Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Agentify has forked the archived TensorZero project, which raised $7.3M, and released Agentify Gateway, an open-source LLM gateway with observability, evaluation, optimization, and experimentation features. The community…
Gabe Pereyra released the Legal Agent Benchmark (LAB), an open-source benchmark for evaluating AI agents on complex legal tasks, and shared initial results on frontier model performance in long-horizon legal-agent work.
OpenAI released GPT-5.4 Mini at $0.75 per million input tokens and $4.50 per million output tokens, while GPT-5.4 costs $2.50 and $15, a 3.3x price multiplier. A developer's task-by-task analysis shows that 50-70% of pro…
Anthropic's prompt caching offers up to 90% discount on cached input tokens via cache_control markers, with a 1.25x write premium for a 5-minute TTL or 2x for a 1-hour TTL. Break-even occurs at the second cache hit for t…
AgentPay, an open-source Python SDK and MCP server, enables AI agents to collect VietQR payments directly to a business bank account without intermediaries. The tool bypasses Stripe and traditional gateways by generating…
A developer describes a practical division of labor between two agentic coding tools: Claude Code for exploratory, conversational tasks and Codex for repetitive, straight-line automation. The workflow, built around Codex…
Loopy, a local terminal meta-agent that watches Claude Code and Codex sessions to identify repetitive patterns and propose automation loops, has been released. The tool aims to help developers shift from manual prompting…
Researchers developed custom Triton kernels that accelerate JumpReLU Sparse Autoencoder inference by 2–14× on real SAEs, exploiting activation sparsity to skip zero entries during matrix multiplication. The kernels use e…
A developer at dualform labs released two open-source skills for Claude Code that address common agent failures: spec, which forces upfront clarification of ambiguous tasks, and review-audit, which performs a read-only a…
A developer refactored a 900-line demonstration file into a reusable Python package called 'harness', which includes modules for action registration, permission budgeting, input sanitization, audit logging, and rollback …
A developer reports that AI is transforming production debugging by rapidly reading and correlating heterogeneous signals, with tools like Datadog's Bits AI SRE claiming time-to-resolution wins of up to 95% in benchmarke…
ScopeKit launched a new tool that creates task-specific context packs for AI coding agents, providing relevant files, symbols, tests, and relationships for a given task. The tool claims to reduce context size by up to 94…
Nylas introduces Agent Accounts with a human-in-the-loop email approval workflow that uses a drafts folder as a safety gate. The system allows an LLM to draft replies automatically while high-risk messages are held for h…
Anthropic disabled access to its advanced AI models, including Mythos, after the Trump administration ordered the company to block foreign nationals from using the technology, citing national security. The unprecedented …
A developer argues that most RAG applications fail because they still parse PDFs as plain text instead of preserving document layout. The recommended approach is to use tools that output structured Markdown, such as IBM'…
A developer identifies four failure vectors that cause LLM prompts to break in production: input distribution mismatch, context contamination over multi-turn conversations, model drift from provider updates, and adversar…
Microsoft released Intelligent Terminal for Windows 11, a fork of Windows Terminal that adds an AI-powered side panel for command-line assistance. The tool supports GitHub Copilot and other Agent Client Protocol-compatib…
SK Group Chairman Chey Tae-won urged employees to adopt AI agents as part of a companywide AI transformation, calling for a "one-person, one-agent" initiative at the 2026 Icheon Forum. Chey said more than 90% of employee…
OpenAI's Codex now completes 85-90% of well-scoped maintenance tasks using its /goal workflow, which loops until a goal is verified complete or a token budget is exhausted. The feature, available in Codex CLI 0.128.0 and…
A cloud architect at an unnamed company reduced their recommendation engine's AI spend by 60% without sacrificing quality by switching from a single-model architecture to a tiered routing system. The team now uses cheape…