🤖

Machine Learning Engineer: 30th September 2025

Newsletters sent once a week, unsubscribe anytime.

Published 30th September 2025

🔧 Company Engineering Blogs

How to estimate correlation between metrics from past A/B tests (booking​.ai). Combining naive proxy–goal correlations with a Total Covariance Estimator to correct measurement noise in A/B tests

Beyond Winning: Spotify’s Experiments with Learning Framework (engineering​.atspotify​.com). Spotify's Confidence platform powers Experiments with Learning (EwL) to maximize information gain from tests across hundreds of teams

Networking at the Heart of AI — @Scale: Networking 2025 Recap (engineering​.fb​.com). Summary of @Scale Networking 2025: AI-focused fabric, Prometheus/Hyperion clusters, RNGs, NICs, and end-to-end GPU-like abstraction

GitHub Copilot gets smarter at finding your code: Inside our new embedding model (github​.blog). Copilot embeds a new code/documentation embedding model to improve code search in VS Code with faster retrieval and smaller memory footprint

Accelerating Qwen3-8B Agent on Intel® Core™ Ultra with Depth-Pruned Draft Models (huggingface​.co). Speculative decoding and depth-pruned drafts accelerate Qwen3-8B on Intel Core Ultra using OpenVINO, with integration into 🤗smolagents

🛠️ Careers, Craft, and Developer Tools

My experience of interview preparation as MLE (andlukyane​.com). MLE interview prep covering Leetcode, ML system design, ML theory, statistics, and behavioral rounds with Obsidian notes and LLM-assisted practice

Associate Research Scientist – Machine Learning, Center for Computational Mathematics (Joint with with Cooper Union) (bayesian​.org). Joint CCM/Cooper Union machine learning faculty position at Flatiron Institute, with 50-50 appointment, covering ML, Bayesian stats, optimization, and software for science

How to Accelerate Community Detection in Python Using GPU-Powered Leiden (developer​.nvidia​.com). GPU-accelerated Leiden in cuGraph speeds up community detection on large graphs using NetworkX nx-cugraph backend

Release pydantic-optuna-bridge (alexdong​.com). Ships pydantic-optuna-bridge v0.1.1 to sync Optuna search spaces with a Pydantic model using optuna_config and Annotated types

I gave a talk about machine learning and math (quomodocumque​.wordpress​.com). Talk on machine learning and math, with slides, a CMSA Big Data conference session, and discussion of generating mathematical material using ML

“Kernel Ridge Regression with Cholesky Inverse Training Using C#” in Visual Studio Magazine (jamesmccaffreyblog​.com). Kernel ridge regression with Cholesky inverse training using C# and SGD vs matrix inverse approaches in Visual Studio Magazine

An Example of Coefficient of Determination Using JavaScript (jamesmccaffreyblog​.com). JavaScript kernel ridge regression with SGD, R2 metric, and synthetic data experiments using Node.js

🛡️ Security, Privacy, and Responsible AI Ops

Membership Privacy Risks in LLMs (brave​.com). CAMIAContext-Aware Membership Inference Attack analyzes token-level uncertainty to reveal LLM memorization risks

CVE-2025-23298: Getting Remote Code Execution in NVIDIA Merlin (thezdi​.com). Zero Day Initiative details CVE-2025-23298 in NVIDIA Merlin Transformers4Rec, unsafe pickle deserialization, and remote code execution risks

AI/Machine learning is amazing if … (katharinabrunner​.de). Explores where AI/ML shines (prefetching, branch prediction, spam filters) and warns about harms in translation, summarisation, coding, and liability

Machine Learning for Safety-Critical Applications Opportunities, Challenges, and a Research Agenda (nap​.edu). Research priorities and safety engineering practices for integrating machine learning into safety-critical systems like autonomous vehicles and surgical robots

A Modern Approach to Engineering (differentshelf​.com). Modern engineering practices, trunk-based monorepos, chaos engineering, inner sourcing, and data as a product drive faster, safer software at scale

⚡ Inference, Retrieval, and Acceleration

A Paradigm Shift: Reasoning at Enteprise Scale (nuit-blanche​.blogspot​.com). Reasoning-first retrieval stack for enterprise-scale documents using late-interaction models, PyLate, ModernBERT, FastPlaid, and PyLate-rs in browser-ready RAG workflows

