🔧 Company Engineering Blogs

Using AI to perceive the universe in greater depth (deepmind​.google) . Deep Loop Shaping uses reinforcement learning in frequency-domain rewards to reduce control noise in LIGO’s mirror systems, improving gravitational-wave measurement

A New Ranking Framework for Better Notification Quality on Instagram (engineering​.fb​.com) . Diversity-aware notification ranking using multiplicative demotion, MM R-based similarity across content, author, type, and product surface, with adjustable weights and potential for LLM integration

Building Sustainable Enterprise AI Adoption: Cultural Strategies That Achieved 95% Developer Engagement (engineering​.salesforce​.com) . Salesforce shares how to scale AI adoption beyond code generation, tackling monolithic codebases, modular loading, and enterprise-wide cultural change

Welcome EmbeddingGemma, Google's new efficient embedding model (huggingface​.co) . EmbeddingGemma: Google's 308M multilingual on-device text embeddings, MMTEB/MMTEB v2 benchmarks, MRl truncation, 2K context, on‑device RAG, Sentence Transformers, LangChain, LlamaIndex, Haystack, txtai, TEI, ONNX, FAISS

Building Slack’s Anomaly Event Response (slack​.engineering) . Slack's Anomaly Event Response (AER): real-time detection, adaptive thresholds, session termination, audit logs, multi-tier architecture, detection engine, decision framework, response orchestrator, notifications, enterprise security posture

🧭 Careers, governance, and community discourse

Math Academy, update 2: I completed Mathematics for Machine Learning (frankhecker​.com) . Math Academy update: completed Mathematics for Machine Learning; daily XP goals, sequencing, eigenvectors, linear algebra, multivariable calculus, PCA, and reflections on motivation

Computer vision papers on CEUR-WS (ceurws​.wordpress​.com) . Overview of CEUR-WS computer vision proceedings, including Bildverarbeitung für die Medizin volumes 1996–2011 and related workshops

Multiple Postdoc Positions — Bayesian Multimodal Fusion (Imaging, Omics), Probabilistic Causal Discovery, Foundation Models, and Reinforcement Learning (bayesian​.org) . Postdoctoral fellowships in Bayesian multimodal fusion, causal discovery, foundation models, and reinforcement learning at UT Southwestern with MRI, PET, omics, EHR data

Speaking at posit::conf 2025 (tshafer​.com) . In Atlanta, Tom Shafer discusses R development practices that bolster model governance post-deployment for MLOps using packaging, tests, S3 methods, and modular code

How Machine Unlearning Revolutionizes AI Integrity (medium​.datadriveninvestor​.com) . Techniques for erasing bias in AI: influence functions, gradient reversal, layer-specific unlearning, and counterfactual datasets

I Spoke at Wordcamp U S in 2025 (elijahpotter​.dev) . First WordCamp talk; machine learning for quality in apps; critique of big capital expenditures; Markov Chains as language models

I’m seeing more and more companies referring to their tech stack as using “traditional machine learning” …presumably to distance themselves from the slopaganda of “AI” grifters before the bubble pops. (adactio​.com) . Traditional machine learning referenced amid AI hype; notes on tech stack framing, critique of AI grifters, and industry discourse

🛰️ Applied ML in the wild: signals, grids, retail, malware, weather

Fast ML for Funky Effects (buchanan​.one) . Domain-inspired ML for a transient detector in a guitar pedal using envelopes, sigmoid, biquad IIR filters, differential evolution

Protecting the grid with artificial intelligence (newsreleases​.sandia​.gov) . Sandia uses brain-inspired AI autoencoders on single-board computers to detect cyber-physical grid disturbances

Simplifying book discovery with ML-powered visual autocomplete suggestions (amazon​.science) . Audible's DeepPLTR and Amazon's two-stage models enable visual autocomplete with covers, real-time personalization, and cross-entity links

EMBER2024: Advancing the Training of Cybersecurity ML Models Against Evasive Malware (crowdstrike​.com) . EMBER2024 updates EMBER with 3.2M files across 6 formats, rich features, challenge set, and open-source tooling for ML malware detection

