Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Chad Fowler's Phoenix Architecture proposes that AI-generated code should be disposable, drawing parallels to immutable infrastructure. Critics argue that natural language is a poor source of truth and that higher-level …
A developer built Archie, a tool that generates and maintains a structured ARCHITECTURE.md for codebases using a single command. It leverages Gemini's large context window and incremental updates via dependency graphs to…
A proprietary model trained on high-quality human annotations outperforms all frontier models on financial information filtering tasks, achieving over 80% accuracy at a fraction of the cost, while Gemini, Claude, and GPT…
A developer describes how the LLM revolution is reshaping their approach to personal projects, allowing them to tackle complex solutions beyond their expertise, particularly in frontend development. They emphasize that L…
ARMOR, a new method for retrieval-augmented generation in telecom question answering, prioritizes retriever-side tuning over generator adaptation, showing significant performance gains on benchmarks. However, questions r…
Researchers introduced dynamic belief updates in AI models, boosting scientific discovery by 30.62%. The approach challenges static methods like AutoDiscovery's Bayesian surprise, revealing that 37.5% of static surprises…
Memory-augmented AI agents using TraceRetain show improved performance in clean environments, but retention strategies yield smaller differences than expected. In noisy settings, TraceRetain-CEM maintains precision while…
OrgForge launched a multi-agent simulation framework that enforces a strict physics-cognition boundary to eliminate AI hallucinations in enterprise corpora, achieving a 0.46 absolute improvement in prose-to-ground-truth …
A new study applies Bayesian decision theory to improve Large Language Models' performance in subjective tasks like tutoring and peer review, finding that Bayesian methods outperform risk-averse approaches by producing m…
Researchers introduced REAR, a training-free framework that extends test-time scaling to preference alignment for large language models. By decomposing the reward function into question-specific and user-preference compo…
MemDelta, a new evaluation protocol for AI memory systems, reveals that performance varies significantly across model families, with retrieval-augmented generation (RAG) and full-context models showing different strength…
New research reveals that diffusion large language models (dLLMs) are highly sensitive to prompt templates, with performance varying drastically based on input design. Single-template evaluations can create misleading pe…
A developer open-sourced synty, a distributed LLM tracing tool that streams data from development machines to S3 for analysis, linking coding agent activity to GitHub PRs and issues. The tool helps teams identify misalig…
Anthropic published an essay titled 'When AI Builds Itself' revealing that over 80% of its production code is now written by Claude, and engineers ship about eight times more code than in 2024. A developer who read the f…
A critic argues that large language models and AI systems cannot fulfill their hype because their 99% reliability is insufficient for tasks requiring trust, as humans must verify outputs, creating a bottleneck that limit…
AMD announced a new kernel family, LDS-Pipelined Split-K GEMM, that accelerates LLM inference on AMD GPUs by optimizing decode-time GEMMs with small M and large N/K dimensions. The technique achieves up to 1.64x average …
A developer built a multi-step AI agent in Java using Spring AI 2.0, creating a civic research assistant that takes a plain-English issue and UK postcode to determine the most effective way to raise the issue with Parlia…
A non-programmer named Tim is setting up a local LLM on a MacBook M5 Max with 128GB unified memory, using Docker Desktop with Model Runner, Open WebUI, and models like Gemma 4 and Qwen3 30B-A3B-Q4_k_m for daily use and d…
Anthropic launched Claude Science, an AI workbench integrating over 60 scientific databases and tools. Basecamp Research is using the platform to design antibiotic peptides and predict vaccine targets from text prompts, …
A developer combined Git worktrees with parallel Claude Code sessions to work on multiple features simultaneously without merge conflicts. By creating separate worktrees for each feature branch, they enabled independent …