Networking at the Heart of AI — @Scale: Networking 2025 Recap (engineering​.fb​.com). Summary of @Scale Networking 2025: AI-focused fabric, Prometheus/Hyperion clusters, RNGs, NICs, and end-to-end GPU-like abstraction

Generate a Video Knowledge Graph: NVIDIA VSS Blueprint with GraphRAG on ArangoDB (arangodb​.com). NVIDIA VSS Blueprint for video search and summarization integrates GraphRAG on ArangoDB to generate a video knowledge graph with multi-stream ingestion and hybrid retrieval

Together with SGLang: Best Practices for Serving DeepSeek-R1 on H20-96G (lmsys​.org). Optimizing DeepSeek-R1 serving on H20-96G with SGLang: load balancing, FP8 attention on Hopper, SBO, DeepXTrace, and tiered online inference

Tech dive: Comprehensive compression leveraging quantization & dimensionality reduction (redis​.io). Redis Query Engine now supports Quantization and Dimensionality Reduction using LVQ and LeanVec with SVS-VAMANA for memory-efficient vector search

Day-0 Support for the SGLang-Native RL Framework - slime on AMD Instinct™ GPUs (rocm​.blogs​.amd​.com). Day-0 ROCm support for slime on AMD Instinct GPUs, detailing kernel/memory optimizations, Docker images, APRIL rollouts, and synchronous/asynchronous RL frameworks

Accelerating Qwen3-8B Agent on Intel® Core™ Ultra with Depth-Pruned Draft Models (huggingface​.co). Speculative decoding and depth-pruned drafts accelerate Qwen3-8B on Intel Core Ultra using OpenVINO, with integration into 🤗smolagents

How to GPU-Accelerate Model Training with CUDA-X Data Science (developer​.nvidia​.com). GPU-accelerated training with cuML, cuML FIL, XGBoost, LightGBM, CatBoost, Treelite, and feature-importance techniques in manufacturing data

📈 Product ML: Experimentation, Recsys, and Applied Modeling

Beyond Winning: Spotify’s Experiments with Learning Framework (engineering​.atspotify​.com). Spotify's Confidence platform powers Experiments with Learning (EwL) to maximize information gain from tests across hundreds of teams

Slop Machines (notes​.hella​.cheap). Overview of bandit problems in feeds, exploration vs exploitation, and collaborative filtering for content recommendation

How to think about ML Evaluation? (lepisma​.xyz). How to think about ML evaluation: focus on product problems, metric graphs, approximate metrics, and production debugging over dashboards

User foundation models for Grab (engineering​.grab​.com). Grab builds a specialized user foundation model unifying tabular and time-series data with modality adapters and hierarchical ID classification for embeddings

Build a scalable feature engineering pipeline with polars (fullstory​.com). Scalable feature engineering for behavioral data using Polars, with dbt groundwork, feature categories, and pipeline design

When in doubt, batch requests (janmeppe​.com). Batch requests reduce network hops and optimize inference by batching similar-sized inputs for faster per-request latency

Global Modeling with XGBoost: Gold vs. Silver (datageeek​.com). XGBoost global modeling of gold and silver futures with tidymodels, modeltime, and time-series splits

🧩 Model Architectures: Vision, Transformers, and MoE

DINOv3 Paper Explained: The Computer Vision Foundation Model (aipapersacademy​.com). DINOv3 scales to 7B params on 1.7B images, introduces Gram Anchoring to preserve dense feature consistency for vision tasks

Expressing an Indicator in Neural Net form, Part 2. (dekalogblog​.blogspot​.com). Experimenting with a neural-net indicator: sparse 40x20 input, 4 outputs, 16 trainable weights, tanh and sigmoid activations, and mixed cross-entropy with Sortino loss

Applied introduction to Categorical treatment of CuTe (veitner​.bearblog​.dev). Explores categorical treatment of CuTe Layouts using morphisms, composition, projections, expansions, and coalesced maps with Python CuTe tooling

Mixture-of-Experts (MoE) Explained: How Trillion-Parameter AI Models Actually Work (omps​.in). MoE routing to 1T parameters with 32B active per forward pass; context length, MuonClip optimizer, fp8 quantization, and token-based pricing examples

supplement for supplement for 0.412 (aarnphm​.xyz). Decoder-only transformers dissected: embeddings, attention, residuals, gating, MOE, and Pareto-sized design considerations

tools for 0.412 (aarnphm​.xyz). Helper scripts and latent projection tooling for 0.412 lecture: WOV, Qwen/Qwen3-0.6B, latent projection.py, and heatmaps

