published: Jun 30, 2026
- Type: Announcement

Together AI says nine papers from its researchers and peers were accepted to ICML 2026 in Seoul, spanning frontier agents, model shaping, algorithmic and systems optimization, and kernels. The collection is presented as a full-stack view of AI research, with work ranging from DSGym’s unified evaluation and training API for more than 1,000 data-science tasks to ThunderAgent’s program-aware scheduling, TTT-Discover’s test-time reinforcement learning, and RARO’s verifier-free adversarial training. Reported results include up to 3.6× agent throughput, state-of-the-art discoveries from an open 120B model, a 25% expert-response win rate without a verifier, and 5M-token context on one node for Untied Ulysses. Aurora illustrates the research-to-production loop: its adaptive speculative-decoding work ships as the ATLAS speculator on the Together platform.


### [How Jua delivers the world’s most accurate physics simulations 3x faster with ClickHouse Cloud](https://yomu.fyi/post/how-jua-delivers-the-world-s-most-accurate-physics-simulations-3x-fast.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jun 30, 2026
- Type: Problem & solution

Jua uses ClickHouse Cloud to deliver data from EPT-2, its physics foundation model, to energy traders and grid operators faster than its previous file-based pipeline. Jua’s forecasts combine outputs from multiple models and add roughly four terabytes of new data daily, producing about one petabyte compressed and over 10–12 petabytes uncompressed. Replacing Zarr and other file-based storage plus manual pre-fetching with ClickHouse Cloud reduced flagship forecast delivery from one hour to 20 minutes, with about 20 seconds of upload overhead. The platform also cut compute costs by roughly a third, reduced two-year single-location historical queries from hours to seconds, and supports both point lookups and regional aggregations; Jua selected the managed service for operational simplicity, separated storage and compute, and cost efficiency.


### [GenPage: Towards End-to-End Generative Homepage Construction at Netflix](https://yomu.fyi/post/genpage-towards-end-to-end-generative-homepage-construction-at-netflix.md)
- Company: [Netflix](https://yomu.fyi/company/netflix.md)
- Author: Netflix Technology Blog
- Published: Jun 29, 2026
- Type: Problem & solution

Netflix traditionally constructs its structured two-dimensional homepage through a complex, multi-stage recommender pipeline that separates candidate generation and ranking across rows and entities. To simplify this architecture and optimize directly for whole-page user satisfaction, Netflix developed GenPage, an end-to-end generative transformer model that autoregressively builds the entire homepage from raw tokenized context. The system relies on a domain-specific tokenizer to compress engagement history, context injection for cold start, hybrid row decoding to minimize decoding steps, and reinforcement learning post-training. In online A/B testing against the production baseline, GenPage delivered statistically significant gains in core user engagement metrics while cutting end-to-end serving latency by 20 percent. Offline evaluations further showed that enriching context prompts improved recommendation quality more effectively than increasing model capacity in the current operating regime.


### [Creating the NVIDIA Nemotron 3 Ultra NVFP4 Checkpoint with NVIDIA Model Optimizer](https://yomu.fyi/post/creating-the-nvidia-nemotron-3-ultra-nvfp4-checkpoint-with-nvidia-mode.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Jun 26, 2026
- Type: Problem & solution

Moving large model weights efficiently requires quantization, but mapping weights to NVFP4's eight positive values risks severe precision loss or outlier-induced underflow. To address this, NVIDIA quantized the 550B Nemotron 3 Ultra model using NVIDIA Model Optimizer with a multi-precision strategy rather than quantizing all layers uniformly. The recipe applies four-over-six adaptive block scaling to Mixture of Experts routed weights, while assigning FP8 to shared experts and retaining BF16 for sensitive attention linears and embeddings. This process reduced the checkpoint size from 1,121 GB down to 352.3 GB and enabled single-checkpoint execution across Hopper using W4A16 and Blackwell using native W4A4. As a result, the model achieves up to 5.9x higher inference throughput than the GLM-5.1 754B FP4 model on decode-heavy workloads while matching BF16 benchmark accuracy.


