Generative AI: 1st July 2025

Boosting Developer Productivity with AI: Faster Dashboards, Automated Testing, and 70% Less Setup Time (engineering​.salesforce​.com). Salesforce enhances developer productivity with AI tools like Code Builder, automated testing, and warm pool optimization, achieving a 70% reduction in setup time

From pair to peer programmer: Our vision for agentic workflows in GitHub Copilot (github​.blog). GitHub Copilot evolves from an assistant to a peer programmer with independent AI agents, enhancing developer workflows through multi-step reasoning and collaboration

Gemma 3n fully available in the open-source ecosystem! (huggingface​.co). Gemma 3n launched as an open-source multimodal AI model supporting diverse inputs, featuring E2B and E4B variants for efficient local performance

Normalizing Flows Are Capable Generative Models (machinelearning​.apple​.com). TarFlow, a Transformer-based Normalizing Flows model, achieves state-of-the-art results in likelihood estimation and image generation with advanced techniques for improved quality

Scaling Pinterest ML Infrastructure with Ray: From Training to End-to-End ML Pipelines (medium​.com/pinterest-engineering). Pinterest optimizes ML infrastructure using Ray, enhancing feature development, sampling, and labeling while achieving faster iteration and reduced costs

A Survey of LLM Inference Systems (arxiv:cs). Survey of LLM inference systems like vLLM, SGLang, Mooncake, and DeepFlow, addressing techniques for request processing, optimization, execution, and memory management

Towards Transparent AI: A Survey on Explainable Large Language Models (arxiv:cs). Survey on explainable AI methods for large language models, focusing on transparency, evaluation, applications, and future challenges in high-stakes domains

Language Modeling by Language Models (arxiv:cs). Genesys uses multi-agent LLMs for discovering novel LM architectures through genetic programming, achieving competitive designs outpacing known benchmarks

A Dual-Layered Evaluation of Geopolitical and Cultural Bias in LLMs (arxiv:cs). Evaluating bias in large language models through factual and geopolitical scenarios, assessing model training and query language effects across diverse cultural contexts

Breaking the Boundaries of Long-Context LLM Inference: Adaptive KV Management on a Single Commodity GPU (arxiv:cs). LeoAM enables efficient long-context LLM inference on commodity GPUs using adaptive KV management, achieving significant speedup while maintaining response quality

TopK Language Models (arxiv:cs). TopK LMs enhance interpretability of transformer architectures using TopK activation, improving efficiency, stability, and neuron analysis without the need for post-hoc training

A foundation model with multi-variate parallel attention to generate neuronal activity (arxiv:cs). Multi-variate parallel attention (MVPA) in MVPFormer enables efficient iEEG signal prediction and seizure detection, leveraging the SWEC dataset of nearly 10,000 hours of recordings

GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling (arxiv:cs). GPAS enhances Pre-LayerNorm Transformers by scaling activations while preserving gradients, improving training dynamics and performance across various model sizes and architectures

Normalizing Flows Are Capable Generative Models (machinelearning​.apple​.com). TarFlow, a Transformer-based Normalizing Flows model, achieves state-of-the-art results in likelihood estimation and image generation with advanced techniques for improved quality

The AI infrastructure stack with Jennifer Li, a16z (complexsystemspodcast​.com). Jennifer Li from a16z discusses AI’s impact on software infrastructure, middleware evolution, and the future of SaaS in a Complex Systems podcast episode

Reward Models (cameronrwolfe​.substack​.com). Exploration of Reward Models in LLMs using the Bradley-Terry model, preference scoring, and reinforcement learning techniques for improved output generation

Paper Review: ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models (andlukyane​.com). ProRL enhances reasoning in large language models by employing prolonged reinforcement learning to develop novel reasoning strategies beyond base model capabilities

Context engineering (simonwillison​.net). Context engineering emerges as a refined approach to prompt engineering, emphasizing the importance of filling context windows for LLM tasks, with insights from industry leaders

Life of an inference request (vLLM V1): How LLMs are served efficiently at scale (ubicloud​.com). vLLM is an open-source inference engine optimizing large language model serving with GPU deployment, continuous batching, and sophisticated token processing

Building a React-Style LLM Tool with LangChain and OpenAI: A Minimal Working Example (blog​.devgenius​.io). Explore a minimal example of a React-style LLM tool using LangChain and OpenAI's gpt-4o-mini, featuring a tool-augmented agent framework

Escaping LLM piping mess with nifty engineering (tinystruggles​.com). Upgrading tangled Python notebooks into a resilient content-adaptation studio using FastAPI, Next.js, and improved LLM processes for language learning

2025-06-27: Paper Summary: MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery (ws-dl​.blogspot​.com). MemoRAG enhances RAG with memory-inspired architecture, improving long-term information retrieval and multi-hop reasoning for complex queries and summarization tasks

Building Production-Grade RAG at Scale (thedataexchange​.media). Douwe Kiela discusses RAG 2.0, document intelligence, multimodal challenges, reasoning models, and strategies for effective retrieval-augmented generation systems

Agentic search for dummies (benanderson​.work). Agentic search using models, corpus preparation, search indexing with Tantivy, and offline document augmentation for enhanced AI queries and results

Reliability for unreliable LLMs (stackoverflow​.blog). Strategies to add determinism and reliability to workflows using non-deterministic large language models, focusing on inputs, outputs, and observability measures

Introducing vitals, a toolkit for evaluating LLM products in R (tidyverse​.org). vitals, an R toolkit for evaluating large language model (LLM) products, simplifies assessments of custom chat and query chat apps using datasets of challenging coding problems

How to Build Bulletproof LLM Eval Systems: The Complete Implementation Guide (joshpitzalis​.com). Implement evaluation frameworks like Analyze-Measure-Improve to achieve 99% reliability in LLM applications, utilizing structured data and systematic coding

Can AI Judge the Quality of AI Generated Design (designforam​.com). Kristen Edwards discusses AI’s role in evaluating design quality and advancements in vision-language models for engineering applications

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.