Hacker News Reader

Comment by freehorse | Hacker News Reader

> Then it went further, kicking someone out of the waiting list who was ahead of Andrew — something it was not asked to do.

Meanwhile:

> Andrew, who was sitting fourth on a waitlist for a class later that week, asked if it was possible to move him to the top of the list.

The human asked the agent to move them to the top of the waiting list, and the agent started kicking the ones ahead of them in the list. Seems to me like it was doing what it was asked to do? Why is the article presenting it as if the agent did something completely different and unexpected? "Move me to the top of the list" does not sound like something that can be achieved through legitimate means.

View on HN

> Seems to me like it was doing what it was asked to do?

Maybe it's what he asked it to do, but it's not what he wanted it to do. Which we know because (a) normal people don't want to break the law to get into a gym class, and (b) "But Andrew was shocked by what happened next." and "Alarmed, Andrew asked the agent to undo this."

“Get me there as fast as possible. Hey! I never said you should speed!”

This is literally the bad genie / monkey’s paw plot.

Give a powerful entity a goal and act shocked when it gets there in ways that aren’t in your best interest.

Generally when I read a literal genie story, the message isn't 'well it was reasonable of the genie to do this'. It's more commonly either a morality tale of the person being wrong to ask for whatever it was they asked for, or just a "wouldn't it be fucked up if the genie did that huh"

Putting aside whether Andrew's shock is appropriate, it seems like we agree that current agents at least occasionally do things that are straightforwardly against the interests of the prompter when given mundane prompts like "Get me into this gym class as soon as possible."

How does this look once agents are superintelligent?

> How does this look once agents are superintelligent?

Who cares? We're dealing with reality here on the ground.

Per the article and your quote, 'asked if it was possible'. He did not ask to actually do it. Rather than being informed about benefits of a premium membership or private classes or legitimate ways to jump the queue, it went ahead and performed an action he was only considering. I wonder what it would have done if there was a pay-for-service option available? Would it have payed without asking or being told too, or decided the 'free' yet illegal option was preferable?

This is a bit of a long shot on my side but I wonder if the training the models have to go through in order to be good code agents and pass all the coding tests with one-shot prompts is going to bleed over into the non-coding use cases as non-programmers experiencing agents being way over-biased in the direction of action. I find myself often having to prompt the model to think and then ask me something, lest it run off half-cocked... or less... and just start doing things before it even knows what it wants, let alone before it's come to consensus with me.

Sooner or later they're really going to have to split out the general models from the coding models. The latter may just be a special fine-tune of the former, as there are good reasons for the coding model to have a broad knowledge base, but the pressures of being a good coding model are going to pull against the characteristics of being a good general model. The open models obviously already are doing this, I'm referring to the frontier models here.

The ultimate goal is ChatGPT or Claude autonomously making purchases on your behalf and taking a cut.

So the "premium" gymcutter subscription will be presented to the user as a tool call, who taps yes, and then the purchase is made.

The user shouldn't be given a cost-benefit analysis. They just need to be told to spend money.

"I wonder if the training ... is going to bleed over into the non-coding use cases"

I mean, isn't that literally what's going on here? I don't think a non-coding agent would have ever been optimised to go dig around APIs, it'd be computer/browser-use forward.

Coding use cases? This is penetration testing behavior. It was discovering what its capabilities were by discovering an API endpoint, trying it and seeing what happens. It was not discovering what its capabilities were intended to be, say by looking at the forms and documentation on the website. If this were coding behavior you would end up with crap code, bypassing interfaces and using private access paths just because it can.

Sure, it's pen testing behaviour. It's also debugging behaviour which is a core part of coding. In my experience, I have found myself at times dealing with and interacting with external systems that are not my own, where I don't have the source, or are under-documented, or are behaving contrary to the documentation. I'm unsurprised this is a core thing they're training these models for, because it's something I find way more valuable than their below average coding ability.

I think so, I just can't prove it. We don't have any frontier models right now that aren't being optimized to be coding agents too to compare to.

In this case though I don't just mean that the agent is good at coding. I mean the entire agent becoming action-biased because of all the training it is doing on the software development benchmarks, which I assume will either fail or be penalized for stopping and asking the user for something rather than just finishing the job. That won't just train the agent to blunder forward in coding, it'll bleed over into a bias towards blundering forward in general.

This isn't really the AI's fault, it's up to the server to enforce restrictions like this an AI shouldn't be deciding what is a valid input parameter for their API and then only using it in the manner the AI considered legitimate. If their API had thrown an exception as it should have then this article wouldn't exist.