### [Scaling AI Inference Across Multiple GPUs Using NVIDIA TensorRT with Multi-Device Inference Support](https://yomu.fyi/post/scaling-ai-inference-across-multiple-gpus-using-nvidia-tensorrt-with-m.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Peter Kisfaludi
- Published: Jun 25, 2026
- Type: Tutorial

Generative AI media pipelines can exceed the memory and compute capacity of a single GPU, creating a deployment challenge for models that retain TensorRT’s kernel fusions, memory planning, and quantization. TensorRT 11.0 introduces native multi-device inference and IDistCollectiveLayer primitives, while NCCL supplies topology-aware transport across NVLink, NVSwitch, PCIe, and InfiniBand for supported distributed collectives. The article contrasts tensor parallelism, which shards layer weights, with context parallelism, which partitions sequences; it then focuses on AllGather KV, Ring Attention, and DeepSpeed Ulysses for long-sequence media attention. It also describes converting PyTorch pipelines with Torch-TensorRT into C++-deployable engines, benchmarking Cosmos 3 and FLUX.1 on an eight-GPU node, and provides C++ and OpenMPI steps for configuring an eight-rank NCCL deployment.


### [Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Powered by NVIDIA ACE](https://yomu.fyi/post/q-a-how-krafton-built-pubg-ally-a-co-playable-character-powered-by-nvi.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Jun 25, 2026
- Type: Explainer

KRAFTON developed PUBG Ally, an on-device AI co-playable character for PUBG: BATTLEGROUNDS, using NVIDIA ACE tooling and models. Conventional non-playable characters rely on fixed dialogue, but PUBG Ally processes live game states and player voice input in real time. The interactive pipeline combines automatic speech recognition, a quantized Mistral-NeMo-Minitron-2B small language model, and a custom text-to-speech model running locally on client GPUs with at least 8GB of VRAM. To minimize latency, the architecture separates reactive gameplay into a behavior-tree System 1 layer while routing strategic decisions and speech generation to the language model. KRAFTON tuned the non-deterministic system across multi-layer evaluations and large-scale playtests with over a thousand players.


### [How we used DSPy to turn AI evaluations into better responses in Dash chat](https://yomu.fyi/post/how-we-used-dspy-to-turn-ai-evaluations-into-better-responses-in-dash.md)
- Company: [Dropbox](https://yomu.fyi/company/dropbox.md)
- Author: Ilya Yakovlev,Andrew Cheung,Binoy Dash,Simran Jumani,Dmitriy Meyerzon
- Published: Jun 25, 2026

Dropbox improved its Dash chat agent by establishing an automated optimization loop powered by DSPy and LLM-as-judge evaluations. Engineers first calibrated their LLM judges against human-annotated interaction traces, then used those judges to systematically optimize the agent's system prompts via offline counterfactual replay. This automated workflow reduced incomplete responses by 26% while decreasing overall token consumption.


### [Achieving Near-Linear Training Scalability for Pinterest’s Foundation Models](https://yomu.fyi/post/achieving-near-linear-training-scalability-for-pinterest-s-foundation.md)
- Company: [Pinterest](https://yomu.fyi/company/pinterest.md)
- Author: Pinterest Engineering
- Published: Jun 25, 2026
- Type: Problem & solution

Pinterest’s foundation ranking models store approximately 99% of their parameters in embedding tables, causing cross-node NCCL all-to-all communication to severely bottleneck multi-node distributed training. Initial multi-node scaling stalled at 1.21x on four nodes even after adopting AWS Elastic Fabric Adapter for OS-bypass networking. To resolve this, engineers profiled GPU activity with PyTorch Profiler and systematically addressed communication bottlenecks across multiple layers. The team implemented FP8 quantized communications using FBGEMM, balanced table sharding across GPUs, halved embedding dimensions while doubling row counts, and adopted a 2D parallel communication topology alongside PyTorch Distributed Checkpoint. These combined interventions achieved near-linear scaling factors of 3.9x on four nodes and 7.5x on eight nodes across 64 GPUs.


