Mid-August is shaping up as an infrastructure moment: the big labs are no longer just shipping bigger models but redesigning the hardware and safety architecture underneath them. Meanwhile the open-source community on GitHub has quietly shifted toward treating AI agent tooling as production engineering rather than research prototyping.
Anthropic Starts Building Its Own Chips
After years of relying entirely on Nvidia, Google TPUs, and AMD hardware, Anthropic confirmed it is assembling an in-house custom silicon team aimed at running Claude models faster and at lower cost per token. The company posted roles spanning hardware and software — with salaries up to $485,000 — and described the group as a "custom silicon team" that will co-design chips and models together rather than treating them as independent concerns. Samsung is reportedly a candidate manufacturing partner. Anthropic made clear it isn't abandoning its existing cloud and chip-supplier relationships; the in-house effort is additive, driven by the arithmetic of serving billions of tokens a day at a $30 billion revenue run rate.
techcrunch.com — Anthropic is hiring an AI chip design team
OpenAI Expands Daybreak With a Dedicated Cybersecurity Model
OpenAI shipped GPT-5.6-Cyber on August 10 as part of an expanded Daybreak program, giving approved security professionals access to a model trained specifically for offensive and defensive security work. Daybreak now runs two tiers: Blue, which removes system-level guardrails from GPT-5.6 Sol for defensive tasks like vulnerability discovery and incident response; and Red, which handles riskier scenarios including exploit-chain research and authentication bypass testing. In internal evaluations, GPT-5.6-Cyber completed 95% of advanced cybersecurity prompts while the standard GPT-5.6 Sol answered just 1.5%. Partners getting early access include Accenture, IBM, CrowdStrike, Cisco, Sophos, and Cloudflare.
qz.com — OpenAI expands Daybreak cybersecurity program, launches GPT-5.6-Cyber
ChatGPT Drops Text Chat Limits for Free Users
OpenAI removed the cap on text-based conversations for free ChatGPT accounts, with GPT-5.6 Luna now serving as the default model for all free and Go-tier users. Unlimited text access has been a paid-tier differentiator since the earliest days of ChatGPT; the move is a clear signal that OpenAI is repositioning the free tier as the top of a funnel for voice, image, and agentic features rather than as a metered introductory experience.
techcrunch.com — OpenAI brings unlimited ChatGPT text chats to free users
cathrynlavery/diagram-design: 29 Diagram Types, No Mermaid Slop
A repository of 29 editorial-style diagram types for Claude Code, built as self-contained HTML and SVG files, hit the trending chart after gaining nearly 8,000 stars. Install it as a Claude Code skill and every diagram Claude generates automatically inherits your site's color palette and fonts — producing clean, opinionated output rather than the generic default charting look. Each diagram type is a standalone file with no external dependencies or runtime JavaScript.
github.com/cathrynlavery/diagram-design
addyosmani/agent-skills: Engineering Standards as Reusable Agent Artifacts
Addy Osmani's production-grade agent skills repository crossed 86,000 stars, making it one of the highest-starred JavaScript repositories currently trending. It packages engineering best practices — code review protocols, performance budgets, accessibility checks, test-coverage standards — as skill files consumable by Claude Code, GitHub Copilot, and other AI coding agents. The framing is the interesting part: it treats the practices good engineers follow as versioned, distributable artifacts rather than as implicit knowledge each new AI session has to reinvent from scratch.