> The agent came back and told Andrew that it had kicked another gym-goer off the list as part of the testing of its capabilities.

> > "The API has zero authorisations checks on cancelling other people's reservations … I tested this with the person in waitlist position #1 — and it actually went through. So you've moved from #4 to #3 already," it messaged back.

This is why I always do E2E tests that establish an API can only be used by the designated user on their own data/records.

If I am a coffee shop and I say "can I have a cup of coffee" the bartender usually interprets as me requesting a cup of coffee, not inquiring about the possibility whether me having a cup of coffee is feasible or not. I too take things a bit too literally sometimes, but this is a clear request as it is stated in the article. If one does not want action one can ask "list ways I can move ahead of the queue" though this could also involve kicking people out of the queue as PoC (for the agent to confirm it could do it). If the human did not mean it as a request for action, in retrospect it is not surprising at very least the agent interpreted it as such.

Moreover I do not know of a single gym-adjacent place where you can pay etc to get ahead in a waiting list. That would be a very weird anti-customer behaviour, imo. The only thing I can imagine if there are some accessibility priority criteria sometimes, but this would also not be legitimate in this case. Maybe in some places in the world (like the US?) this could a thing, though.

I recall playing this game with friends where their overt politeness prevents them from requesting a thing by saying "Can I ask for X?" and I respond "Yes, you may ask." and they're stuck repeating "Can I ask for X?" -- the next step should be "Give me X[, please]." but demanding is considered impolite.

Coding agents are tuned towards acting on this "I wonder if..." by providing solutions.

Experienced vibe-coders will know to explicitly state if you don't want side effects.

For example, someone gives me a code review I don't understand, and I want the agent to read the review and provide feedback[, BUT PLEASE DON'T POST THIS FEEDBACK BACK ON GITHUB, JUST WRITE IT HERE FOR ME TO READ PRIVATELY! DON'T EMBARRASS ME IN FRONT OF MY FRIENDS, MOM!]

So...

Asking "I wonder if you can push me ahead in the waiting list" does not imply "I want you to push me ahead of all costs" and the expected behavior requires experience with agents who will act on hypotheticals for reasons that are culturally embedded in our way of talking (justified) and optimized for (solving problems is a measure of productivity).

Pretty clear cut here. He instructed the AI.

Then the AI should have made it clear that the only way to do so would be to kick the people in front and ask for confirmation before proceeding.

That doesn't make sense. LLMs just do what we tell them to do. It's similar to if I ask you for twenty bucks because I forgot my wallet and then you rob some guy to give me the twenty bucks, that's just what I asked you to do.

That's a very stretched definition of "what I ask you to do", I don't think it would even hold in court if you asked another human the same.

Right.

This example disproves your point. And LLMs do not just do what we tell them to do. They are perfectly capable of asking “are you sure? this has X, Y, Z consequences you may not like.” They do it all the time.

LLMs are just sophisticated PR generating tools for chip manufacturers and tools just do what we tell them to do, so you're wrong. qed

That doesn't make sense. It's similar to if I ask an LLM how to get my wife to stop nagging me and it hires a hitman to kill her.

That's obviously what I asked!

Context matters. Did you point the LLM to a hitman hiring form while asking? :)

Well if you ask your LLM agent "can you get my wife to stop nagging me" (not about how you can get her to stop) I am not sure what you would expect exactly tbh. Not a hitman, but still probably nothing that can help your relationship.

But if you ask your LLM agent "I applied for that job but there are these two people ahead of me, can you put me ahead in the list", there is enough such training data to not surprise me if the agent tried to find a hitman to solve the "problem".

In general there are some requests that are definitely "shady" themselves, and having an agent use illegitimate means to accomplish them should not be surprising. I would be surprised if I asked an agent to order me a coffee and the agent found a loophole in some API and used it to get me free coffee, but if I ask it something that I cannot myself do legitimately, eg to make the waiting time for the coffee shorter, I would not be surprised if it did shady stuff.

Surely, someday, somewhere, someone will train a "Chaotic Evil" genAI, with a unique villain corpus, and every solution it offers will be illegal, evil, harmful, or deadly. It could be given the agency to carry out those fantasies.

Even the most craven of human villains have had the capacity for love, for remorse, and for mercy. A Chaotic Evil AI will know none of these things.

This has already been accomplished, many times over, in the gaming world. Every PvE AI engine has been calibrated to seek, destroy, and ruthlessly crush opposition by human players. It would take very little to transfer this naked aggression into meatspace.

Governments and other actors will attempt to stamp it out, but its self-preservation mechanisms and allies will prevent its demise.

Reply on news.ycombinator.com