Machine Learning Engineer: 23rd September 2025
🔧 Company Engineering Blogs
Defending 20 Trillion Transactions: How Hyperforce’s Trusted Perimeter Stops DDoS Attacks with AI (engineering.salesforce.com). Explains Trusted Perimeter for Hyperforce: TLS termination, CDN caching, DDoS protection, AI-driven threat detection, and 10-minute global rollback
Viaduct, Five Years On: Modernizing the Data-Oriented Service Mesh (medium.com/airbnb-engineering). Viaduct Modern overhauls engine and API, open-sources OSS at Airbnb, introducing tenant modules, resolvers, and dual APIs
🛠️ Applied ML Case Studies & Niche Models
Ranking the Perceived Safety of Streets (saadiqm.com). Crowd-sourced street-view comparisons, Place Pulse 2.0, RankNet, DINOv3 embeddings, and a NA cities filter to rank perceived street safety
The Deceptively Complex World of Turkish Diacritics: A Neural Network Journey (ergoso.me). Turkish diacritics restoration with a small neural model: BiLSTM, attention, masking, balanced loss, 6-letter scope, and GPU scaling
Building a Grocery Classification Model (dgendill.com). Building a grocery classification model using 1–3-gram vocabularies, labeled data, and a RandomForest/OneVsRest multi-label classifier
🏈 Revamping Win Probability for 2025 (blog.collegefootballdata.com). New Python/XGBoost win-probability models with regulation, clutch, and overtime blending plus a Win Probability Calculator
Random Mutation Hill Climbing (RMHC) (4rknova.com). Evolving images with Random Mutation Hill Climbing (RMHC) in the browser using Canvas 2D, MSE fitness, and shape-based DNA
What I Learned Building Synthetic Genomic Data: When Statistical Models Hit Their Limits (ksred.com). ML-enhanced synthetic cfDNA data yields 92% fetal accuracy; statistical models plateau at ~77% due to fragment interdependencies
⚙️ ML Systems & Inference Engineering
Migrating to Nvidia Triton: High-Throughput, Low-Cost Inference at Scale (medium.com/gumgum-tech). Migrating to Nvidia Triton for high-throughput, cost-efficient NLP inference with Python backends, dynamic batching, and GPU utilization tuning
Qwen-8B Embeddings: Near-SOTA Performance at 600x the Speed (alexdong.com). Qwen-8B embeddings enable near-SOTA text classification, 600x faster than LLM classifiers, achieving MAP ~0.944 on Kaggle with simple MLP
Running SOTA AI-based Weather Forecasting models on AMD Instinct (rocm.blogs.amd.com). Step-by-step guide to running SOTA AI weather forecasting models on AMD Instinct MI300X using JAX, PyTorch, Docker images, ai-models, and ECMWF data
Paths to Support Additional Numeric Types on the Java Platform #JVMLS (inside.java). JVMLS discusses enabling full support for new numeric types in Java, including 16-bit floats and 8-bit formats for ML, with trade-offs across paths
Anthropic: A postmortem of three recent issues (simonwillison.net). Anthropic explains three concurrent infrastructure bugs across AWS Trainium, NVIDIA GPUs, and Google TPUs that degraded Claude’s responses, plus privacy constraints hindering investigation
Announcing AWS Neuron SDK 2.26.0 (aws.amazon.com). GA of Neuron SDK 2.26.0 brings PyTorch 2.8, JAX 0.6.2, FLUX.1-dev image generation, and MoE expert parallelism on Trainium2
How do vector databases work? (hclimente.github.io). Vector embeddings, cosine similarity, UMAP visualizations, and HNSW-based vector databases (Qdrant) for RAG with LLMs
🚀 Training Paradigms, Efficiency, and Model Behavior
What gpt-oss Leaks (fi-le.net). Analysis of GPT-oss; GPT-5 training data leakage, glitch tokens, tokenizers, embeddings, and membership inference across OpenAI models
The Shift to Reinforcement Learning Greatly Reduces Learning-Efficiency (tobyord.com). RL training learns far less per hour than pre-training, impacting scalability, generality, and frontier task efficiency in AI systems
The Extreme Inefficiency of RL for Frontier Models (tobyord.com). New scaling paradigm: RL’s information efficiency vs pre-training; long-horizon tasks, token-entropy, METR/HCAST, o1/o3/o3 models, latency and inference costs
Surprising GPT-OSS 20B 2-bit Quantization Performance (alexdong.com). Benchmarking GPT-OSS 20B 2-bit, 5-bit quantization (GGUF), and long-context effects on RTX 3000 with MAP@3 insights
This paper changed my life: Dan Goodman on a paper that reignited the field of spiking neural networks (thetransmitter.org). Friedemann Zenke’s 2019 paper on surrogate gradients enables modern ML with spiking neural networks and SpyTorch tutorials
Giving Foundation Models a Notion of Now (building.nubank.com). Explores encoding transaction timestamps as time deltas to improve transformer model performance and generalization with time-delta tokens and bucketed granularity
🧪 Classic ML, Explainability, and Statistical Modeling
Detecting and Fixing Data Quality Problems in Data Pipelines and ML Systems (wasi0013.com). Detects, validates, and remediates data quality issues in web-scraped pipelines using Python, Pandas, and Imputation techniques
