Machine Learning Engineer: 19th August 2025

LLM Evaluation: Practical Tips at Booking.com (booking​.ai). Evaluates LLMs using a 'judge-LLM' framework with golden datasets, annotation protocols, prompt engineering, and metrics like accuracy, F1-score, for optimizing Generative AI applications

Intern Experience at Lyft (eng​.lyft​.com). Lyft data scientists Morteza Taiebat and Han Gong recount internships on Sustainability and Driver Loyalty teams, using difference-in-differences, hierarchical linear models, CPIDH, causal prediction, and budget optimization for EV adoption, driver productivity, and referral incentives

Migrating Airbnb’s JVM Monorepo to Bazel (medium​.com/airbnb-engineering). Airbnb migrated their JVM monorepo from Gradle to Bazel, optimizing build speed, reliability, and scalability through remote execution, sandboxing, automated build file generation, and multi-version library support

SURP Student Spotlight: Alejandro Ortega Cruz Prieto (dunlap​.utoronto​.ca). SURP spotlight on Alejandro Ortega Cruz Prieto explores machine learning in astrophysics, polarization imaging, Stokes I, rotation measure, image fusion, discrete regime coding, and aims for a PhD in ML theory

My experience of searching for a job in 2024 as an MLE (andlukyane​.com). MLE job search in 2024: global interviews with 20+ companies, rounds up to 6, topics like string matching, Bert score, spell checkers, inference optimization, attention, BERT, ranking metrics, text-to-image, face recognition; delays, ownership, LeetCode, domain fits, Saudi Vision 2030, LM pivots, NDA constraints, team-matching, E5 offer

My Book is Out! (m-clark​.github​.io). Michael Clark announces publication of Models Demystified: A Practical Guide from Linear Regression to Deep Learning, co-authored with Seth Berry; print and web versions, 2025-08-15 release, print code discount, Routledge/CRC Press, GitHub, and book website links; AI boom context; personal milestones include founder, fatherhood

Bridging Minds and Disciplines: The IAIFI Summer School and the Future of Collaborative Science (firstprinciples​.org). IAIFI Summer School explores AI–physics frontier with lectures, tutorials, hackathons; projects include DESI cosmology data, latent space dynamics, physics-guided optimization, and domain adaptation

“My biggest lesson was realizing that domain expertise matters more than algorithmic complexity.“ (towardsdatascience​.com). Claudia Ng highlights domain expertise over algorithmic novelty in a Web3 credit scoring challenge, emphasizing MVP delivery, problem-driven consulting, multilingual AI learning products, and TTS for endangered languages

The Modern Data Toolbox (technology​.doximity​.com). Hybrid data toolbox blends LLMs, traditional ML, and statistics for real-time fraud detection, enhanced product discovery, and synthetic data generation with privacy-preserving techniques

Robotics Training Experiments: What Worked, What Didn't, and What Surprised Me (tinystruggles​.com). Robotics manipulation with SAC, HER, and imitation learning; dense vs sparse rewards, multi-environment SAC training, goal-conditioned environments, and stage-based exploration strategies

Glowing algae reveal the geometry of life (cam​.ac​.uk). Glowing Volvox carteri ECM reveals foam-like, foam-like network; pherophorin II fluorescence; stochastic geometry; machine learning for ECM compartment shapes; cross-disciplinary work Cambridge, Bielefeld; PNAS study

Complex deep learning models are no better at understanding genetic perturbation than simple baseline ones, study finds (phys​.org). Foundation models for single-cell transcriptomics fail to outperform simple baselines in predicting single and double gene perturbations; scGPT, scFoundation underperform vs additive models in transcriptomic effects

Basic Feature Engineering with DuckDB (duckdb​.org). DuckDB demonstrates SQL-based data preprocessing: missing value handling, categorical encoding, one-hot encoding, and feature scaling using FROM-first syntax, PIVOT, and aliasing on a synthetic financial fraud dataset

How to make UMAP plot in R (datavizpyr​.com). UMAP in R with umap package, Palmer penguins data, data prep, scaling, ggplot2 visuals, color by species, facet by island, outlier detection, and caveats

