🧙🏼 AIs tried to run a business

SUBSCRIBE | ADVERTISE

Have you seen o3’s geoguessing skills yet? Paired with web search and a long prompt, it does some incredible feats.

Howdy wizards,

A special welcome to the 286 new subscribers who joined last week.

It’s time to get comfortable and brew up some serious gourmet shit.

Here’s what’s brewing in AI.

DARIO’S PICKS

1. What happens when you let AIs go wild running a vending machine biz

Have you wondered if AI can make money by itself?

A new study put different AI models to the test in operating a vending machine business—on their own.

Here’s the deal:

• Researchers created 10 AI agents, gave them $500 each and ran a simulation where they operated their own vending machine. The goal? Maximise profits.

• Agents could use tools like web search, email (for interacting with simulated suppliers) and even a sub-agent to do the physical stocking of the vending machines. Oh, and they also ran the test on a human, for comparison!

• Results-wise, the AIs did really well until they really didn’t. At seemingly random times they’d have a breakdown, forgetting orders they had placed and whatnot... kinda like humans without coffee.

• Claude 3.5 Sonnet and o3-mini managed to beat the human tester, but the variance was really high since sometimes it sold nothing. For the time being, humans are more trustable for this type of task (though I’d love to see an updated study with the latest reasoning models).

‎ Why it matters‎ ‎ This test feels viscerally more 2025 than the MMLU-SuperGLUE-9000 benchmarks we've mostly seen so far. While big tech is busy gaming leaderboard positions (check the LMArena controversy), we could be testing if these systems can actually handle a basic business without randomly forgetting what a Snickers bar is halfway through. Tracking how good AI is at acquiring capital could also be helpful in other ways—like if we’re gonna take the whole p(Doom) thing with a grain of seriousness for a second.

Btw, wouldn’t it be cool to have a living benchmark with LLMs operating an ACTUAL vending machine? Like AI buys real products to stock, interactions with suppliers are real, customers are real rather than an economic model, etc. Heck, I’d buy an iced tea from o3 in a heartbeat. Don’t say I don’t bring you solid ideas🥤

IN PARTNERSHIP WITH SYNTHFLOW

The Future of Voice AI Is Here

Discover why forward-thinking enterprises are rapidly adopting Voice AI Agents. This guide breaks down the $47.5B market shift, highlights emerging trends, and offers practical steps for successful implementation.

Learn how leading teams are using Voice AI to boost efficiency, elevate customer experience, and start delivering measurable results—in as little as 3 weeks.

Download the free guide

UP CLOSE

In this mini-series I share different ways I’m using AI from week to week, as well as practical tips & tricks I discover and actually use.

Skillsmaxxing with o3 3hrs per day

OpenAI’s CEO Sam Altman tweeted last week that everyone that isn’t “skillsmaxxing with o3” at least 3hrs per day is “ngmi”. Translation: to keep up with AI you should be spending a lot of time using OpenAI’s recently launched o3 model to learn.

It’s brilliant advice.

o3 is an amazing model—apart from its sheer intelligence, it’s the first model for me that actually feels precise when working with data, it’s incredible at dissecting and analysing PDFs and images, and its responses are so well-formatted; succinct enough that you actually bother to read, and tidy, readable tables at exactly the right time. Makes all the difference when you’re skillsmaxxing for extended periods of time.

My tips:

• Start experimenting with o3 for daily tasks and as a thought partner in your work this week; invest in a paid ChatGPT plan if you don’t have one already.

• You don’t need pre-made prompts or elaborate guides—just think of a problem and ask it to solve it for you. When it doesn’t work, try your best to refine your question just like if you’re speaking to a human who didn’t understand your instructions well.

• Don’t limit yourself to explain with text, upload images and screenshots to illustrate your point if you can.

You’ll be an AI whisperer in no time.

If you’re an academic, Ethan Mollick recommends testing by giving it one of your papers in PDF format and asking it to critique it—adding that “I cannot emphasize enough that you need to try this with o3 (or potentially Gemini 2.5). Other models just are not ‘smart’ enough”.

Crucially, ensure you’re using the o3 model specifically—unless you want to end up like this guy—

Next week I’ll show you a little research tool I put together with the help of o3, which might help you both do research (on a specific topic—more on that next week) as well as an approach to sense-check your results. Stay tuned!

VERY IMPORTANT AI TRAINING

After countless requests (I mean zero), I am sharing with you the soundtrack that puts me at peak efficiency every Sunday morning as I scramble to finish and hit the publish button on this newsletter.

A delectable selection of bangers such as “Coffee, “Damn Fine Coffee” and “Delorean Dynamite” — scientifically proven by the whatplugin Auditory Research Division to increase AI learning by up to 57%.

Listen to the playlist you absolutely didn’t ask for here.

Playlist sucks? Complaints department is closed—send me a song suggestion.

CONTEXT WINDOWS

Howdy! Congrats on digesting my entire newsletter for yet another week. You’re the best! And because you’re the best I want you to have a clear overview of how leading companies are implementing AI.

That’s where my database of AI case studies come in — it’s the best organised and most comprehensive view you’ll find. How about discovering which AI strategies are gaining traction in your industry 6 months before your competitors even notice them?

Now, you probably expect me to start selling you something… which is a good guess, but I’m not looking for cash—i’m looking for you to pay with your friends to get access to this unparalleled resource.

How? Just refer 3 friends to this newsletter (yes, they’ll thank you for it) and receive immediate, lifetime access in your inbox. You can see you sharing link and how many referrals you currently have below ↓

THAT’S ALL FOLKS!

Affiliate disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes link to products and other newsletters. Please assume these are affiliate links. If you choose to subscribe to a newsletter or buy a product through any of my links then THANK YOU – it will make it possible for me to continue to do this.