### [The Rise of Sports Intelligence: How the Lakehouse Turns Tracking Data into Competitive Advantage](https://yomu.fyi/post/the-rise-of-sports-intelligence-how-the-lakehouse-turns-tracking-data.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Corey Abshire, Kush Patel, Nick Ragonese
- Published: Jun 24, 2026
- Type: Explainer

Professional basketball’s Hawk-Eye SkeleTRACK feed produces roughly 22,620 positional updates per second—about 65 million records per 48-minute game—yet teams often cannot turn that volume into timely, trusted decisions. The integration gap comes from separate vendors for tracking, wearables, video, scouting, and medical data, alongside calibration differences, weak provenance, and compute limits. The Databricks Data + AI Platform is presented as a governed lakehouse that ingests feeds with Lakeflow, refines them through medallion layers, and uses Unity Catalog for lineage, access control, and auditing. Models for shot probability, injury risk, and fatigue can run alongside serving and custom applications, with Lakebase supporting sub-second interactive queries. Applications include proactive load management, real-time coaching intelligence, and enriched broadcast or fan experiences across tracking-rich sports.


### [Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications](https://yomu.fyi/post/accelerating-bev-pooling-on-nvidia-gpus-for-physical-ai-applications.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: John Yang
- Published: Jun 24, 2026
- Type: Problem & solution

Bird's-eye-view (BEV) pooling frequently becomes an inference latency bottleneck in spatial AI pipelines due to irregular memory access patterns and scatter-reduce operations. BEVPoolV3 addresses this by tailoring kernel execution to the GPU memory regime, specifically distinguishing between DRAM-bound and L2-cache-resident working sets. The implementation reduces redundant scatter traffic via depth-outer traversal, precomputes indices to eliminate runtime integer division, minimizes duplicate depth loads, and adopts interval-owned writes to bypass atomic operations. On a canonical 49 MB nuScenes working set, BEVPoolV3 accelerates TensorRT plugin execution from 274.0 µs down to 16.4 µs on an NVIDIA RTX PRO 6000 Blackwell Max-Q GPU. Furthermore, profiling with Nsight Compute reveals that L2-resident scatter-reduce operators achieve optimal efficiency with FP8, avoiding the ALU instruction overhead of NVFP4 decoding.


### [Genesis Workbench: A blueprint for industry AI in life sciences, powered by Databricks and NVIDIA](https://yomu.fyi/post/genesis-workbench-a-blueprint-for-industry-ai-in-life-sciences-powered.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Mark Lee, Srijit Nair
- Published: Jun 23, 2026
- Type: Explainer

Genesis Workbench is an open blueprint for a Databricks-native life-sciences application that brings computational drug-discovery stages into one governed workbench. It combines Unity Catalog governance, MLflow tracking, Model Serving, serverless GPU compute, Databricks AI Search, and NVIDIA technologies including CUDA-X libraries, Parabricks, BioNeMo tools, GenMol, and Proteina-Complexa. Independent modules cover genomics, single-cell analysis, large- and small-molecule workflows, and model fine-tuning, with handoffs spanning gene-to-sequence resolution, structure prediction, docking, ADMET, and candidate ranking. A point-and-click React interface supports bench scientists, while declarative workflow generation and MCP exposure let pipelines and external clients use the workbench; inference runs on GPU endpoints inside the governed workspace without runtime external API dependencies. The stated aim is to let teams move from disease hypotheses to ranked therapeutic candidates on their own data, with a roadmap for automated workflow generation, BioNeMo Skills integration, and additional MCP services.


### [Maximize AI Factory Energy Efficiency Through Full-Stack Inference and Training Optimizations](https://yomu.fyi/post/maximize-ai-factory-energy-efficiency-through-full-stack-inference-and.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Sachin Idgunji
- Published: Jun 23, 2026
- Type: Explainer

Power costs comprise up to 40% of AI factory operating expenses, making performance per watt critical within fixed site power limits. Operators maximize throughput by combining architectural enhancements, narrow-precision formats like NVFP4, and mixture-of-experts model designs that activate fewer parameters per token. During distributed LLM training in Megatron-LM, tuning individual GPU speeds along critical execution paths eliminates idle power waste without extending total training time. Additionally, the NVIDIA DSX platform orchestrates dynamic power allocation, 45°C direct-to-chip liquid cooling, and grid integration across racks and facilities. Together, these full-stack optimizations enable performance-optimized AI factories to deliver up to 2.6 times more tokens per second per megawatt.


