Sunday, July 26, 2026

Pulse (last 3d) · Agents 59 · Research 42 · Models 23 · Products 19 · Industry 18 · Open Source 12 · Legal 12 · Enterprise 11 · Inference 8 · Media 8

Trending · OpenAI 12 · Codex 9 · Claude 7 · Anthropic 6 · ChatGPT 6 · Claude Code 5 · Nvidia 5 · DeepSeek 4 · GPT-5.6 4 · AGENTRY FIND 3 · AMD 3 · FDA 3

---

Claude Opus 5: The System Card. Zvi Mowshowitz publishes a detailed walkthrough of the Claude Opus 5 system card. Source

Introducing Claude Opus 5. Simon Willison notes Opus 5 leads the Artificial Analysis leaderboard while pricing matches the older Opus 4 tier. Source

Anthropic releases Opus 5 with 'close' to Claude 4.5's capabilities. The Verge covers the release, framing it as near-flagship capability at half the price. Source

Opus 5's effort dial is not monotonic. Community testing finds coding scores drop above the "high" effort setting, contradicting Anthropic's own migration guide. Source

Kimi K3 cyber-capable open-weight release. UK AISI/CAISI find Kimi K3 outperforms GLM-5.2 on ExploitBench (32% vs 24%) and is the first open-weight model to solve "The Last Ones" cyber range; weights drop July 27. Source

SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals. Combines hybrid linear attention with attention residuals to make video diffusion cheaper. Source

Dataset Distillation by Influence Matching. Distills datasets by matching per-example influence rather than outputs or gradients. Source

Multi-Turn On-Policy Distillation with Prefix Replay. Stabilizes student training in multi-turn on-policy distillation by replaying shared prefixes. Source

llama.cpp gains full MCP support. Native MCP support (including stdio servers) lands in llama.cpp, enabling fully agentic chat in its WebUI. Source

Cloudflare's new AI traffic options. Cloudflare adds controls to block Googlebot-style training crawlers and manage AI traffic by default. Source

AutoDev Studio: open-source multi-agent SDLC harness. Multi-agent coding tool that learns a repo once, reportedly 7–75% cheaper than cold Claude Code on 6/6 localized tasks. Source

Running a 28.9M parameter LLM on an $8 microcontroller. Demo project runs a small LLM on an ESP32. Source

White House science blueprint: more AI, less life sciences. Report proposes shifting billions of federal research dollars from universities toward AI at tech companies. Source

One fallen power line exposed an AI data center problem. TechCrunch examines how a single downed line revealed growing grid-reliability risks for AI data centers. Source

Bipartisan bill would require AI disclosure. US lawmakers propose mandating that companies tell users when they're interacting with AI. Source

You can't ignore Google Zero anymore. The Verge examines "Google Zero" and its impact on publishers and AI-driven traffic. Source

Open-weight AI is having its Kubernetes moment. Essay drawing parallels between open-weight AI distribution and the early Kubernetes adoption pattern. Source

TorchWright: compile computation graphs into transformer weights. A "compiler" that turns arbitrary Python graphs into vanilla Phi-3 transformer weights with no training. Source

Why good AI agents still produce bad system outputs. Discussion thread on the gap between agent capability and end-to-end output quality. Source

Google comes out in favor of open-weight models. Reddit summary claiming Google has publicly aligned with the open-weight camp against Anthropic. Source