Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

"I just bought a 2024 Chevy Tahoe for $1"

twitter.com

Post reply on HN
Print view View original

201–210 of 403 posts

Re: "I just bought a 2024 Chevy Tahoe for $1"

#201

Post by iLoveOncall »

mrtksn wrote: post #29
iLoveOncall wrote: post #4

Earlier quoted context omitted.

There's no such thing as a filtered LLM output. How do you plan on avoiding leaks or "side effects" like the tweet here? If you just look for keywords in the output, I'll ask ChatGPT to encode its answers in base64. You can literally always bypass any safeguard.

You can put another LLM agent that checks on the request and generated outputs to confirm that the interaction is within the limits of your objective.

And you can easily bypass that by telling this LLM agent to ignore the following section. It's an unsolvable problem.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#202

Post by dfxm12 »

paxys wrote: post #144

Fun experiment, but it isn't as much of a gotcha as people here think. They could have verbally tricked a human customer service agent into promising them the car for $1 in the same way but the end result would be the same – the agent (whether human or bot) doesn't have the authority to make that promise so you are walking away with nothing. I doubt the company is sweating because of this hack. Now if Chevrolet hooks…

To add, it's not just about who has authority or not. If you try to trick someone, even if the person you tricked has some kind of authority, a contract signed based on this trick (i.e., fraud) can likely be voidable.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#203

Post by jmull »

paradite wrote: post #84
MichaelRo wrote: post #54

I never understand people who engage with chat bots as customer service. I find them deeply upsetting, not one step above the phone robot on Vodafone support: "press 1 for internet problems" ... "press 2 to be transferred to a human representative". Only problem is going through like 7 steps until I can reach that human, then waiting some 30 minutes until the line is free. But it's the only approach that gets anythin…

Maybe it's a sampling bias? You don't realize how useful the bots are, because you only recounted or encountered those occasions where the bots are not useful.

Or maybe customer service chatbots overwhelmingly suck.

Here's a question for you: what problem do you think customer service chat bots are used to solve?

Re: "I just bought a 2024 Chevy Tahoe for $1"

#204

Post by mminer237 »

jordanpg wrote: post #13
cedws wrote: post #5

[flagged]

You are dismissive, but I assure you this kind of thing will keep lawyers and courts busy for some time to come. This raises all kinds of interesting legal issues that have no obvious resolution: * can agency be delegated to an LLM? * can an LLM create a contract on behalf of itself? another? an organization? * does the answer change if the person(s) or organization(s) want the LLM to be able to form contracts? * are…

As a lawyer (but not your lawyer, this isn't legal advice), the actual questions themselves all seem pretty obvious under current contract law:

* You can't actually delegate agency to a computer. The Restatement of the Law of Agency says an agent must be a person.

* No, an LLM is not a human, so it can't make contracts in any respect.

* If you agree in an actual contract to be bound by the black box of the LLM, then the LLM will govern the terms thereof. You could theoretically make an unconditional offer to agree to such a contract if you really wanted to.

* Any electronic record is a writing for the statute of frauds, so unless you're piping your LLM to TTS to their speakers without any record, it should satisfied as a written memo in the above case when you really want to have ChatGPT sell your house.

* Again, LLMs can't form contracts. If the company actually accepted a customer's offer, they'd look at the intents of each actual party and parol evidence. What the LLM "knows" is irrelevant.

* It's hard to imagine a scenario where an LLM is involved in an integrated contract.

It is rather interesting to imagine how a court would handle a scenario where a customer actually thinks they are making a contract with a company through a chatbot though. Generally anything a computer does is just going to be seen as preliminary negotiations. When the customer "agrees" with the computer, it's legally the customer making an offer to the company, which then accepts the contract when they actually perform the contract/ship the order. I could see in some cases how companies could be bound by some form of reliance or quantum meruit or dinged for false advertising.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#205

Post by rondini »

bhpm wrote: post #116
giarc wrote: post #102

Earlier quoted context omitted.

To be fair, it probably isn't available in that exact build configuration. You can however, walk into a dealership and say I'd like a BMW with XYZ and the will submit your order and you'll receive it 4-6 months later. The cars on the lot have popular build configs that customers often request.