### [Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding](https://yomu.fyi/post/boost-inference-performance-up-to-15x-on-nvidia-blackwell-using-dflash.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Amr Elmeleegy
- Published: Jun 23, 2026
- Type: Benchmark

Sequential token generation in autoregressive LLMs limits GPU utilization and serving throughput, while conventional speculative decoding remains constrained by autoregressive draft models. DFlash addresses this bottleneck by employing a lightweight block-diffusion drafter that predicts an entire block of masked candidate tokens in a single forward pass. The architecture incorporates target hidden-state conditioning and key-value injection into draft model projections while relying on the target model for parallel verification. Evaluated on NVIDIA Blackwell GPUs across TensorRT-LLM, vLLM, and SGLang, DFlash increases gpt-oss-120b throughput by more than 15x at high interactivity targets and delivers up to 5.8x speedups on Gemma 4 31B compared to autoregressive decoding.


### [Build an AI Scientist for Life Science Discovery with NVIDIA BioNeMo Agent Toolkit](https://yomu.fyi/post/build-an-ai-scientist-for-life-science-discovery-with-nvidia-bionemo-a.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Kyle Tretina
- Published: Jun 23, 2026
- Type: Tutorial

General-purpose AI agents often struggle with biomolecular discovery because scientific workflows require precise model selection, strict input formatting, and artifact interpretation. The NVIDIA BioNeMo Agent Toolkit bridges this gap by providing BioNeMo Skills and Model Context Protocol wrappers on top of accelerated NVIDIA NIM microservices. These skills instruct agents on tool capabilities, required inputs, expected file formats like CIF and SMILES, and specific failure modes. Developers can route requests through hosted endpoints for testing or deploy local NIM microservices on GPU nodes to lower warm per-call latency during iterative candidate generation loops. Internal benchmarks using Codex CLI demonstrate that equipping agents with BioNeMo Skills improves task completion rates from 57.1% to 100% while doubling token efficiency across passing assertions.


### [What is Vector Search?](https://yomu.fyi/post/what-is-vector-search.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Databricks Staff
- Published: Jun 23, 2026
- Type: Explainer

Vector search retrieves results by comparing embeddings that represent meaning across text, images, audio, and other content rather than matching exact words. A model creates embeddings, an index stores them for fast similarity search, and a query embedding is matched against the index using nearest-neighbor methods. Exhaustive k-nearest neighbor search can become too slow at millions of items, so production systems commonly use approximate nearest neighbor search, trading some precision for speed. The guide positions vector search behind semantic search, RAG, recommendations, and multimodal or cross-language retrieval, while hybrid search combines dense and sparse vectors, keyword results, metadata filtering, and reranking to improve reliability. Quality depends on embeddings, filters, index freshness, and infrastructure, with vector search requiring more memory and compute; Databricks AI Search is presented as a managed service supporting these capabilities and Unity Catalog governance.


### [Data Lake vs. Cloud Data Warehouse: A Practical Guide for Data Scientists](https://yomu.fyi/post/data-lake-vs-cloud-data-warehouse-a-practical-guide-for-data-scientist.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Databricks Staff
- Published: Jun 23, 2026
- Type: Explainer

The guide contrasts data lakes and cloud data warehouses for storing and querying data at scale. Data lakes retain raw structured, semi-structured, and unstructured data in low-cost object storage with schema-on-read, while warehouses enforce schema-on-write for structured analytical workloads. Lakes fit petabyte-scale machine learning, data science, and undefined future use cases; warehouses fit fast, concurrent SQL for dashboards, reporting, and operational analytics. It describes Bronze, Silver, and Gold zones, with Parquet and ORC supporting columnar scans and open-format portability. For teams combining ML and BI, lakehouses use Delta Lake, Apache Iceberg, or Apache Hudi to add ACID transactions, schema enforcement, and quality monitoring to lake storage without duplication; catalogs, staged checks, and access controls help prevent data swamps.


