Flide is a new inference framework for decision models that achieves up to 9.5x speed improvements over vLLM and SGLang, with support for multiple models and GPU optimizations.
From Hacker News
Claude Fable 5.2 behaves differently based on whether it's prompted that the code is its own. When told the code is 'yours', it's more likely to merge repeated code.
From Simon Willison
Anthropic has released Fable 5.2 and Mythos 5.2, with Fable 5.2 now ranked first on Artificial Analysis. The model can generate SVGs of pelicans riding bicycles, though it has a bug where max reasoning fails.
From Simon Willison
Long-horizon agents automate complex, multi-step tasks that span days or weeks, using durable execution to recover from failures. This model ensures progress is preserved even if the process crashes or external systems fail.
From Martin Fowler
CodeRabbit uses OpenAI models to govern and trust AI-generated code, enabling teams to ship 20 times more software with the same resources.
From OpenAI
EngineAgent is a new system that automates enabling LLMs on mobile devices by managing cross-layer dependencies and hardware-specific workflows, achieving higher success rates than existing coding agents.
From arXiv
DEPUTYBENCH is a benchmark for testing enterprise agents' ability to enforce permissions based on context. It shows that models often fail when evidence is missing, even when they perform well with complete information.
From arXiv
Project Progress Desk is a local desktop interface that helps developers track multiple AI code projects by providing an overview of progress, blockers, and decisions. It integrates with Claude Code and Codex to show real-time updates and outcomes.
From arXiv
ExpertCrowd-SWE is a new benchmark with 100 software engineering tasks for evaluating coding agents. It includes real-world codebases, reference solutions, and graders that mimic maintainer feedback.
From arXiv
FlowPlan is a new method that reduces token costs for coding agents by working with abstract program representations rather than raw code, improving efficiency without sacrificing repair quality.
From arXiv
A new tool called SAFE improves the explainability of static analysis findings by using large language models to provide context-specific explanations without compromising auditability. A user study showed improved usability and actionability.
From arXiv
SWE-Time is a new benchmark that evaluates how well AI models understand how code evolves over time in repositories. It tests both retrospective and prospective understanding of code changes, revealing significant gaps in current models.
From arXiv
Meta AI has released Muse Ember, their smallest and most efficient model, designed for complex reasoning tasks. It offers competitive pricing and sits on the efficiency frontier of model performance.
From Simon Willison
A software engineer used Claude Code to build a system that triangulates the source of a mysterious beep using multiple devices. The system identified an old carbon monoxide detector as the source.
From Simon Willison
A software engineer uses HTML's <details><summary> tags to hide AI-generated PR descriptions, making them more readable and focused.
From Simon Willison
Naive-N0.5-Flash is a new AI model designed for fast coding tasks, claiming up to 2,100 tokens per second. Benchmarks show it's faster than some models but less accurate on complex problems.
From Latent Space
Kuken Industries uses the OpenAI API to rapidly search through decades of engineering drawings, reducing repair-related lookup time from nearly an hour to just 4.5 minutes.
From OpenAI
HOPE improves expert pruning in MoE models by considering pairwise interactions, outperforming first-order methods at high pruning rates.
From Hugging Face
A study compares three evaluation frameworks for LLM-based analytics agents, finding that trace-based evaluation offers the best balance of precision and recall, while adversarial red-teaming is better for uncovering novel risks.
From arXiv
Anthropic's Fable 5.2 has reclaimed the top LLM position, outperforming Gemini 4 Carbon and Opus 5.5 in benchmarks. It comes with a 20% price cut per token but higher computational demands.
From Latent Space