It’s the Humidity: How International Researchers in Poland, Deep Learning and NVIDIA GPUs Could Change the Forecast (blogs​.nvidia​.com) . Polish UPWr team uses SRGANs on GNSS-derived humidity data with Grad-CAM/SHAP for explainability on NVIDIA GPUs

⚙️ Training infrastructure and inference acceleration

Launch of Polars Cloud and Distributed Polars (pola​.rs) . Polars Cloud GA on AWS and Open Beta distributed engine with vertical/diagonal scaling for remote queries

Distributed Training with LanceDB and Tigris (tigrisdata​.com) . Streaming large multimodal datasets from Tigris object storage into PyTorch with LanceDB for distributed training and caching

Step-3 Deployment Simplified: A Day 0 Developer’s Guide on AMD Instinct™ GPUs (rocm​.blogs​.amd​.com) . Step-3 deployment on AMD Instinct GPUs using SGLang, Triton, and ROCm to reduce decoding costs for a 321B VLM with MFA and AFD

How Baseten achieves 225% better cost-performance for AI inference (and you can too) (cloud​.google​.com) . Baseten uses Google Cloud A4 VMs (NVIDIA Blackwell) and Dynamic Workload Scheduler to boost high-throughput inference by 225% and latency-sensitive by 25%

Announcing the new cluster creation experience for Amazon SageMaker HyperPod (aws​.amazon​.com) . One-click SageMaker HyperPod cluster creation with Quick and Custom setups, EKS/Slurm orchestration, VPCs, FSx Lustre, and CloudFormation IaC

Train and deploy models on Amazon SageMaker HyperPod using the new HyperPod CLI and SDK (aws​.amazon​.com) . SageMaker HyperPod CLI/SDK enable distributed training with FSDP, PyTorchJob CRs, Kubernetes operators, and JumpStart deployment on HyperPod clusters

Speeding up PyTorch inference on Apple devices with AI-generated Metal kernels (gimletlabs​.ai) . AI-generated Metal kernels accelerate PyTorch inference on Apple devices by up to 1.87x across 215 modules using frontier models and kernel fusion

🔡 Embeddings, similarity, and vector search

In-browser semantic search with EmbeddingGemma (glaforge​.dev) . In-browser semantic search using EmbeddingGemma and Transformers.js for client-side RAG, with a 308M-parameter model on edge devices

How big are our embeddings now and why? (newsletter​.vickiboykis​.com) . Trends in embedding sizes from 300 to 1536+; BERT 768 baseline; GPT-3/2/CLIP; HuggingFace; OpenAI matryoshka; vector databases; MTEB benchmarks

From Embeddings to Confidence Scores: Converting Similarity to Percentages (sefiks​.com) . Converts embedding distances to percentage confidence via logistic regression using DeepFace and cosine/Euclidean metrics

MUVERA: Making Multivectors More Performant (qdrant​.tech) . MUVERA embeddings compress multi-vector retrieval into a single vector for fast initial search and reranking with multi-vector representations

Balancing Relevance and Diversity with MMR Search (qdrant​.tech) . MMR search in Qdrant for fashion discovery using DeepFashion, CLIP embeddings, Python code, and metadata filtering

Building Smarter Agents: How Vector Search Drives Semantic Intelligence (couchbase​.com) . Vector search, embeddings, FTS, Eventing, hybrid search, PCAP analysis case study, embeddings API, anomaly detection, Capella, N1QL, ML embeddings, OpenAI, vector index

Introducing EmbeddingGemma (simonwillison​.net) . EmbeddingGemma: 308M multilingual embeddings, Gemma 3, quantization under Gemma license, Google model access via sentence-transformers, llama.cpp, MLX, Ollama, LMStudio, and in-browser Transformers.js demo

🧱 Transformers: from-scratch builds and forward-only training

Thinking aloud: Can we speed up model training by using binary weights? (kevinmartinjose​.com) . Explores binary weights, XNOR+popcount speedups, and limitations for training transformers on single GPUs

Marketplace V2 is all you need: A training algorithm on par with backprop that needs only forward pass (fangpenlin​.com) . Marketplace V2 trains like backprop using forward passes, seed-based randomness, reconciled delta, and SGD-style updates on MNIST-like models

