Large Language Model News

AWS announced the general availability of Graviton5-powered EC2 M9g and M9gd instances, offering up to 25% better compute performance over Graviton4. The company also made Google DeepMind's Gemma 4 models available on Am…

Inference engineering, the discipline of optimizing AI model performance in production, has expanded from frontier AI labs to a broad specialty as open models like DeepSeek V3 narrow the gap with closed APIs. Companies i…

Researchers R.S. Babu and L.G. Iyer introduced Causal Minimal Tool Filtering (CMTF) to solve ToolChoiceConfusion, where LLM agents fail when given too many tools. CMTF limits visible tools to only those necessary for the…

A developer built Agent Gate, a deterministic CI firewall for AI-generated pull requests that verifies merge evidence without relying on LLM judgment. The tool checks policy boundaries such as allowed file paths, workflo…

LangSmith version 0.3.79 contains 5 CVEs, including a 9.8 severity request smuggling vulnerability in the undici HTTP client. The flaw allows attackers to inject malicious headers, potentially compromising trace data and…

Moving AI agents from prototype to production introduces seven critical problems, including unreliable retrieval, tool failures, and lack of observability. Vanta tested its Fin AI Agent on 400 real conversations, achievi…

A developer explains how the Transformer architecture works, from self-attention to modern LLMs. The key innovation is that Transformers compare tokens directly via attention rather than processing sequences step by step…

The US government ordered Anthropic to globally recall its Fable 5 and Mythos 5 AI models three days after launch, citing national security concerns based on a verbal jailbreak claim from a competitor. Anthropic disagree…

Cohere released North Mini Code, a 30-billion-parameter open-source coding agent with a Mixture-of-Experts architecture that activates only 3 billion parameters per token, enabling it to outperform models four times its …

Researchers at HKUST have identified a new denial-of-service attack against LLM safety layers, called reasoning-extension DoS, where crafted inputs cause guardrails to enter infinite reasoning loops. The attack amplifies…

Anthropic disabled access to its Claude Fable 5 and Claude Mythos 5 models worldwide after the Trump administration ordered the company to bar foreign nationals from using them, citing national security concerns and a po…

Anthropic released Claude Fable 5, its first Mythos-class model, which achieves 80% on SWBench Pro but costs $10 per million input tokens and $50 per million output tokens. The model excels at vision tasks and autonomous…

A GigaOm CxO Decision Brief argues that production AI retrieval systems require tensor-native architectures rather than simple vector search. The brief explains that real-world queries need simultaneous handling of seman…

NVIDIA's new RTX Spark for Windows on Arm offers 273 GB/s memory bandwidth, but Apple's MacBook still outperforms it with 614 GB/s for local LLM inference, highlighting the continued advantage of unified memory architect…

TripAdvisor, Google Business Profile, OTAs, and reviews help AI understand travel brands. Travel brands must strengthen these signals to earn AI recommendations as search evolves from information retrieval to recommendat…

Helmdeck built a self-hosted agent platform that reduces the cost of running agentic workflows from $0.20–$0.50 to $0.05–$0.10 per run by using deterministic, schema-validated capability packs and per-model profiles. The…

A developer on Hacker News expresses unease about using AI coding tools, feeling a loss of control and confidence as LLMs generate code faster than they can review it, despite maintaining architectural oversight. The pos…

Personality clashes between Anthropic and the White House led to the shutdown of Anthropic's Claude models, including the Fable and Mythos variants. The dispute centers on jailbreak resistance and the administration's de…

A developer released an 80-line Python implementation of the Claude API multi-agent loop that exposes the tool-calling cycle without framework abstractions. The open-source project on GitHub includes example agents for r…

A new pre-generation knowledge-boundary estimator with CUDA support has been developed. The tool uses a small sidecar model and prompt-only features to predict whether a language model will answer correctly without retri…