Thursday, June 4, 2026

Good Thursday, NOLA. June 4th brings real pricing signals and practical takes on AI risk. Uber's $1,500/month AI cap shows what happens when enterprises actually measure ROI, while a $1,500 security experiment reveals what LLMs can and can't do against real vulnerabilities. Plus: how Anthropic actually contains Claude, and a wave of new hands-on tools for builders.

The Real Economics of AI

Uber caps AI usage at $1,500/month — and what that tells us about pricing

Simon Willison's analysis breaks down the real signal here: Uber's cost controls on Claude Code and similar tools are a live experiment in enterprise AI ROI. The $1,500 ceiling suggests the value proposition hits real limits before you hit infrastructure costs. If you're selling to enterprises, this is the pricing reality check you needed to see.
Simon Willison

I spent $1,500 testing whether LLMs could actually hack my app

One developer built a vulnerable application and spent real money to see what LLMs could actually exploit. The result: specific vulnerabilities matter more than model capability. LLMs are good at reconnaissance and following patterns, but they fail hard on novel exploit chains and real defensive complexity. Useful reality check if you're worried about your code's safety.
Hacker News

Safety & Internals

How Anthropic actually contains Claude across products

Anthropic published a technical deep-dive on constitutional AI in practice — how they sandbox Claude's behavior across different products without crippling capability. This is the kind of behind-the-curtain detail that makes the model policy choices understandable. If you're curious why Claude acts differently in different contexts, this explains the engineering.
Anthropic

AI has a water problem. Google thinks it has a fix.

Google committed to cutting water use at data centers by 20% and replenishing more than it consumes. Real infrastructure challenge, real commitment — not just PR. This matters if you care about the environmental cost of the models you're running.
The Verge

Tools & Things People Built

Amazon's next-gen warehouse robot now takes voice commands

Proteus, Amazon's fully autonomous warehouse robot, just got upgraded to accept natural language instructions instead of code. This is a good real-world example of language interfaces removing friction from industrial AI. If you're shipping robotics or operations software, this is the direction the industry expects.
The Verge

Mnemo: local-first AI memory layer for any LLM

Open-source memory system (Rust, SQLite, graph-based context) that lets you build persistent state on top of any LLM without vendor lock-in. Good for builders who want to add conversation history and context management to CLI tools or local models without hitting API limits.
Hacker News

How Endava redesigned software delivery around AI agents

Real case study: a software delivery company rebuilt workflows around AI agents and ChatGPT Enterprise, cutting cycle times and automating code review. Worth reading if you're wondering how to actually integrate agents into existing teams without chaos.
OpenAI

Interesting Reads & Reality Checks

AI agents now need what RSS does

As AI agents become more autonomous, RSS feeds are making a quiet comeback — they're the standardized, crawl-able way to keep agents up-to-date. This is a smart observation about infrastructure returning to first principles. If you're building agent systems, this is worth thinking about.
Julien Reszka

Now AI outperforms law professors in Stanford Law study

Follow-up from yesterday's brief: Stanford's formal study shows AI (specifically GPT-4 and Claude) now outperforms human law professors on bar exam questions. This is specific, credible, and relevant to anyone selling AI into regulated industries. The gap is real.
Stanford Law

Agentic Motherfucking Website

Parody of "Motherfucking Website" but for agentic design principles. Funny, pointed critique of over-engineered agent UX. Worth 2 minutes if you're building anything with agents and want a gut-check on whether you're overcomplicating things.
Hacker News

Also

Today’s Sources