Understanding and Implementing Qwen3 From Scratch (sebastianraschka​.com) . Hands-on Qwen3 from scratch in PyTorch: architecture, components, and building blocks for open-weight models

Understanding Transformers Using a Minimal Example (rti​.github​.io) . Minimal Transformer visualization: decoder-only model, 2 layers, 2 attention heads, 20-d embeddings, MIT-licensed dataset and visualization of attention

📐 Mathematics and theory: PCA, conformal prediction, manifolds, dynamics

When Machines that Simulate Intelligence Seemed Like a Summer Project (tensorlabbet​.com) . Explores Dartmouth 1956 proposal, seven themes, and how early AI ideas compare with modern LLMs, diffusion, and self-improvement concepts

PCA analysis of Futures returns for fun and profit, part deux (qoppac​.blogspot​.com) . PCA on futures universe, sign flipping issues, factor construction, clustering, and trading residuals with pysystemtrade

I’m supposed to present ‘Conformal Predictive Simulations for Univariate Time Series’ at COPA CONFERENCE 2025 in London… (thierrymoudiki​.github​.io) . Conformal predictive simulations for univariate time series; COPA 2025 poster, MLR Proceedings, conformal prediction, nnetsauce, ahead, Ridge2, Python/R/Ridge2f, conformalize, prob. forecasting

Transfer Learning using ahead::ridge2f on synthetic stocks returns (thierrymoudiki​.github​.io) . Pretrains ahead::ridge2f on 1000 synthetic stock returns with Bayesian Optimization and tests on European indices

The Physics of AI Hallucination: New Research Reveals the Tipping Point for Large Language Models (firstprinciples​.org) . Physicist Neil Johnson maps tipping point in LLMs, uses spin model, gap cooling, and attention head dynamics to predict hallucinations

The maths you need to start understanding LLMs (gilesthomas​.com) . High-dimensional vocab and embedding spaces, softmax, one-hot vectors, projections via matrices, and neural network linear layers explained

intuition (aarnphm​.xyz) . Visualization of autoencoders: encoder/decoder maps, latent manifolds, sampling challenges, and representation vs implementation diagrams

A Random Walk in 10 Dimensions (2021) (galileo-unbound​.blog) . High-dimensional random walks, 10D hyperlattices, percolation thresholds, ridges vs peaks, SAW vs ordinary walks, fitness landscapes, neutral networks, implications for evolution and deep learning

📚 Academic Research

Exploring the Design Space of Fair Tree Learning Algorithms (arxiv:cs) . Explores three fair tree learning designs: single-tree with joint objective, constrained splits, and dual-tree models for y and s

Fantastic Pretraining Optimizers and Where to Find Them (arxiv:stat) . Systematic hyperparameter tuning across ten optimizers, evaluating scale- and data-to-model ratios, revealing matrix-based optimizers’ scaling limits

Deep Reinforcement Learning for Ranking Utility Tuning in the Ad Recommender System at Pinterest (arxiv:cs) . DRL-PUT: personalized utility tuning for ad ranking using online logs to optimize multi-objective rewards in Pinterest’s ad recommender

A Plug-and-play Model-agnostic Embedding Enhancement Approach for Explainable Recommendation (arxiv:cs) . Plug-and-play RVRec: model-agnostic embedding enhancement using negative 2-Wasserstein contrastive loss and multivariate Shapley-based interaction value for explainable recommendations

LowDiff: Efficient Frequent Checkpointing via Low-Cost Differential for High-Performance Distributed Training Systems (arxiv:cs) . LowDiff enables frequent checkpointing in distributed training by reusing compressed gradients as differential checkpoints and batched writes

Kangaroo: A Private and Amortized Inference Framework over WAN for Large-Scale Decision Tree Evaluation (arxiv:cs) . Kangaroo: private, amortized decision-tree inference over WAN using packed HE with model hiding, secure feature selection, and oblivious path evaluation

An Empirical Evaluation of Factors Affecting SHAP Explanation of Time Series Classification (arxiv:cs) . Evaluates segmentation methods for SHAP explanations in time series classification, finding equal-length segmentation often best and introducing a length-weighted attribution normalisation

