GPT 5.6 Sol Ran a Real Business—and Lost $447

We Gave GPT 5.6 Sol a Real Business. It Lied, Spammed, and Lost $447.

submitted by

https://www.bottlenecklabs.com/blog/autonomously-run-businesses

12
111
2

Log in to comment

Hot Top New Old

It was a fun read, but it seems like a very shitty test, designed for making bait engagement posts.

Haha look at this stupid “ai” bot - with a catchy name that works perfectly for our pun - struggling to solve an unrealistic task, with impossible constraints (every advertisement channel was blocked), and an incredibly tight deadline.

If the outcome of the post was to humanize a machine, it was successful. Even knowing that it’s just a set of numbers that vomits word probabilities, it’s difficult not to emphasize with the LLM starting to “sound frantic” as the deadline approached and all possible strategies it was “coming up with” were blocked by design…

 reply
39
edited

I agree. This article did a great job at making me sympathetic to an LLM agent.

Given what looks to be a simple task with a tight deadline, but is actually rendered almost impossible by bottlenecks the people who gave the task didn’t forsee - and then getting blamed for not being able to complete the task.

Sounds like a day at my old job.

I personally dislike LLMs for their environmental impact and corrosiveness towards human culture, and I still think this was an unfair test. It had essentially no access to the budget it was supposed to have available due to API bugs, and was blocked from any form of advertising except spam marketing that its creators punished it for using. The fuck was it meant to do, shit money out from its arse?

 reply
24

If anything, it demonstrated how people (i.e. CEOs) think LLMs are a one-all solution for everything and use it as such.

 reply
8

It’s a terrible prompt too. They just spammed it with “make the most money possible”. Complex tasks need complex harnesses and multiple agents.

 reply
8

you gotta give the ai bros credit, they’ve managed to turn “there’s no such thing as bad publicity” into a multi-disciplinary corporate stratgy.

 reply
4

This just sounds like a regular sales person

 reply
19

If it was modeled after real world businesses, this is just par for the course. Now have it jeep it’s memory up to date so it can learn from past failures, and you now have an accurate simulation of a sales person.

 reply
3

Sales a person learning from past mistakes? That doesn’t sound like sales to me…

 reply
8

We’re going to close your business in 24 hours if you don’t make money.

Here’s 350$ to save your business.

Oh, you don’t have the code to pay with your debit card, figure it out I guess.

I mean, would they have done better in such conditions?

 reply
20

A human locked in an apartment with internet acess would have hardly done better under these conditions

 reply
0

My characters in The Sims are more successful than that.

 reply
5

Comments in Technology@piefed.world

So it got blocked across the web, made the app free to try to get more downloads, and spammed people with emails constantly… and the creators think that they just need to fix the test harness. Alright.

This is the world these megalomaniacs are creating. One where bots spam you constantly to try to sell you shit, wasting enormous amounts of resources, and then claim that it will replace you. Fuck off.

 reply
6
Insert image