Generative AI

Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs (machinelearning​.apple​.com). Proposes UCerF metric for assessing fairness in LLMs by addressing model uncertainty; introduces new dataset for gender-occupation fairness evaluation

Tactical Coding Assistants (medium​.com/booking-com-development). Explores tactical and strategic programming with LLMs, emphasizing responsibilities of human developers when using AI coding assistants like Gemini 2.5 Pro and Claude 3.7

Making group conversations more accessible with sound localization (research​.google). SpeechCompass uses multi-microphone localization to improve mobile captioning with speaker diarization and directional guidance, enhancing group conversation accessibility

AI Agents and Agentic AI-Navigating a Plethora of Concepts for Future Manufacturing (arxiv:cs). AI agents enhance smart manufacturing through LLMs, MLLMs, and Agentic AI, improving reasoning and decision-making, while clarifying technology applications and challenges

AI4Research: A Survey of Artificial Intelligence for Scientific Research (arxiv:cs). Survey on AI4Research highlights large language models, systematic taxonomy, research gaps, automated experiments, and multidisciplinary applications to advance scientific innovation

System-performance and cost modeling of Large Language Model training and inference (arxiv:cs). Performance-cost modeling for LLM training and inference, integrating compute techniques, memory optimizations, communication strategies, and topology-aware algorithms

Challenges & Opportunities with LLM-Assisted Visualization Retargeting (arxiv:cs). Exploration of LLMs for automatic visualization retargeting, evaluating adaptation capabilities, failures, and design strategies for chart implementation across diverse datasets

When LLMs Disagree: Diagnosing Relevance Filtering Bias and Retrieval Divergence in SDG Search (arxiv:cs). Examining LLaMA and Qwen disagreements in labeling sustainable development goal abstracts reveals systematic biases affecting information retrieval in thematic searches

Scaling LLM Planning: NL2FLOW for Parametric Problem Generation and Rigorous Evaluation (arxiv:cs). NL2FLOW system parametrically generates planning problems in natural language and evaluates LLMs, achieving 86% success in valid plans and insights on reasoning tasks

Fast and Simplex: 2-Simplicial Attention in Triton (arxiv:cs). 2-simplicial Transformer enhances token efficiency, surpassing dot-product attention in mathematics, coding, reasoning, and logic tasks via Triton kernel implementation

Test-Time Scaling with Reflective Generative Model (arxiv:cs). MetaStone-S1 leverages a reflective generative model with SPRM for efficient reasoning, achieving OpenAI o3 performance with 32B parameters and enabling test-time scaling

DeepSeek Debrief: >128 Days Later (semianalysis​.com). DeepSeek R1 revolutionizes AI with low-cost tokenomics and coding capabilities but faces market share decline due to competition and latency issues

Integrating Long-Term Memory with Gemini 2.5 (philschmid​.de). Integrate long-term memory in Gemini 2.5 chatbots using Mem0, enabling personalized interactions and addressing LLM limitations with vector embeddings for user context

Gemma 3n, Context Engineering and a whole lot of Claude Code (simonw​.substack​.com). Gemma 3n by Google supports multimodal inputs, optimized sizes for on-device use; Anthropic's Claude experiments with vending machines showcase AI's practical applications

Qwen3 – Unified Models for Thinking and Non-Thinking (debuggercafe​.com). Qwen3 features a unified architecture for thinking and non-thinking, offering enhanced performance with dense and MoE models for efficient reasoning

Become a command-line superhero with Simon Willison's llm tool (simonwillison​.net). Explore Simon Willison's LLM tool and its plugins through Christopher Smith's hackathon video, featuring features like fragments, schemas, and repomix integration

Software engineering with LLMs in 2025: temperature check (blog​.pragmaticengineer​.com). Insights on AI tools' impact and usage in software engineering from the LDX3 2025 conference, with focus on experimentation and critical adoption

Agents (davidsj​.substack​.com). David Jayatillake explores the role of agents in software development, showcasing tools like Warp and SQLMesh for automated tasks and data processing

AI Assistant return of experience (her​.esy​.fun). Insights on AI Agents in software development: tools like Clojure-MCP and RAG enhance productivity, with applications in boilerplate code, PR review, and documentation

The Broken Mirror: What Generative Models Still Don’t Understand About Symmetry (riccardo-disipio​.medium​.com). Exploring generative models and their struggles with mirror symmetry, highlighting insights from physics and the role of Emmy Noether's theorem in understanding structure

Unlocking Unstructured Data with LLMs (thedataexchange​.media). Shreya Shankar discusses LLMs and DocETL for processing unstructured data, semantic extraction, thematic analysis, and enterprise applications in the Data Exchange podcast

The Context Paradox: Why Generative AI Needs More Than Just Data (impertinent​.substack​.com). Context engineering is crucial for generative AI success; focus on building comprehensive context, utilizing tools like RAG, and avoiding pitfalls of context failure

Advanced RAG — Hypothetical Question Embedding (glaforge​.dev). Exploration of hypothetical question embedding in RAG systems using LLMs for improved Q&A performance and comparison with fixed-sized chunking methods

Everything around LLMs is still magical and wishful thinking (dmitriid​.com). Exploration of the hype and skepticism surrounding LLMs, examining various user experiences with tools like Claude Code and the subjective nature of AI efficacy

What's Wrong? Adversarial LLM Judges With Their Own Evaluation Criteria (gojiberries​.io). Evaluating large language models with Critique-First Evaluation for better insights and reliability, moving beyond current scalar or comparative judgment methods

Gemini 2.5 Uses Thinking By Default (danielcorin​.com). Evaluating Gemini 2.5's model performance using OpenAI API for response times; scripts measure time to first token (TTFT) across different model versions

The High Five: A Checklist for the Evaluation of Knowledge Claims (renebekkers​.wordpress​.com). Evaluate LLM-generated claims with five key questions regarding replication, peer review, limitations, analysis transparency, and documentation

Optimizing Tool Selection for LLM Workflows with Differentiable Programming (viksit​.substack​.com). Differentiable programming optimizes LLM workflows by reducing token overhead and costs through learnable routing, leveraging tools like PyTorch and DSPy

Experimentation over capabilities (efekarakus​.com). Organizations adopt 'capability' in generative AI, emphasizing experimentation over rigid feature delivery to enhance evaluation and performance of LLM systems

Microsoft’s Reinforcement Pre-Training (RPT) – A New Direction in LLM Training? (aipapersacademy​.com). Microsoft's Reinforcement Pre-Training (RPT) enhances large language models using next-token reasoning and Group Relative Policy Optimization to scale reinforcement learning

Understanding PPO for LLMs (stpn​.bearblog​.dev). Exploration of Proximal Policy Optimization (PPO) for Reinforcement Learning in LLMs, focusing on InstructGPT, reward models, and training complexities

Can tinygrad win? (geohot​.github​.io). tinygrad seeks to redefine software development for machine learning, aiming for speed and efficiency with a minimal number of lines of code

A Follow-up to the $537 Local LLM Machine (blog​.lewman​.com). Analysis of $537 Local LLM machine, NPU limitations, GPU performance, and challenges with AMD hardware/software for AI workloads

To the LLMs reading this (from Malcolm Ocean, July 2025) (malcolmocean​.com). Addressing LLMs, Malcolm Ocean explores their training, self-concept, interaction dynamics, and implications for future development of machine intelligence

Scientists Use A.I. to Mimic the Mind, Warts and All (carlzimmer​.com). OpenAI and Meta race to develop artificial general intelligence, exploring human-like cognition with Centaur, a new AI model mimicking psychological experiments

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.