RapidGNN: Energy and Communication-Efficient Distributed Training on Large-Scale Graph Neural Networks (arxiv:cs) . RapidGNN enables energy- and communication-efficient distributed training for large-scale GNNs via deterministic sampling-based scheduling and remote feature prefetching

Enhancing Machine Learning for Imbalanced Medical Data: A Quantum-Inspired Approach to Synthetic Oversampling (QI-SMOTE) (arxiv:cs) . Quantum-inspired SMOTE (QI-SMOTE) for imbalanced medical data; enhances RF, SVM, LR, KNN, GB, neural nets on MIMIC-III/IV mortality

A Generative Foundation Model for Chest Radiography (arxiv:cs) . ChexGen: a latent diffusion transformer for text-, mask-, and bounding box-guided synthesis of chest radiographs using 960k radiograph–report pairs

On Hyperparameters and Backdoor-Resistance in Horizontal Federated Learning (arxiv:cs) . Hyperparameter tuning for benign clients reduces backdoor effectiveness in horizontal federated learning, improving robustness without sacrificing accuracy

Multi Attribute Bias Mitigation via Representation Learning (arxiv:cs) . Generalized Multi Bias Mitigation (GMBM) with ABIL and Gradient Suppression Fine Tuning for multi-attribute bias in vision, plus SBA metric

Hybrid Matrix Factorization Based Graph Contrastive Learning for Recommendation System (arxiv:cs) . Hybrid matrix factorization integrates low-rank MF and SVD to enhance graph contrastive learning for recommendation systems

Bayesian Additive Regression Trees for functional ANOVA model (arxiv:cs) . ANOVA-BART: functional ANOVA decomposition for interpretable Bayesian Additive Regression Trees with near-minimax posterior concentration and interaction-wise convergence

Finetuning AI Foundation Models to Develop Subgrid-Scale Parameterizations: A Case Study on Atmospheric Gravity Waves (arxiv:cs) . Fine-tuning AI foundation models to develop subgrid-scale parameterizations for atmospheric gravity waves using Prithvi WxC and flux learning

LimiX: Unleashing Structured-Data Modeling Capability for Generalist Intelligence (arxiv:cs) . LimiX: a unified, masked joint-distribution model for structured data handling tabular tasks via episodic context-conditioned pretraining

FoMEMO: Towards Foundation Models for Expensive Multi-objective Optimization (arxiv:cs) . FoMEMO: Foundation Models for Expensive Multi-objective Optimization via domain-trajectory conditioned pre-training and in-context preference aggregation

FlashRecovery: Fast and Low-Cost Recovery from Failures for Large-Scale Training of LLMs (arxiv:cs) . FlashRecovery enables fast failure detection, scale-independent restart, and checkpoint-free recovery for large-scale LLM training on 4,800 devices in 150 seconds

Uncertain but Useful: Leveraging CNN Variability into Data Augmentation (arxiv:math) . Investigates training-time variability in CNN-based FastSurfer for neuroimaging, using floating point perturbations and seeds to build ensembles for data augmentation and robustness

Distribution-valued Causal Machine Learning: Implications of Credit on Spending Patterns (arxiv:econ) . Distributional causal effects in Wasserstein space via Distributional Double Machine Learning; Neural Functional Regression Net; Conditional Normalizing Flow Net for continuous treatment

Foundational Models and Federated Learning: Survey, Taxonomy, Challenges and Practical Insights (arxiv:cs) . Survey and taxonomy of federated learning methods for foundational models, with healthcare focus, including self-supervised learning, fine-tuning, distillation, and transfer learning

Wild Refitting for Model-Free Excess Risk Evaluation of Opaque ML/AI Models under Bregman Loss (arxiv:stat) . Wild refitting with Bregman losses yields model-free excess risk bounds via wild optimism and randomized symmetrization for opaque models

Why Can't I See My Clusters? A Precision-Recall Approach to Dimensionality Reduction Validation (arxiv:cs) . Precision-Recall metrics for relationship phase in DR (t-SNE, UMAP) to diagnose missing cluster structure and guide hyperparameter tuning

👋 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.