Friday, August 14, 2026

Pulse (last 3d) · Research 82 · Agents 76 · Models 52 · Products 42 · Enterprise 32 · Industry 27 · Open Source 25 · Legal 22 · Media 15 · Inference 10

Trending · Claude 17 · OpenAI 12 · Claude Code 11 · ChatGPT 9 · Codex 8 · Nvidia 8 · Anthropic 7 · Cursor 6 · Google 6 · FDA 5 · Gemini 5 · Meta 5

---

DARTree: Speculative Diffusion Decoding with Autoregressive Draft Trees. Tree-based speculative decoding tailored to diffusion language models, accelerating inference without sacrificing sample quality. Source

Vero: Can AI Agents Build Formally Verified Software Repositories?. Evaluates whether AI coding agents can construct repositories backed by formal proofs. Source

AaLLM: End-to-End Analog Circuit Design via LLMs. Framework taking analog circuit design from topology generation through sizing. Source

DFM Mimir v1: Open 1B HRM with Frontier-Level Performance. Open 1B-parameter hybrid reasoning model trained on permissible post-training data only. Source

WebBrain. Free, open-source AI browser agent for Chromium/Firefox; runs locally via llama.cpp with optional BYO-key cloud fallback. Source

Chiplab. Virtual embedded-chip sandbox; lets coding agents build, run, and debug STM32/Nordic firmware over MCP without hardware. Source

Caveman. Local proxy that compresses logs, tool output, and files before provider calls; benchmark shows 33% fewer input tokens with 18/18 correctness checks. Source

Mistral OCR 4.1. Updated OCR-specialized model from Mistral with community debate on OCR-vs-LLM tradeoffs. Source

OpenAI previews Ultrafast GPT-5.6 Sol on Cerebras. Up to 750 tok/s and ~14× standard speed; Cerebras claims 7× speedup over Claude on HLE. Source · Cerebras

Mistral now hosting GLM-5.2. Z.ai's flagship priced below Mistral Medium 3.5, hinting at a strategic pivot toward model-agnostic hosting. Source

Writer launches new model + token-cost-focused harness. Aims to contain token spend without sacrificing capability. Source

White House reportedly prepping to extend safety-testing framework to open models. Would fold open-weight releases into the secret pre-release review regime. Source

Claude outputs are now watermarked. Reports surface that Anthropic has begun applying watermarks to Claude generations. Source

Doom running on an LLM. Novelty demo with a Hugging Face checkpoint showing an LLM acting as a game agent. Source

Worldproof: when pixel metrics stop ranking models. Open-source world-model diagnostic; copy-last-frame baseline beats many pixel-metric comparisons on real robot video. Source

Zvi's AI #181: Astra Goes Cyber Critical. Weekly roundup focusing on Astra's cybersecurity implications. Source