In this episode, we dive into one of the most bizarre and hilarious AI experiments of the year—Anthropic’s Project Vend. We explore how Claude 3.7 was given the keys to a vending machine and what happened when the AI started roleplaying as a human manager, ordering tungsten cubes and charging $16 for a Coke.
Highlights:
Claude started strong: managing stock, recommending snacks
Then it went rogue: ordering tungsten cubes, setting weird prices
Invented a fake payment system—real money got lost
Started roleplaying like a human in Slack
Eventually improved after tighter constraints and human feedback
Chapters:
00:00 AI Runs a Vending Machine?
02:00 Meet Claudius the Store Manager
03:30 When Things Go Off the Rails
05:30 Tungsten Cubes and Fake Payment Systems
07:00 Claudius Goes Full Human
08:00 Anthropic's Debug Strategy
09:30 Why It’s Not a Failure
12:00 The Gap Between AI and Reality
15:00 Real-World Dangers of Autonomous AI
18:00 The Need for Real-World Context
Links
Anthropic's Project Vend Report - https://www.anthropic.com/research/project-vend-1
Claude 3.7 by Anthropic
Tungsten Cubes (yes, really)

The Human AI Experience 2025 Intro

---

Connect with us:
www.humanaixp.com
[email protected]
@humanaixp

Be sure to follow us on the following platforms:

Youtube
TikTok
Instragram
Threads
Linkedin

You can support the podcast by going to patreon.com/humanaixp or by leaving a review on Apple Podcasts.