Video models are zero-shot learners and reasoners (simonwillison​.net). Video models like Veo 3 may become unifying foundation models for vision, enabling zero-shot learning and chain-of-frames reasoning

🌍 Scientific ML, Time Series, and Spatial Modeling

Effort.jl : Des simulations cosmiques sur laptop grâce à Julia (sciencedaily​.com). Effort.jl emulator speeds cosmological EFTofLSS simulations on laptops using neural networks and physics-informed gradients

Science in the age of foundation models (amazon​.science). Foundation models in science: probabilistic time series, spatiotemporal forecasting, physics constraints, uncertainty quantification, and Chronos/T-SFM integration

A Universal Model of Urban Street Networks (geoffboeing​.com). Marc Barthelemy and Geoff Boeing propose a two-step generative model starting with a minimum spanning tree and adding edges to match degree distributions

Time series foundation models can be few-shot learners (research​.google). TimesFM-ICF uses continued pre-training with in-context examples and a common separator token to enable few-shot time-series forecasting

Open Ocean #5 (diagrammonkey​.wordpress​.com). Gaussian process interpolation on a globe using lat/long converted to 3D, Matern covariance, and zonal stretching to fill oceans

Open Ocean #6 (diagrammonkey​.wordpress​.com). Bayesian PCA for infilling SST data in Open Ocean #6; pattern-based reconstruction with iterative EOF-like patterns

New tensor network-based approach could advance simulation of quantum many-body systems (phys​.org). New tensor network approach maps symmetric 1D Hamiltonians to dual symmetry-breaking models for efficient ground states

Can a model trained on satellite data really find brambles on the ground? (toao​.com). Field validation of a TESSERA-embedded ABM model predicting bramble hotspots using Sentinel data, iNaturalist inputs, and logistic/knn classifiers

🧠 Theory of ML and Optimization

Beyond Power Laws: Scaling Laws for Next-Token Prediction (francisbach​.com). Scaling laws for linear bigram next-token prediction under Zipf’s law; normalization, time rescaling, and sign descent vs gradient descent

Diffusion Beats Autoregressive in Data-Constrained Settings (blog​.ml​.cmu​.edu). Data-constrained diffusion surpasses autoregressive models when training data is limited but compute is abundant

The QMA Singularity (scottaaronson​.blog). Scott Aaronson and Freek Witteveen study black-box amplification limits in QMA with a quantum oracle separation and AI-assisted proof ideas

To Understand AI, Watch How It Evolves (quantamagazine​.org). Interpretability through training dynamics and evolutionary views; stochastic gradient descent, random initialization, vestigial structures, causality, and training-time variations

the Harvard and Brown school of computer science (xianblog​.wordpress​.com). Harvard-Brown school contrasted with LeCun's neural nets, using Bayesian inference and Markov random fields for pattern learning

Making sense of parameter-space decomposition (lesswrong​.com). Intuitive overview of parameter-space decomposition (SPD) for LLMs, rank-1 subcomponents, ablation, stochastic reconstruction, and potential uses

Modular Manifolds (thinkingmachines​.ai). Co-designs neural net optimizers with manifold constraints on Stiefel, hypersphere, spectral norm; introduces modular manifolds and dual ascent for Muon-style updates

📚 Academic Research

TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees (arxiv:cs). TABFAIRGDT uses autoregressive decision trees and soft leaf resampling to generate fair, high-quality tabular data on CPU

Enhancing Credit Default Prediction Using Boruta Feature Selection and DBSCAN Algorithm with Different Resampling Techniques (arxiv:stat). Boruta feature selection and DBSCAN with SMOTE-Tomek/exact resampling improve GBM-based credit default prediction

Enhancing Credit Risk Prediction: A Meta-Learning Framework Integrating Baseline Models, LASSO, and ECOC for Superior Accuracy (arxiv:cs). Meta-learning framework combines XGBoost, RF, SVM, DT, KNN, MLP, LASSO, and ECOC for credit risk classification and default probability on Corporate Credit Ratings data

Intelligent Algorithm Selection for Recommender Systems: Meta-Learning via in-depth algorithm feature engineering (arxiv:cs). Engineering algorithm features with static code metrics and performance landmarks improves meta-learning for recommender systems and top-k selection

Rethinking player evaluation in sports: Goals above expectation and beyond (arxiv:stat). Residualized metrics for goals above expectation using double machine learning to infer player-specific effects in soccer and beyond

👋 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!