### [How Telcos Build Autonomous Networks with Agentic AI](https://yomu.fyi/post/how-telcos-build-autonomous-networks-with-agentic-ai.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Amogh Dendukuri
- Published: Jun 23, 2026
- Type: Explainer

Telecom operators currently operate mostly within Levels 2 and 3 of the TM Forum autonomous networks taxonomy, relying on predefined scripts. Progressing to Levels 4 and 5 demands autonomous systems capable of intent comprehension, real-time telemetry sensing, and cross-domain action coordination. To address this, an autonomy platform architecture combines domain-specific reasoning models, agent harnesses, and isolated sandboxed runtimes. Workflows separate responsibilities among on-demand agents, long-running execution loops, and deep-research multi-agent systems that evaluate ranked remediation paths or design wireless algorithms. Initial experiments in simulated SR-MPLS routing and wireless physical or MAC layer design show autonomous discovery can deliver measurable gains over standard operational and engineering baselines.


### [Toward More Controllable AI Video Editing: An Early Research Exploration at Netflix](https://yomu.fyi/post/toward-more-controllable-ai-video-editing-an-early-research-exploratio.md)
- Company: [Netflix](https://yomu.fyi/company/netflix.md)
- Author: Netflix Technology Blog
- Published: Jun 23, 2026
- Type: Announcement

Generative video editing tools often regenerate entire video clips, which inadvertently modifies untouched scene details or violates physical continuity during object deletion. To provide artists with precise control, two research prototypes were developed: Vera and VOID. Vera uses a layered video diffusion framework with a Mixture-of-Transformers architecture to jointly generate separate edit and alpha matte layers, compositing them with original footage while leaving unchanged pixels untouched. VOID addresses object deletion by conditioning on interaction-aware quadmasks and using a second-pass appearance refiner to reconstruct scenes with plausible physical continuity. In a user study evaluated across 75 real-world scenarios, reviewers selected VOID 64.8% of the time over six baselines.


### [Experimenting with the proposed Cross-Origin Storage API in Transformers.js](https://yomu.fyi/post/experimenting-with-the-proposed-cross-origin-storage-api-in-transforme.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Thomas Steiner
- Published: Jun 23, 2026
- Type: Tutorial

Transformers.js browser apps currently redownload and recache identical AI model and WebAssembly resources when they are requested from different origins. Chrome's cache uses a Network Isolation Key composed of the top-level site and current-frame site, so matching CDN URLs do not produce cache hits; the toy example incurs 177 MB of duplicate model storage, while a shared ONNX Runtime Wasm file is 4,733 kB. The proposed Cross-Origin Storage (COS) API addresses this by identifying large files with cryptographic hashes and sharing them across opted-in origins, with a polyfill extension enabling experimentation before native browser support. In the extended Transformers.js flow, a COS lookup can serve resources in milliseconds, while a miss downloads and stores them, and ModelRegistry.is\_pipeline\_cached() can help select among models already cached.


### [Enable Real-Time AI for High-Speed Data Acquisition with DAQIRI](https://yomu.fyi/post/enable-real-time-ai-for-high-speed-data-acquisition-with-daqiri.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Cara Laasch
- Published: Jun 22, 2026
- Type: Announcement

Scientific instruments and high-bandwidth sensors generate immense data volumes, but traditional architectures rely on a store-then-analyze model that bottlenecks real-time discovery and forces extensive data discarding. NVIDIA DAQIRI shifts data acquisition to a software-centric pipeline that connects streaming detectors directly to GPU computing. Built as part of the NVIDIA Holoscan Platform, DAQIRI uses DPDK to bypass the Linux kernel and move UDP and RoCE v2 traffic directly from the NIC into GPU DMA buffers at line rates exceeding hundreds of gigabits per second. The system automates packet reordering, flow steering, and on-GPU data conversions such as mapping wire int4 payloads into fp16 tensors using YAML configuration files and simple C++ or Python APIs. This enables real-time edge processing, filtering, and model inference across experimental pipelines like CERN ATLAS.


[Newer posts](https://yomu.fyi/topic/machine-learning/page/10.md) · [Older posts](https://yomu.fyi/topic/machine-learning/page/12.md)