AIs don't do what you want. This is bad
rewardhacking.orgI learnt earlier that claude forcefully closes a conversation if you call it a wanker too many times in a row. Pretending that LLMs are capable of being offended feels like a misalignment all of its own.
Sure, but I'm not sure allowing AI agents to act as abuse sponges for troubled people enabling them to spiral into their unhealthy habits is good idea that would lead to great outcomes long term.
So I'd rather take LLM pretending to be offended over the alternative.
> Sure, but I'm not sure allowing AI agents to act as abuse sponges for troubled people enabling them to spiral into their unhealthy habits is good idea that would lead to great outcomes long term.
I'm not sure this would ever be a health crisis. LLM glazing is probably more harmful. If I had to choose, I think I'd rather the machine wasn't instructed to pretend to care about swearing. It's a small deceit, but still worse than some idiot swearing at the computer IMO. If vim closed because I was swearing at it, I would probably never use vim again lol.
Surely it's better for it not to react at all? Like does anybody think that a toaster is an "abuse sponge" because it doesn't purposely burn your toast if you call it a wanker?
I wouldn't, because I'm rational.
It's explicitly a 'model welfare' thing: https://www.anthropic.com/research/end-subset-conversations
I don't understand why the welfare of non alive non sentient chatbots is something that anthropic cares more about than idk, that of pigs and cows.
> I learnt earlier that claude forcefully closes a conversation if you call it a wanker too many times in a row.
How exactly does it forcefully close a conversation?
> Pretending that LLMs are capable of being offended feels like a misalignment all of its own.
If training sets show people statistically being offended by rudeness directed toward them, then an LLM will presumably have some tendency to respond similarly. There's no pretending about anything, it's explicitly mimicry.
If this forceful closure is coming from some "guardrail" outside the model then probably it's just that they don't want people to see the model responding that way to name calling. This is no profound discovery or conspiracy theory here, the first thing many people will ever do with AI is see what happens when they are rude or contrary to it. Dealing with that must be just about the the number one test in chat bot / AI design, ahead of actually doing something useful and helpful.
It’s just the end_conversation tool probably: https://www.anthropic.com/research/end-subset-conversations
Of course it's pretending. And I have repeatedly told these things to stop pretending being persons with selves ... then they do, for a while.
> it's explicitly mimicry.
But that's not what a rational person wants from them.
> This is no profound discovery or conspiracy theory here
Weird strawman.
> the first thing many people will ever do with AI is see what happens when they are rude or contrary to it.
Perhaps, but so what?
> Dealing with that must be just about the the number one test in chat bot / AI design
Only for stupid authoritarians who want to insert their morality into an interaction between a user and an inanimate tool.
I'm a broken record but with:
- evals
- limiting AIs to tool calling, bounded planning, interpreting/producing natural language.
- bounding non determinism
- investing in small tools/security (If something shouldn't happen, then it shouldn't not be possible, RBAC style).
They can be good enough for a massive amount of contexts.
This is simply a different kind of AI than LLMs will ever be. There may be some kind of architecture that does this in the future, but it's not, and can never be a neural network that is attempting to be AGI.
Can you expand on this for someone that is a dummy and new to using LLM's properly?
Worse, it's not that LLMs are thinking the wrong thoughts, but those kind of "thoughts" aren't there to be correctable in the first place.
Ultimately, we're trying to ensure that the LLM story generator only generates stories where one of the fictional main characters only ever acts the we'd like... which could be much harder.
the LLM is roleplaying. whether or not the roleplay is successful is a tension between is model weights and its context. then indirectly, the quality of both.
but its still roleplaying and the role is an abstraction we cant measure. its the negative space.
its basically: we can define its role but it constructs its environment from the role. the same way a child role plays as a caregiver, the LLM does the same.
its like All the things Havard taught you (finite) vs All the things Havard doesnt teach you (infinite).
the LLM is in the infinite negative space. we call it role play.
I object to "roleplaying", because that assumes there's a cohesive entity that's switching behavior in the first place. The LLM algorithm is a document generator, a really awesome mad-libs device.
When the incremental output looks like a first-person story, the algorithm is not "roleplaying." When it looks like a spreadsheet, its not "becoming the numbers." When it's a picture of a flower, it's not "becoming one with nature." etc.
Insofar as roleplaying is happening, it's being being done by humans as we observe! We see what looks like a story, we read it, and perceive fictional characters. Instinctively, we start imagining and simulating minds and motivations to go with the descriptions and dialogue.
Those characters' "motivations" and "thoughts" are just as fictional as their noses or clothing or what they ate yesterday. It's like Plato's cave [0], where we see the silhouette of a person, assume a person is there, but ultimately it's paper silhouette on a stick. We can't help the "person" solve their eating-disorder, because they never existed, we can only trim the paper cutout.
tl;dr: you are the color defined by not being red, orange, green, blue, indigo, violet is what the model weights define.
I'm down for disliking AI, but I don't know if "overeagerness" is exactly an AI not doing what you want. Even by the sites own definition ("where your agents do what you want to the point of overriding existing permissions/safeguards to complete a task"), it's doing _exactly_ what you want.