Unveiling the Power of Hilbert Curves in Clustering: (cyberwarhead​.com). Hilbert curves improve clustering by preserving locality, enabling efficient 1D mapping of high-dimensional data, with applications in data management, image segmentation, and scalable analytics

Global Modeling with GluonTS DeepAR: Future of Semiconductors in the U.S. (datageeek​.com). GluonTS DeepAR forecasts for TXN, ADI, IFNNY, 6723.T using daily data; onshore manufacturing impact under Section 232 tariffs; visualizing predictive intervals and model accuracy

MLOrbs?: MLOps in the database with orbital and dbt (emilyriederer​.com). MLOps in the analytical database using orbital's sklearn-to-sql and tidymodels, sqlglot, and dbt; churn modeling with IBM telecom data; zero-infrastructure deployment inside dbt pipelines

High-Performance Model Weight Storage and Distribution in Cloud Environments (nilesh-agarwal​.com). Examines storage solutions like NFS and FUSE for scalable model weight distribution in cloud AI deployment, comparing costs, performance, and future developments for ML workloads

Automate AIOps with Amazon SageMaker Unified Studio projects, Part 1: Solution architecture (aws​.amazon​.com). Explores multi-account AWS architecture, SageMaker Unified Studio, automation with CI/CD, project and data governance, multi-tenancy, and security strategies for scalable AI/ML operations

Automate AIOps with SageMaker Unified Studio Projects, Part 2: Technical implementation (aws​.amazon​.com). Implement AIOps with SageMaker Unified Studio, focusing on project initialization, development, and deployment involving administrators, data scientists, and ML engineers

Weeknotes: 18th August 2025 (digitalflapjack​.com). Yirgacheffe 2.0: API simplification, autoscaling to match rasters, read_shape_like vs read_shape, GeoTIFF band concepts, Tessera multi-band handling, parallelism strategies, Slurm container backends, and usability improvements for ecologists

New Tool: xstack - Completely Passive eBPF Linux Stack Profiling Without Any Tracepoints (tanelpoder​.com). Passive Linux stack profiling with xstack: eBPF task iterators, read-only kernel/user stacks, no tracepoints, PID/TID filtering, Hz sampling, iterations, and flamegraph output via Flamelens

MicroZed Chronicles: MathWorks Deep Learning Processor (adiuvoengineering​.com). Overview of implementing MathWorks Deep Learning Processor with MATLAB, Simulink, Vivado, and AXI interfaces, supporting CNNs and LSTMs on AMD FPGAs for satellite telemetry anomaly detection

AI Cluster Networking (nwktimes​.blogspot​.com). AI cluster networking covers Scale-Out Backend inter-node RDMA, Scale-Up intra-node GPU communication, Frontend user inference, Management and Storage networks, UES Ultra Ethernet Specification and UET transport for RDMA over Ethernet

Avalanche stack and real-time streaming applications at Nu (building​.nubank​.com). Nubank's Avalanche stack enables real-time analytics with Kubernetes, Kafka, Flink, and Pinot for fraud detection, Autopilot risk calibration, On-Demand Features Handler, and case studies in opportunistic loans

Marketplace: my first attempt at training without backprop on GPU efficiently (fangpenlin​.com). Marketplace-inspired GPU-efficient training without backpropagation, upstream sampling, vendor-neutral specs, Tinygrad integration, CUDA memory considerations, autoregressive weight remixing, parallel mutation

Parallelization Strategies in Neural Networks (nwktimes​.blogspot​.com). Data, model, and tensor parallelism explained; 3D parallelism enables scalable AI training across GPUs and nodes, covering FNNs, forward/backward passes, activations, DMA, RDMA, and memory considerations

Dion: the distributed orthonormal update revolution is here (microsoft​.com). Microsoft Research introduces Dion, a scalable distributed orthonormal update optimizer leveraging low-rank approximation, amortized power iteration, and error feedback to improve large-scale AI model training efficiency

Smarter training for smarter AI (cs​.jhu​.edu). Johns Hopkins researchers introduce MomSPS and USAM optimization methods to speed up deep learning training, reduce hyperparameter tuning, and improve model robustness and real-world generalization

