Machine Learning Engineer: 12th August 2025

Genie 3: A new frontier for world models (deepmind​.google). Genie 3 is a groundbreaking world model that generates diverse interactive environments, advancing AI capabilities in simulations and engaging with the real world

Vision Language Model Alignment in TRL ⚡️ (huggingface​.co). Introduction of Mixed Preference Optimization, Group Relative Policy Optimization, and Group Sequence Policy Optimization for enhancing Vision Language Models alignment

The Interspeech 2025 Speech Accessibility Project Challenge (machinelearning​.apple​.com). Interspeech 2025 SAP Challenge highlights advancements in ASR for speech disabilities, leveraging 400+ hours of data, evaluating teams on WER and SemScore

Achieving 10,000x training data reduction with high-fidelity labels (research​.google). Google researchers develop a novel active learning method achieving 10,000x data reduction for fine-tuning LLMs while enhancing model alignment with human experts

A better path to pruning large language models (amazon​.science). Prune Gently, Taste Often: Wanda++ scans decoder blocks post-training, calibrating weights on small data to preserve performance while pruning efficiently on a single GPU runtime

AI: great expectations (rodneybrooks​.com). Examines AI hype cycles from GIANT BRAINS to expert systems and neural networks, highlighting Berkeley, Widrow, ADALINE, MADALINE, Amara’s Law, and lessons for manufacturing today

RNLA 2025 (mathsci​.ai). RNLA 2025 convenes IPAM workshop on Randomized Numerical Linear Algebra (Aug 11–15, 2025) featuring groups, travel/housing assistance; applications due Mar 31; led with Riley Murray

507: Turn Our Data Into Predators (embedded​.fm). Chris and Elecia discuss data-driven science, ultrasonic recorders, engineering AI applications, and resources like Data-Driven Science and Engineering and the Datasaurus Dozen

Things I Wish I Had Known Before Starting ML (towardsdatascience​.com). Explore crucial insights on machine learning, including flexible boundaries, the difference between research and production code, and the importance of deep reading

Multi-Dimensional Vector Support in CocoIndex (cocoindex​.io). CocoIndex adds custom targets and multi-dimensional vector support, enabling multi-vector embeddings, patch-based image processing, MaxSim retrieval, Qdrant integration, and typed vector workflows in Python today

Lecture 9: Introduction to Monitoring (medium​.com/marvelous-mlops). Explore Databricks' ML monitoring tools focusing on data and model drift, emphasizing statistical health and performance tracking for machine learning systems

Accelerate ND-Parallel: A Guide to Efficient Multi-GPU Training (huggingface​.co). Efficient multi-GPU training with Accelerate and Axolotl: strategies include Data, Fully Sharded, Tensor, and Context Parallelism for large models

Ask HN: How can ChatGPT serve 700M users when I can't run one GPT-4 locally? (news​.ycombinator​.com). How ChatGPT serves 700M users: inference at scale with GPUs, clusters, sharding, RPC, load balancing, model optimization, MoE, quantization, JAX scaling book, unsloth guides resources

Sim2Real Last Steps (irvin​.quest). Exploration of sim2real challenges using deep reinforcement learning and vision-based tasks, leading to a shift towards world modeling for robotics

What I Learned About Machine Learning – Don’t Use It! (bobbydurrettdba​.com). Bobby Durrett critiques machine learning for Oracle database monitoring, detailing attempts with autoencoders, binary classification, and z-score methods, emphasizing data visualization

Visual Anomaly Detection: Turning HTTP Requests into Bitmaps for Machine Learning (russell​.ballestrini​.net). Transform HTTP logs into grayscale bitmaps, extract features with OpenCV, PyImageSearch-inspired Isolation Forest, train on 1k samples, and compare anomalies by user agents and requests

The spatial join? (spatialists​.ch). Vikram Gundeti redefines geospatial intelligence for data scientists, simplifying access with H3 cells, eliminating traditional GIS barriers, and embracing ML workflows

Please Hold the Bacon: Review of the bacon R Package (replicationindex​.com). Review of the bacon R package, a mixture model correcting bias in z-scores for genomics datasets, highlighting its limitations and applications in high-throughput studies

Stellar Flare Detection and Prediction Using Clustering and Machine Learning (towardsdatascience​.com). Utilizing DBSCAN and XGBoost for detecting and predicting stellar flares, analyzing time-series data from NASA's TESS, enhancing understanding of stellar behavior

Fei Wan (2025) on propensity score matching (andifugard​.info). Fei Wan (2025) revisits King and Nielsen’s propensity-score matching critique, endorsing machine learning estimated scores over logistic regression, while discussing inverse-probability weighting and quasi-experimental methods

Writing Word2Vec from scratch in Rust (lucas-montes​.com). Implementing Word2Vec in Rust using CBOW architecture for relationship mapping in notes, focusing on efficiency, parallelization, and core algorithm steps

Word Embeddings: Theory and Analysis (blog​.sparsh​.dev). Overview of word embeddings, vocabulary discretization, and dense representations; highlights Word2Vec and GloVe, semantic similarity via cosine similarity, analogy examples, subword n-grams, and embedding dimensionality

