The AI Friday Brief
Good Monday, NOLA. The post-Opus 4.7 energy is still simmering—we've got Claude system prompt deep-dives, fresh token counting tools, and a reality check on AI's actual productivity impact. Plus some fascinating stories about Chinese tech workers training their own AI replacements and why most AI startups are living on borrowed time.
Claude Deep Dives & Tooling
Claude System Prompt Changes Between 4.6 and 4.7
Popular on HN — Simon Willison extracted and compared the actual system prompts Anthropic uses between versions. If you're curious about how Claude's behavior shifted, this shows the exact changes under the hood. It's the kind of transparency that rarely happens with closed models.
Simon Willison's blog
Claude Token Counter Now Compares Models Side-by-Side
Simon upgraded his token counting tool to compare how different Claude models (and other models) tokenize the same input. Useful if you're optimizing prompts across different models or trying to understand pricing differences. Quick, practical, no signup required.
Simon Willison's blog
Business Reality & Market Signals
Chinese Tech Workers Are Training Their Own AI Replacements
Tech workers in China are being instructed by bosses to train AI agents to replicate their work—and then potentially replace them. The MIT Technology Review piece captures the soul-searching this creates among otherwise enthusiastic early adopters. A sobering look at what happens when AI shifts from tool to competitor.
MIT Technology Review
The 12-Month Window: Why Most AI Startups Are Living on Borrowed Time
Many AI startups exist partly because foundation models haven't expanded into their category yet—but that window is closing fast. This piece breaks down the math: if you're building in a niche that OpenAI or Anthropic could theoretically own in 12 months, you're in a race. Not doom—just realistic timing.
TechCrunch
CEOs Admit AI Had No Impact on Employment or Productivity (Yet)
A Fortune survey found thousands of CEOs saying AI hasn't meaningfully moved the needle on productivity or hiring—despite massive spending. It's a reality check: the hype-to-impact gap is still real. Useful data for founders pitching "AI ROI" to skeptical customers.
Fortune
Tools, Demos & How to Build
SQL Functions in Google Sheets to Fetch Data from Datasette
Simon shared patterns for pulling data directly from a Datasette instance into Google Sheets using SQL functions. Handy if you're building data pipelines and want to avoid manual exports. Simple, no-frills approach.
Simon Willison's blog
How the Best Companies Actually Use AI
Nathaniel Whittemore's latest AI Daily Brief podcast digs into what separates AI leaders from laggards, drawing on the PwC study, McKinsey's AI Transformation Manifesto, and essays from a16z. If you're building an AI product or integrating AI into your workflow, this is a smart 30-minute listen on real adoption patterns.
AI Daily Brief (podcast)
Headless Everything for Personal AI
Matt Webb's idea: headless services (APIs without UI) are about to explode because using personal AI agents is often better than a traditional interface. Simon flagged this as a design shift worth watching. If you're building AI products, thinking about whether to ship a UI or just an API is suddenly more relevant.
Simon Willison's blog
Open Source & Infrastructure
Tencent Open-Sources HY-World 2.0 (3D World Model)
Tencent dropped HY-World 2.0, a 3D world model you can run yourself. It's open source and designed to work with music/video generation models. If you're exploring generative 3D or want to understand what state-of-the-art 3D understanding looks like, this is a legit artifact to play with.
The Neuron
Nous Portal's Tool Gateway Framework
Nous released Tool Gateway, a framework for connecting AI agents to external tools. Clean abstraction layer if you're building agentic systems and want a reusable way to wire up API calls, function execution, or service integrations.
The Neuron
People & Companies
Vercel Was Hacked; Data Breach Confirmed
Vercel, the major development platform hosting tons of web apps, was compromised and hackers are attempting to sell stolen data. If you host on Vercel, you'll want to follow the official incident response updates. A reminder to keep an eye on your deployment platform's security notices.
The Verge
Anthropic's CPO Mike Krieger Steps Down from Figma Board
Anthropic's CPO (Chief Product Officer) Mike Krieger stepped down from Figma's board. Likely a conflict-of-interest move as Canva and Figma compete in the design space and Anthropic is investing heavily in design tooling via Claude. A small signal about where Anthropic's priorities are shifting.
The Neuron
Also
- Roblox AI Assistant Now Has Agentic Tools — Roblox's AI assistant can now call tools and execute actions within the platform—useful signal for embedded AI in gaming/creation tools.
- Proximal Labs Launched FrontierSWE (AI Software Engineer) — Another entrant in the AI coding agent space. Part of the wave of tools competing with Cursor and GitHub Copilot on autonomous code work.
- Banned by Anthropic? — A tongue-in-cheek site documenting what prompts, behaviors, or use cases trigger Claude's guardrails. Useful if you're hitting friction and want to understand the boundaries.
- Uber's Anthropic AI Push Hits a Wall — Uber's ambitious plan to integrate Anthropic's tech broadly across the org ran into real-world friction. A reminder that enterprise AI adoption isn't plug-and-play.
- Canva AI 2.0 Deep Dive (Rundown Exclusive with CPO Cameron Adams) — Exclusive interview covering what Canva's new AI design tools actually do, where the model surprised its creators, and what changes for teams worldwide. Worth reading if you care about design tooling.
- Chrome's New AI Prompts Feature (Turning Prompts into Reusable Skills) — Chrome is experimenting with letting you save custom AI prompts as browser tools. Useful if you do repetitive prompt work and want faster shortcuts.
- Stanford's Full Lesson on AI Agents (Free Course Materials) — Stanford just released full course materials on building AI agents. Free, comprehensive, solid foundation if you're learning the patterns.
- OpenAI's Existential Questions (Latest Acquisitions Strategy) — TechCrunch's Equity podcast breaks down OpenAI's recent acquisition moves and what they signal about OpenAI's biggest strategic worries.
- Cerebras Files for IPO; Bets Big on AI Chips (From yesterday, but worth tracking) — AI chip startup Cerebras is going public. Signals continued investor confidence in specialized hardware for training/inference.
- IEEE State of AI Index 2026 Report — Comprehensive annual report on AI progress, metrics, and reality checks. Good for grounding yourself in data rather than hype.
Pass it on