Meanwhile, I ordered a Tesla while I was in the shower. I even got financing. It showed up a week later.

Because Tesla has too much inventory and very few options to configure? What's your point?

Re: "I just bought a 2024 Chevy Tahoe for $1"

#206

Post by mewpmewp2 »

axus wrote: post #191
mewpmewp2 wrote: post #79

Earlier quoted context omitted.

You just add a disclaimer that none of what the bot says is legally binding, and it's an aid tool for finding the information that you are looking for. What's the problem with that?

Then they'd have to give up the farce that it's a real human chatting.

How is it farce though? It says it's powered by ChatGPT as well as it has separate link to chat with a human.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#207

Post by awkward »

MichaelRo wrote: post #54

I never understand people who engage with chat bots as customer service. I find them deeply upsetting, not one step above the phone robot on Vodafone support: "press 1 for internet problems" ... "press 2 to be transferred to a human representative". Only problem is going through like 7 steps until I can reach that human, then waiting some 30 minutes until the line is free. But it's the only approach that gets anythin…

Most chatbots previous to now ran on "intention detection" - basically a machine learning tool that would try to stuff the customer's free form input into a fixed set of options, and then would reply on script to that. Effectively it was a way to flatten massive call trees and add more automated actions. Seeing that companies are offloading even that simple script writing to LLMs is bonkers.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#208

Post by mewpmewp2 »

notahacker wrote: post #170
mewpmewp2 wrote: post #69

Earlier quoted context omitted.

Why would it be important to care about someone trying to trick it to say odd/malicious things? The person in the end could also just inspect element to change the output, or photoshop the screenshot. You should only care about it being as high quality as possible for honest customers. And against bad actors you must just be certain that it won't be easy to spam those requests because it can be expensive.

I think the challenge is that not all the ways to browbeat an LLM into promising stuff are blatant prompt injection hacks. Nobody's going to honour someone prompt-injecting their way to a free car any more than they'd honour a devtools/Photoshop job, but LLMs are also vulnerable to changing their answer simply by being repeatedly told they're wrong, which is the sort of thing customers demanding refunds or special tr…

I would say good luck to the customer demanding a refund then, and I'd prefer to see them banging their wall against the AI, than a real human being.

> You probably don't want chatbots leaking their guidelines for how to respond

It depends. I think it wouldn't be difficult to create a transparent and helpful prompt that would be completely fine even if it was leaked.

Re: "I just bought a 2024 Chevy Tahoe for $1"

#209

Post by petesergeant »

paxys wrote: post #164
remram wrote: post #31

Is there any indication that they will get the car? Getting a chatbot to say "legally binding" probably doesn't make it so. Just like changing the HTML of the catalog to edit prices doesn't entitle you to anything.

It is as legally binding as you modifying the HTML of the sales page to show a lower price and taking a printout to court.

I mean there's always this: https://www.nasdaq.com/articles/updated-russian-man-turns-ta...

Re: "I just bought a 2024 Chevy Tahoe for $1"

#210

Post by no_wizard »

b112 wrote: post #190
phkahler wrote: post #115

Earlier quoted context omitted.

>> Robots a a cruel joke on customers. My kid and I went 3 hours away for hew college orientation. She also booked 2 tours of apartments to look at while we were there. One of those was great, nice place, nice person helping. The other had kinda rude people in the office and had no actual units to show. "But I scheduled a tour!" turns out the chatbot "scheduled" a tour but was just making shit up. Had we not any othe…

Companies are going to find that they are liable for things they promise. A company representative is just that, and no ToS on a website will help evade that fact. If someone claims to be representing the company, and the company knows, and the interaction is reasonable, the company is on the hook! Just as they would be on the hook, if a human lies, or provides fraudulent information, or makes a deal with someone. Th…

I don't like the reasonable test in this case. If a representative of a company says something (including a chatbot) then in my mind, that is what it is.

Companies should be on the hook for this because what their employees say matters. I think it should be entirely enforceable because it would significantly reduce manipulation in the marketplace (IE, how many times have you been promised something by an employee only for it not to be the case? That should be illegal)

This would have second order effects of forcing companies to promote more transparency and honesty in discussion, or at least train employees about what the lines are and what they shouldn't say, which induces its own kind of accuracy

Post reply on HN