Neurosymbolic AI: The 3rd Wave (muratbuffalo​.blogspot​.com). Neurosymbolic AI integrates learning and reasoning, utilizing Logic Tensor Networks to enhance interpretability and modularity for robust AI systems

modded-nanogpt: Analyzing value-embedding-, UNet-, and x0-lambdas (snimu​.github​.io). modded-nanogpt analyzes value-embedding-, UNet-, and x0-lambdas, detailing three residual-mixing tricks, learned lambda dynamics, layer skipping, training effects, and links to learning-rate and sequence-length schedules, patterns

Exploring fun parts of Neural Network (shivasurya​.me). Explores neural networks from XOR basics in NumPy to sigmoid versus ReLU, training dynamics, 3Blue1Brown insights, MNIST hints, and implications for security reviews and LLMs

Deep linear networks (danmackinlay​.name). Exploration of deep linear networks, gradient flow, singular value dynamics, and gated models with a focus on feature learning and hierarchical structures

Generalization Gap in Over‑Parameterized Models (gojiberries​.io). Explores generalization gap in over-parameterized models, focusing on concepts like double descent, sampling error, under-optimization, and implicit bias in machine learning

The challenge of defining a neural population (thetransmitter​.org). Proposes dynamical boundaries for neural populations, highlighting subspace communication, and null space concepts; measurement scales (electrodes, calcium imaging, fMRI), region independence, and Mark Humphries' perspective

A ML Model is a Decent First-Order Approximation of a Human Learner (justinmath​.com). Machine learning models are similar to human learners in their incremental updates and dependency on feedback and task pre-training

Using geometry and physics to explain feature learning in deep neural networks (phys​.org). Spring-block phenomenology models feature learning in deep neural networks, linking data separation across layers to friction, noise, and training dynamics; revealing relations akin to thermodynamics

Supervised Machine Learning Methods with Uncertainty Quantification for Exoplanet Atmospheric Retrievals from Transmission Spectroscopy (arxiv:astro). Comparative study of ML regression methods for exoplanet atmospheric retrievals using transmission spectroscopy, assessing accuracy, speed, and uncertainty quantification

Algorithm Selection for Recommender Systems via Meta-Learning on Algorithm Characteristics (arxiv:cs). Per-user meta-learning for recommender systems improves NDCG@10 by 8.83% using user meta-features and algorithm characteristics from source code

X-VFL: A New Vertical Federated Learning Framework with Cross Completion and Decision Subspace Alignment (arxiv:cs). X-VFL introduces Cross Completion and Decision Subspace Alignment to handle non-aligned VFL data with missing features, enabling locally independent inference and on CIFAR-10 and MIMIC-III

Efficient Multimodal Streaming Recommendation via Expandable Side Mixture-of-Experts (arxiv:cs). Expandable Side Mixture-of-Experts (XSMoE) for streaming recommendations attaches expert modules to frozen multimodal encoders, enabling gated routing and pruning to adapt visual and textual preferences

Scaling DRL for Decision Making: A Survey on Data, Network, and Training Budget Strategies (arxiv:cs). Explores scaling strategies in deep reinforcement learning, addressing data efficiency, network architecture, and training budget to enhance decision-making performance

eSASRec: Enhancing Transformer-based Recommendations in a Modular Fashion (arxiv:cs). Modularly enhances SASRec with LiGR Transformer layers and Sampled Softmax Loss, benchmarking additive improvements; identifies eSASRec as strong production-ready baseline with open-source implementation across datasets

SLA-MORL: SLA-Aware Multi-Objective Reinforcement Learning for HPC Resource Optimization (arxiv:cs). SLA-MORL optimizes dynamic resource allocation for ML in cloud environments, balancing training time, costs, and SLA compliance using multi-objective reinforcement learning

Stacked Hybrid RNN-CNN Reconstruction of X-ray Influence on 21-cm Brightness Temperature (arxiv:astro). Stacked hybrid LSTM-GRU-CNN emulator reconstructs X-ray flux effects on global 21-cm brightness during the EoR, combining CNN with LSTM and GRU for accurate parameter inference

PSEO: Optimizing Post-hoc Stacking Ensemble Through Hyperparameter Tuning (arxiv:cs). PSEO optimizes post-hoc stacking ensembles through hyperparameter tuning, achieving superior predictive performance in Automated Machine Learning on 80 public datasets

A Scalable Pretraining Framework for Link Prediction with Efficient Adaptation (arxiv:cs). A pretraining framework for link prediction using Mixture-of-Experts, combining node and edge information, achieving state-of-the-art performance with low computational costs

HiD-VAE: Interpretable Generative Recommendation via Hierarchical and Disentangled Semantic IDs (arxiv:cs). HiD-VAE proposes a hierarchical and disentangled approach for generative recommendations, addressing semantic ID issues and enhancing interpretability and diversity in recommender systems

Decorrelated feature importance from local sample weighting (arxiv:cs). Introduces local sample weighting (losaw) to decorrelate features from others, improving feature importance under correlation; applicable to random forests, CNNs, and neural networks, with tradeoffs

Advanced Multi-Architecture Deep Learning Framework for BIRADS-Based Mammographic Image Retrieval: Comprehensive Performance Analysis with Super-Ensemble Optimization (arxiv:cs). BIRADS mammographic image retrieval using DenseNet121, ResNet50, VGG16 with metric learning, super-ensemble optimization, significant precision improvements, and statistical validation

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.