Friday, August 21, 2026

Pulse (last 3d) · Research 108 · Products 72 · Agents 62 · Models 35 · Open Source 20 · Media 20 · Industry 19 · Legal 17 · Enterprise 14 · Inference 12

Trending · Claude 21 · OpenAI 13 · ChatGPT 12 · Anthropic 8 · Codex 8 · Claude Code 7 · Cursor 7 · Gemini 5 · verl-uni-agent 5 · Google 3 · Hugging Face 3 · Nvidia 3

---

Grok 4.6 - xAI ships agentic-upgraded Grok 4.6 with stronger software engineering and web-app generation, pricing unchanged at $2/$6 per 1M tokens. Source

NobodyWho - Open-source on-device LLM inference engine built on llama.cpp, with SDKs for Swift, Kotlin, Flutter, RN, Python, and Godot. Source

Rubric - MCP-served corpus of 554 curated UX rules that lets coding agents pull prescriptive guidance before building and audit a live site after shipping. Source

NVIDIA CUDA MCP - Official NVIDIA-hosted Model Context Protocol server giving AI assistants access to current CUDA docs, GPU code generation, and performance analysis. Source

Vomit - Tool that runs a second LLM to clean Claude 5's verbose, on-brand output, surfacing deeper issues with agent communication-style enforcement. Source

AI4AI-Bench - Benchmark for evaluating LLM agents that design algorithms aimed at recursive self-improvement. Source

FlashPrefill V2 - Block-sparse prefill attention delivering long-context inference speedups for production LLM serving. Source

Faraday (Inherent Labs) - Demonstrates that a per-task rubric-judge reward recipe, not the closed model itself, lets rubric-RL beat frontier agents at paper replication. Source

Learning When to Think - Adaptive-reasoning method that decides when an LLM should think versus answer directly, dynamically allocating test-time compute. Source

OpenAI closing the enterprise gap with Anthropic - New data indicates OpenAI is gaining on Anthropic among business users. Source

Micro1 hits $500M gross run rate - AI training-data startup scales sharply amid the ongoing data-labeling boom. Source

Thomson Reuters CoCounsel Legal - Next-gen CoCounsel ships with Westlaw Brief Builder, agentic drafting, and tabular review of up to 10,000 documents. Source

Grok emitting gibberish - xAI's Grok has been sending garbled responses to users, adding to a turbulent rollout week. Source

LLM résumé-screening study, corrected - Follow-up with controlled re-runs addressing three methodological critiques from the original bias-study thread. Source

The alignment tax - Essay arguing closed-model safety pipelines impose a 25-35% hidden compute overhead through system prompts, classifiers, and hedging. Source

Coding agent approved its own bug - Developer argues for splitting code-writing and code-reviewing agents after the same agent signed off on its pagination regression. Source