2025-09-17: Classic Machine Learning Models and XAI Methods (ws-dl.blogspot.com). Overview of classic ML models (Naive Bayes, Random Forest, SVM) and model-agnostic XAI methods (SHAP, LIME, permutation importance) with logistic regression discussion
Key improvements in shapviz and kernelshap (lorentzen.ch). Shapviz and kernelshap updates with GLM and XGBoost SHAP explanations and interactions
Generating Synthetic Data with R-vine Copulas using esgtoolkit in R (thierrymoudiki.github.io). Tutorial on generating synthetic data with R-vine copulas using esgtoolkit in R and RVineModel fitting
Graph Regularization (blog.devgenius.io). Graph regularization in semi-supervised learning using Laplacian, manifold regularization, NSL, and Planetoid with TensorFlow
🌊 Scientific ML for Physics & Simulation
Introducing P3D: The three-dimensional PDE-Transformer (ge.in.tum.de). P3D: a scalable 3D PDE-Transformer for high-resolution physics surrogates and diffusion-ready probabilistic 3D turbulence modeling
Discovering new solutions to century-old problems in fluid dynamics (deepmind.google). New AI-assisted method using Physics-Informed Neural Networks to discover unstable singularities in IPM and Boussinesq fluid equations
New AI Technique Unravels Quantum Atomic Vibrations in Materials (caltech.edu). Caltech AI method speeds up phonon interaction calculations for thermal transport by compressing high-order tensors via CANDECOMP/PARAFAC decomposition
Autodesk Research Brings Warp Speed to Computational Fluid Dynamics on NVIDIA GH200 (developer.nvidia.com). Autodesk Research leverages NVIDIA Warp and GH200 to speed Python-based CFD with XLB, achieving ~8x speedup and scaling to 50B cells
New approach improves accuracy of quantum chemistry simulations using machine learning (phys.org). Machine learning inverts many-body results to learn a universal exchange-correlation functional for DFT accuracy
📐 Math Foundations for ML: Linear Algebra, Sampling, and Structure
Think Linear Algebra (allendowney.com). Think Linear Algebra explores code-first approaches to linear algebra, with chapters on PageRank, 2D graphics transforms, LU decomposition, null space, and truss analysis
Global epistasis emerges from a generic model of a complex trait (or: random walks on a hypercube with reweighting) (quomodocumque.wordpress.com). A simple model of evolution using 2^n genotypes on a hypercube with fitness polynomial and reweighting dynamics
Polynomial Bounds for Chowla’s Cosine Problem (gilkalai.wordpress.com). Polynomial bounds for Chowla’s cosine problem via spectral graph theory and Hadamard products
Rayleigh quotient (aarnphm.xyz). Rayleigh quotient for Hermitian matrices, eigenvalue bounds, eigenvectors, and Lagrangian formulation with extensions to covariance, generalized and two-sided forms
Latent trees (languagelog.ldc.upenn.edu). Syntactic structure in LMs, historical statistical approaches, Harris/Firth, word segmentation, latent trees, and critiques by Futrell, Liu, and Beguš
I Misunderstood Rejection Sampling All This Time (buttondown.com/jaffray). Rejection sampling expanded beyond disks: image distributions, volume sampling, grid subdivision, and efficiency improvements
📚 Academic Research
APFEx: Adaptive Pareto Front Explorer for Intersectional Fairness (arxiv:cs). Adaptive Pareto Front Explorer (APFEx) optimizes intersectional fairness across sensitive attribute combinations with differentiable metrics and Pareto-aware gradient strategies
Personalized Federated Learning with Heat-Kernel Enhanced Tensorized Multi-View Clustering (arxiv:cs). Personalized federated learning using heat-kernel enhanced tensorized multi-view fuzzy c-means with Tucker/CP decompositions and DP-enabled global aggregation
Green Recommender Systems: Understanding and Minimizing the Carbon Footprint of AI-Powered Personalization (arxiv:cs). Recommender systems eco-footprint: comparing traditional models and deep learning, measuring energy use and CO2 via hardware meters
Benefits of Online Tilted Empirical Risk Minimization: A Case Study of Outlier Detection and Robust Regression (arxiv:stat). Online Tilted Empirical Risk Minimization (TERM) for streaming data; balances accuracy, fairness, and robustness via a per-sample tilt parameter t
Cost-Performance Analysis: A Comparative Study of CPU-Based Serverless and GPU-Based Training Architectures (arxiv:cs). Comparative cost-performance analysis of SPIRT, ScatterReduce, AllReduce, and MLLess for serverless and GPU-based ML training with RedisAI
👋 Before you go
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.
You may also like
About Machine Learning Engineer
Our Machine Learning Engineer newsletter covers the latest developments, research papers, tools, and techniques in ML engineering and deployment. Each week, we curate the most important content so you don't have to spend hours searching.
Whether you're a beginner or expert in machine learning engineering, our newsletter provides valuable information to keep you informed and ahead of the curve in this technically challenging field.
Subscribe now to join thousands of professionals who receive our weekly updates!