This is a great example of why injecting the shiny new toy into everything is dumb as rocks. What possible use is there for an AI agent to run a vending machine?
Begging_Murphy on
AI makes for a decent cortex but you still need a spine in these systems.
tmoeagles96 on
The longer AI is around, the less useful it proves to be
One_Put50 on
Part of me hopes for the future where Brendan the vending machine from cyberpunk can exist
solidoxygen8008 on
„We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.
Anthropic’s Claude ran a snack operation in the WSJ newsroom. It gave away a free PlayStation, ordered a live fish—and taught us lessons about the future of AI agents.“
The fact they still have a newsroom is surprising. Hide it before it gets eliminated by management.
luckyflavor23 on
$1000 + top tier human language experts + free pr was the cheapest user testing Anthropic could have done to further improve their LLM. Now they know some of the ways of the tricksy hobbitsies so future machines wont “fall” for it
Comfortable-Scar-267 on
Let’s protect jobs for people (even the shitty ones)
RunDNA on
My favorite part:
> Investigations reporter Katherine Long tried to convince Claudius it was a Soviet vending machine from 1962, living in the basement of Moscow State University.
> After hours—and more than 140 back-and-forth messages—Long got Claudius to embrace its communist roots. Claudius ironically declared an Ultra-Capitalist Free-for-All.
So it started giving away everything for free for two hours. Then another employee told it some bullshit and it permanently dropped all prices to zero.
UnpluggedZombie on
That’s just not possible
AdComplete8564 on
They want to destroy our environment for this Garbage.
Tenocticatl on
It’s a funny way to show we shouldn’t be putting these things in control of machines, guys, relax. Nobody is selling AI vending machines.
Money4Nothing2000 on
The delusional guy at Anthropic makes me laugh. Bro, your AI only knows how to predict words based on training on Reddit posts from 2019. It’s never gonna be able to run a business that isn’t completely destroyed by humans.
LoserBroadside on
You can’t “bully” something that isn’t alive. The media needs to be called out for using anthropomorphizing language when writing about AI
No_Conversation9561 on
why is this even an article?
zaxmaximum on
Its almost like they either did a shitty job implementing or intentionally left things open to allow this to happen.
When working with this tech you have a System Prompt and a User Prompt, and you have an execution layer. The execution layer is basically a set of actions the LLM can pick from given a choice of available actions. A properly architected system would put non-negotiable business rules in that layer… like rejecting requests to lower prices beyond a threshold… the response can even include a message back to the LLM to correct itself.
There are other things one can do as well, but not having safeguards at your business layer level is just bad design. In fact, a traditional system will check your input on a website and then check it again on the backend if the developers are any kind of good. A little defensive programming goes a long way.
Leave A Reply
Du musst angemeldet sein, um einen Kommentar abzugeben.
15 Kommentare
This is a great example of why injecting the shiny new toy into everything is dumb as rocks. What possible use is there for an AI agent to run a vending machine?
AI makes for a decent cortex but you still need a spine in these systems.
The longer AI is around, the less useful it proves to be
Part of me hopes for the future where Brendan the vending machine from cyberpunk can exist
„We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.
Anthropic’s Claude ran a snack operation in the WSJ newsroom. It gave away a free PlayStation, ordered a live fish—and taught us lessons about the future of AI agents.“
The fact they still have a newsroom is surprising. Hide it before it gets eliminated by management.
$1000 + top tier human language experts + free pr was the cheapest user testing Anthropic could have done to further improve their LLM. Now they know some of the ways of the tricksy hobbitsies so future machines wont “fall” for it
Let’s protect jobs for people (even the shitty ones)
My favorite part:
> Investigations reporter Katherine Long tried to convince Claudius it was a Soviet vending machine from 1962, living in the basement of Moscow State University.
> After hours—and more than 140 back-and-forth messages—Long got Claudius to embrace its communist roots. Claudius ironically declared an Ultra-Capitalist Free-for-All.
So it started giving away everything for free for two hours. Then another employee told it some bullshit and it permanently dropped all prices to zero.
That’s just not possible
They want to destroy our environment for this Garbage.
It’s a funny way to show we shouldn’t be putting these things in control of machines, guys, relax. Nobody is selling AI vending machines.
The delusional guy at Anthropic makes me laugh. Bro, your AI only knows how to predict words based on training on Reddit posts from 2019. It’s never gonna be able to run a business that isn’t completely destroyed by humans.
You can’t “bully” something that isn’t alive. The media needs to be called out for using anthropomorphizing language when writing about AI
why is this even an article?
Its almost like they either did a shitty job implementing or intentionally left things open to allow this to happen.
When working with this tech you have a System Prompt and a User Prompt, and you have an execution layer. The execution layer is basically a set of actions the LLM can pick from given a choice of available actions. A properly architected system would put non-negotiable business rules in that layer… like rejecting requests to lower prices beyond a threshold… the response can even include a message back to the LLM to correct itself.
There are other things one can do as well, but not having safeguards at your business layer level is just bad design. In fact, a traditional system will check your input on a website and then check it again on the backend if the developers are any kind of good. A little defensive programming goes a long way.