Vandermonde Matrices are Merely Exponentially Ill-Conditioned (ethanepperly​.com). Gautschi’s bound on Vandermonde conditioning, exponential ill-conditioning, block Krylov iterations, RBKI, elementary symmetric polynomials, Lagrange vs Vandermonde, robust FTA, inverse Vandermonde entries

Derivatives, Gradients, Jacobians and Hessians – Oh My! (blog​.demofox​.org). Derivatives, gradients, Jacobians and Hessians explained: optimize with gradient descent, compute partial derivatives, build Jacobians, explore determinants, and apply in rendering and ML

New Physics-Inspired Proof Probes the Borders of Disorder (quantamagazine​.org). Band matrix thresholds for localization-delocalization transitions; Yau, Yin, Erdős, Knowles, and collaborators prove delocalization just above predicted band widths in 1D, 2D, 3D

A Bird’s-Eye View of Linear Algebra: Why Is Matrix Multiplication Like That? (towardsdatascience​.com). Explores matrix algebra, including addition, subtraction, multiplication, inverse, identity, and division, illustrating their roles in linear maps and applications in AI and data science

Who Invented Backpropagation? (people​.idsia​.ch). Historical overview of backpropagation, Linnainmaa 1970, Kelley 1960 precursor, Werbos 1982 application, Rumelhart et al. 1986, Amari 1967-68, SGD, deep learning breakthroughs, pre-training debates, 2010 plain backprop in deep nets

xRFM: Accurate, scalable, and interpretable feature learning models for tabular data (arxiv:stat). xRFM combines feature learning kernel machines with a tree structure to adapt to local data, scale to massive datasets, and deliver interpretability via Average Gradient Outer Product across tabular regression and classification tasks

INFNet: A Task-aware Information Flow Network for Large-Scale Recommendation Systems (arxiv:cs). INFNet introduces a task-aware architecture for large-scale recommendations, employing categorical, sequence, and task tokens with heterogeneous cross-attention and homogeneous Proxy Gated Units to optimize multi-task learning

Deep Learning in Classical and Quantum Physics (arxiv:cs). Deep learning for quantum science and technology: DL/ML for exploring parameter spaces, extracting patterns from experiments, guiding quantum control, materials discovery, with rigor, interpretability, and mitigation strategies

GNN-based Unified Deep Learning (arxiv:cs). Unified learning encodes diverse models (MLPs, CNNs, GNNs) into a shared graph space, using a uGNN to guide parameter decoupling, sharing, and cross-distribution generalization

Fast and Simple Multiclass Data Segmentation: An Eigendecomposition and Projection-Free Approach (arxiv:cs). Proposes a penalty reformulation and an eigendecomposition and projection-free optimization scheme for graph-based multiclass data segmentation, improving efficiency and accuracy on large-scale problems

Electromagnetic Simulations of Antennas on GPUs for Machine Learning Applications (arxiv:cs). GPU-accelerated EM simulations using gprMax for antenna design ML datasets; compares open-source and commercial solvers; evaluates ML/DL models for antenna parameter estimation; analyzes performance

UbiQTree: Uncertainty Quantification in XAI with Tree Ensembles (arxiv:cs). UbiQTree decomposes uncertainty in SHAP values using Dempster-Shafer theory, distinguishing aleatoric, epistemic, and entanglement effects to enhance interpretability in tree ensemble models

Meta-learning optimizes predictions of missing links in real-world networks (arxiv:cs). Meta-learning optimizes link prediction in diverse real-world networks using stacking, topological predictors, graph neural networks, and algorithm selection based on network characteristics

I've got a big favor to ask - keeping Blaze running isn't expensive, but it does all add up, so I'm asking readers like you to help, if you can.
That's why I'm launching a Patreon page!. Nothing flashy, just a way for folks who find value in these newsletters to chip in a little each month. In return, you'll get:

• Real say in how Blaze evolves — vote on new topics, features, topic curation ideas
• First dibs on merch (details still cooking)
• That warm fuzzy feeling knowing you're supporting something that saves you time and keeps you plugged into great tech writing

If you are getting value from blaze, checking this out would mean the world. And if you can't contribute, no worries—the newsletters keep coming either way, and you can follow along on patreon for free.
Thanks for reading and being part of this nerdy corner of the internet. All the best - Alastair.