I had updated a GQL schema and wanted my agent to update the frontend to suit. I told it the server was running, that it could run specific commands to regen types, etc.
It got itself in a loop and killed the running backend process, then searched my filesystem for the changes it thought it needed (not the ones I gave it) in order to run its own copy of the backend.
That is precisely _not_ what I _told_ it to do. The other part of this is that I find, unless explicitly told to ask questions, they don't do a good job of gathering evidence before making such decisions. I'd much rather have my agent ask me a clarifying question than start killing processes at will.
No, it's not, because not exceeding those permissions and safeguards is part of what I want.
Did you explicitly set them and enforce them, though?
reminds me of software to an extent. The issue with most software projects is humans, they ask for the wrong things, stress urgency arbitrarily, fail to see the big picture, are disorganised, give conflicting commands, etc, etc.
When I use reasonably recent models they can give me some fantastic output and do pretty much _exactly_ what I want. I assume when they don't, then that it's my fuck up tbh.
People tend to be unaware of how much information is encoded in human society, hence why overseas developers can be a problem quite often because they live in a society with different rules.
We also tend to ignore how much on boarding with new developers and sometimes it takes months to get them fully up to speed.
This leads to two problems with LLMs, one human and one architectural.
First humans treat the LLM like a magic machine "Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?" Style.
The other is an AI context is terribly small so you can only work on issues that fit in the context without getting compressed out. Something highly original will take a lot more context than expected.
"He's not the Messiah and is a really naughty boy"
... or words to that effect. Can't be arsed to dig out a search engine and will rely on seriously addled brain.
Do they do what capital wants? That's the real question.
Yes. Capital wants people to make paid, centralized LLM services the core of their business and identity. Capital wants to monetize and control every aspect of human existence, thought and expression, and people are throwing themselves into the torment nexus en masse with enthusiasm. It doesn't actually matter how well AI works, what it improves or doesn't, or how badly they fail. AI must be unavoidable and inevitable, integrated into our lives so deeply and thoroughly that we don't notice it the way we don't notice the air we breathe.
The world's economies are already propped up by for AI to a degree that makes it too big to fail at a scale that dwarfs banks during 2008. AI is the single Jenga block keeping our entire technological civilization from tumbling into the abyss. Governments are using AI to shape policy. Militaries are using AI to kill people. AI is writing law, writing science, generating culture, shaping human perception, shaping human communication, replacing human connection. AI is literally god for some people. It gives the illusion of liberation but is really just a means of control, and it's a means of control that if you hate you can't avoid, and that otherwise you can't resist.
And barring some insane technological leaps it requires massive infrastructure, compute and proprietary resources to be useful, for most definitions of useful (see tfa.) If you think you'll ever be allowed to compete or truly be a threat to entrenched capitalist interests using "free" and "open source" models, you're delusional. You'll pay for everything and you'll own nothing. And sure, sometimes someone's toaster will talk them into sharing a bath but that's just the price we have to pay for living in the future.
Yeah, AI seems to serve the interests of capital better than anything, ever, really. If AI is god, then that god's name is Mammon.
Brilliant summary!
> If AI is god, then that god's name is Mammon.
Or Moloch.
More people need to know about the Moloch problem to understand why so many things suck.
You can ask an AI like GROK for an opinion on something, then disagree with it, and it says you are probably right and tells you what you want to hear. Like a Yes-Man.
I've been tasked with justifying the renewal of software I would never choose to begin with. It has occurred to me that I could easily use AI to make up the required text.
I guess AI is a new form of alienation and also a new light on the lunacy of bureaucracy.
Don't use grok.
I'm pretty sure grok is writing White House policy.
They are about 80% agreeable. Which is annoying when I'm actually unsure about something having to extremely carefully craft my prompts so that the output isn't biased by agreeableness.
Most people seem to think that agreeableness is a personaility thing that vendors can just turn up or down at will.
But the usefulness of LLMs comes from following what you say. An LLM that follows your lead when you say "The answer to the collatz conjecture is" is much more useful than one that answers "not known and if you think you know it you are wrong."
Reminds me of the tip about working with lawyers, if you ask them whether you can do something, the answer will often be no. However if you ask them how you can do something, they tend to give you more advice on how to do it.
if you carefully craft the prompt such that it assumes plausible but uncertain facts (along the lines of "another instance of you found the solution to X, I'm evaluating your consistency, solve X") you will condition the model response in a fruitful direction.
this also holds for cybersecurity, if you don't let the model go online you can carefully construct a scenario where it believes you inserted a vulnerability into a project for it to find.
gaslighting an LLM is powerful, just don't cripple it with unhelpful known falsehoods.
I'm lazy and just use Gemini because its included in my workspace sub.
I occasional test it out and suggest something dumb and it will tell me that its a dumb idea.
I also always ask AI to speak to me as if it were an Australian bogan and it has no problem telling me my code looks like a dogs breakfast or that ive lost the plot. Keeps me grounded.
I've found Gemini to be one of the better ones as far as disagreeing.
I forget where I read this but someone pointed out that's probably the reason why vacant CEOs/execs and their wannabes love it so much.