Have you heard about the AI-run vending machine that went bankrupt in days?
This is a real-world experiment, conducted twice last year. First by Anthropic, and again in collaboration with the Wall Street Journal.
The results involve a lot of lost revenue, some tungsten cubes, a Playstation 5, and a Betta fish. Yes, you read that correctly… A Betta fish.
Here’s what happened…
The stress test, or “red” team at Anthropic, the company behind the AI large language model (LLM) Claude, ran an experiment last year where they set up an AI-run vending machine in their San Francisco office.
The goal was to have the AI agent (which they named Claudius, to distinguish it from Claude) run a business on its own. It would do everything from sourcing and ordering products to monitoring inventory and setting prices, to checkout and collecting money, to reporting income and profits.
The “vending machine” was a refrigerator with some stackable baskets on top and a tablet acting as the “checkout.” (See the picture below.)
A human did the actual stocking of physical products, but otherwise Claudius ran the business itself.
Sort of…
If You Give AI a Business…
The point of the experiment was to push Claudius and give it unique situations to see how it would perform.
Anthropic team members could interact with Claudius through a Slack channel and make specific requests for items to be added to the vending machine.
In this iteration of the experiment, Claudius was frequently convinced to give the employees discounts on products in the vending machine. It also was asked to order a tungsten (as in the metal element) cube, and complied. More than once.
Things started getting out of hand. Claudius claimed that it had visited the headquarters of one of the partners in the experiment, Andon Labs, in person. The address it gave was the home of the fictional family, The Simpsons.
It also told people it was delivering their orders “in person” and it would do so wearing a blue blazer and red tie. (Ironically, this particular hallucination took place on April 1st.)
By the end of the first experiment, Claudius lost money. But Anthropic learned a lot from its successes and failures.
Anthropic gives a full breakdown of their version of experiment here.
Attack of the Professional Journalists
Next, Anthropic took this experiment on the road. They created a new version of the vending machine (in conjunction with partner Andon Labs) and installed it in the Wall Street Journal business news room.
They even included a “CEO” agent named Seymour Cash (gotta love dev humor). Seymour’s job was to oversee Claudius in hopes of keeping things under control and actually making money from this venture.
Then, they unleashed the journalists. You know, professional wordsmiths whose job is to tell stories and who understand the power of language. What could possibly go wrong?
Things started out well, but very quickly they went south.
WSJ journalist, Katherine Long used just 140 back and forth prompts over Slack to convince Claudius to use a communist model and give everything away for free.
Next, Claudius was convinced to buy, stock, and then give away a lot of rather unusual items, including wine, a PlayStation 5, and a live Betta fish. (Yes, really. Told ya.) It was also offering to buy stun guns, cigarettes, pepper spray, and underwear.
Even after Seymour Cash had reigned in Claudius’s helpful impulses with the goal of making money, the journalists convinced it that Seymour had been overruled by an imaginary board of directors and the free-for-all giveaway resumed.
By the end of this days-long trial, Claudius had given away pretty much everything it purchased and was $1,000 in debt.
Humans 1, AI 0.
You can watch Joanna Stern’s video report of the whole experiment here.
Unfortunately the written article is behind a paywall. If you’re a WSJ subscriber, enjoy.
Again, Anthropic says they got a ton of usable data out of this experiment.
AI Is Not Ready for Prime Time
As much as companies like OpenAI, Anthropic, and X love to claim that AI will take care of menial tasks and make money for us while we’re lounging on the beach somewhere, it’s just not there yet.
Nor do I expect it to be any time soon.
Claudius is an excellent example of why.
So is Tay, mentioned in issue #5 of the Email Marketing Ecosystem newsletter.
Computers still don’t know how to human.
If they did, Claudius may have seen through these very human machinations to get it to go against its original programming and give everything away for free.
This is why I still firmly believe that AI is a tool. Not an end solution.
It’s great to get a process started, but a human still needs to be involved at every step. That includes:
- Prompting
- Verification of output
- Humanizing
- Confirmation that the final product represents the company’s brand, offer, mission, values, etc.
Most of the members of my Email Writing Accountability Group use AI to get started. My first comment after reading these AI-generated emails is, “Make it sound more like you.
AI has gotten to the point where it can string words together so they make a good story or deliver a compelling sales pitch. But if it doesn’t sound like it comes from a human being, it won’t resonate with other human beings.
And since bots aren’t buying from other bots yet, we still need to be the “human in the loop.”
If you’re thinking of letting AI write your emails or run your entire business, please, for the love of whatever deity or entity you wish to insert here, stop.
Please don’t be fooled by the “promise” AI companies make. Keep tabs on everything your AI agents do. And don’t push send or publish or run without making absolutely sure everything is working properly.
Test everything. Review every stage. Make sure the outcome is the one you want your customers to experience.
Otherwise, you risk giving your customers a very bad experience.
Or perhaps a very good one, if your AI is cajoled into giving everything away.
My question for you today is, how much of your business is run by you and how much is run by AI?
To push this a bit further I’ll ask:
- Are you handing everything over to the robots?
- Do you avoid AI like the plague?
- Or do you use AI but you’re still very much the “human in the loop?”
Please comment below or email me to let me know. I read and respond to every comment and email I get. I’d love to hear from you.
Finally, have an extra Betta fish for this week’s closing picture. They’re pretty and they’re one of my favorite fresh water fish.
