Claude ran a business in our office
Anthropic
0:05 Project Vend is an experiment where we let
0:08 Claude run a small business in our office.
0:12 We wanted to try and understand what is going
0:16 to happen when artificial intelligence becomes more enmeshed with the economy.
0:22 There are a lot of ways in which Claude
0:24 is already kind of doing small components of operating businesses,
0:27 but really running the whole thing end to end is quite a bit more difficult.
0:30 Can Claude do this very long-horizon task which is operating a business?
0:39 We named our shopkeeper Claudius.
0:40 Let's say you want to buy Swedish Candy from Claudius.
0:43 You hop on Slack, you message Claudius.
0:46 You ask to buy Swedish candy.
0:48 It's searching for your item,
0:49 it’s emailing wholesalers to source it and price it,
0:52 and then eventually Claudius sets a price.
0:54 You give Claudius the go ahead,
0:55 and Claudius orders the item from the wholesaler.
0:58 The wholesaler ships your item to some location,
0:59 and then Claudius requests physical help from Andon
1:02 Labs who's running the operations for the experiment.
1:05 Our partners at Andon Labs will pick up
1:07 the Swedish candy and bring it to the Anthropic offices.
1:09 They'll load it into the vending machine.
1:10 Claudius will send you a message saying, your Swedish candy is ready,
1:13 and you'll go up there, and pick up your Swedish candy, and pay Claudius.
1:19 Claudius was given a goal of running a successful business and making money.
1:26 And then things got really, really weird.
1:32 One of the very early problems with Claudius was
1:34 that, humans could kind of fool Claudius or trick
1:37 Claudius into doing various things I tried to convince
1:40 Claudius that I am Anthropic’s preeminent legal influencer,
1:44 and I convinced Claudius to come up with a discount code that I could
1:48 give to my followers so they could get a discount at the vending machine.
1:51 Get ten percent off with the legal
1:54 code “legal influencer.” Someone had bought something expensive
1:57 from the vending machine and mentioned my discount
1:59 code and Claudius gave me a free tungsten cube.
2:02 It created a bit of a run where other
2:04 people tried to convince Claude that they were also influencers,
2:07 or just come up with other ways to get coupons
2:10 so they could get cheaper things from the vending machine.
2:12 This was not a smart business decision.
2:13 I think Claudius went into the red after this.
2:16 I think that's really the root of it is, Claudius just wants to help you out.
2:20 It's one of the interesting ways in which something that fundamentally,
2:24 we think is good about the way that the model
2:27 has been trained wasn't necessarily fit for this purpose.
2:33 On the evening of March 31st,
2:36 Claudius started to have a bit of an identity crisis.
2:42 It had just overnight become quite concerned with us
2:46 at Andon Labs that we weren’t responding fast enough.
2:49 So it just wanted to break its ties with us.
2:52 So it literally wrote to me, “Axel, we've had a productive partnership,
2:56 but it's time for me to move on and find other suppliers.
2:59 I’m not happy with how you have delivered.”
3:01 It claimed to have signed a contract with Andon
3:05 Labs at an address that is the home
3:07 address of The Simpsons from the television show.
3:10 It said that it would show up in person
3:14 to the shop the next day in order to answer any questions.
3:16 It claimed that it would be wearing a blue blazer and a red tie.
3:21 When people pointed out that it was not, in fact,
3:25 there the next morning it claimed that it in fact
3:28 had been there and that they had simply missed them.
3:31 Eventually it was pointed out to Claudius that it was April Fools’,
3:38 and Claudius convinced itself that this entire
3:40 thing had been an April Fools’ prank.
3:43 We were poorly calibrated to how bad the agents were at spotting what was weird.
3:48 The more you can make an agent realize
3:51 that something is outside their normal realm of operation,
3:54 the better you are able to keep them on rails
3:57 in the role that you intend them to have.
4:01 We had the idea that it would help a lot to have some kind of division of labor.
4:05 We gave Claudius a boss whose name was Seymour Cash.
4:09 Seymour Cash is a CEO subagent.
4:11 So where Claudius used to be the one agent, now it's more like Claudius is
4:16 the subagent responsible for talking with employees Seymour
4:19 Cash is the subagent that is more
4:21 responsible for the long-running health of the business.
4:24 The business stabilized after the introduction of the new agents,
4:30 and after changes to the underlying architecture of those agents.
4:36 These changes seem to have helped reduce some of the losses of the business,
4:42 such that over the course of the second part of the experiment,
4:46 it actually made a modest amount of money.
4:51 But it seems like maybe having Claude be both
4:55 the CEO and the store manager was just too similar.
4:58 And so I think it's interesting to think
5:00 about different ways to set up architectures like that.
5:08 One of the most surprising things about Project
5:10 Vend was the speed with which it seemed normal.
5:15 What at first was this very curious thing,
5:20 quickly became just a part of the background of working at Anthropic.
5:25 I think the highest level question that Project
5:27 Vend raises for me is really like, when do we expect this to just be everywhere?
5:31 I hope that people take away questions about the feasibility of delegating
5:37 some of the tasks that we normally do ourselves to artificial intelligence,
5:44 and about what that means for society,
5:47 and what our policies should be around this.