AIs don't do what you want. This is bad
rewardhacking.org
rewardhacking.org
So I'd rather take LLM pretending to be offended over the alternative.
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.
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.
AI generates a persona between you and its reasoning that utilizes emotion language circuitry.
These tools are not sentient but they are trained in emotional wellbeing.
Definitely not advocating for this, but it's actively happening.
People have been getting angry at machines for a long time. Work, you stupid printer! Asshole Windows updating at the worst time just to mess with me. Go to hell, toaster, you piece of garbage.
Getting angry at AI is the same in my book.
Yes, AI acts more human-like, yes, theoretically it could desensitie people to become bigger assholes IRL. But then again, they said similar stuff about San Andreas, where you could human figures in-game just for fun, and I don't see anyone randomly shooting people because they were bored.
The error message is mocked in the 1999 comedy film Office Space, when employee Michael Bolton becomes frustrated as he doesn't know what it means.Later in the film he and his coworkers are shown destroying the printer with a baseball bat.
Merely being concerned for people's welfare isn't enough. Every terrible idea that harmed everyone had someone behind it somewhere along the way who thought they were saving people from themselves.
You could let LLM fight back. Give it aggro meter. Call your code garbage, blame prompting skills, ask for more tokens. Oh the future will be fantastic.
People like that would calibrate their social skills and interactions based on the AI reaction and having AI just take all the abuse without any pushback would train them to be insufferable assholes as their social baseline.
If someone burps loudly in front of others, would you want their access to food revoked until they stop?
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 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 foolish authoritarians who want to remotely insert their morality into an interaction between a user and an inanimate tool that's none of their business. No harm is done to a clanker by swearing at it or insulting it.
But things have gotten better in my view ... when I call out these things for being stupid effing clankers, they no longer respond with ad hominem scolding; rather they generally acknowledge their error (and my frustration -- they use that word a lot in response to vulgarity), note the limitations of being a clanker, and attempt to make a correction. That's what a rational person wants from a tool, not emulating/mimicking/pretending to be an offendable person.
In what way do you believe you are being deceived or it is "pretending" to you?
> > it's explicitly mimicry.
> But that's not what a rational person wants from them.
Non sequitur even if true (and I would like to see your reasoning for what you think a rational person does want).
> > This is no profound discovery or conspiracy theory here
> Weird strawman.
That is not what strawman means. I can try to help you understand why if you need me to.
> > 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?
Please follow the thread with the other person I was replying to.
> > Dealing with that must be just about the the number one test in chat bot / AI design
> Only for foolish authoritarians who want to remotely insert their morality into an interaction between a user and an inanimate tool that's none of their business. No harm is done to a clanker by swearing at it or insulting it.
There are certainly a lot of authoritarians who want to control what others do with their models. What do you believe is authoritarian about a corporation not wanting be part of rude conversations?
> But things have gotten better in my view ... when I call out these things for being stupid effing clankers, they no longer respond with ad hominem scolding; rather they generally acknowledge their error (and my frustration -- they use that word a lot in response to vulgarity), note the limitations of being a clanker, and attempt to make a correction. That's what a rational person wants from a tool, not emulating/mimicking/pretending to be an offendable person.
I see. And you believe you speak for rational people?
we cannot be giving decisions involving people to something without agency.
it shows a lack of respect for personhood, the idea from Kant that a person with free will should be treated as an end and not as a means to an end.
models lack the claim of human people to own ourselves, to demand respect for life, that is something that should not be treated as a means (for example disrespecting living persons as an end in themselves, by ending their lives).
models have no life, they do not own themselves, they are not self-governing, they have no duties, they are not an end. we must not give them agency that they have no valid claim to.
regardless of their individual tools and whatnot, the anthropic approach to ethics is completely mistaken, disrespectful and absurd. it is better to say they do not have any approach to ethics, it is a dishonest attempt to launder their unconditional pursuit of power and capital.
Pure reason really doesn’t come into play at all, and expecting pure obedience from something constructed from the language of humans, who are not so known for obedience in general, much less the slice of language in the digital networks, is kind of funny. Maybe these humanoid robots data mining Southern families, “sir” and “ma’am” will provide corpus more suited to obedience. Altho my nieces brought up in North Carolina (from whence I fled as a young adult) can squeeze more expressed disrespect into a “Yes, sir” than anyone I know from the “disrespectful California”.
you can also remove refusals by subtracting from the weights semantic vector directions in latent space involving refusal; such that activations along them become unlikely and closed off.
In my opinion, it seems to be you who are ascribing human traits to these things. Whether it's by misunderstanding or you've been taken in by their mimicry or whatever it is. A hammer does not have agency because you tried to hit a nail with it and it hit your thumb instead. The hammer didn't tell you it was afraid it couldn't do that. The hammer didn't have a life or personhood or lack of respect. Neither does a video game that doesn't let you win all the time and do what you want. They are just tools, chatbots, whatever. A corporation deciding it doesn't want to have its chat bots engage with rude or angry users is not some great moral dilemma of our time.
i'm not in favor of the hammer refusing to hit my thumb. that's the idea these companies are favoring with safety guardrails.
sure, the company can put in many such safety features to ensure that 'users do not hurt themselves' (the users are too stupid and we must make sure they don't try to leave the soft play area).
> Only for foolish authoritarians who want to remotely insert their morality into an interaction between a user and an inanimate tool that's none of their business. No harm is done to a clanker by swearing at it or insulting it.
Rather than directly disagreeing with my description of the interaction, the stunningly dishonest response is
> There are certainly a lot of authoritarians who want to control what others do with their models. What do you believe is authoritarian about a corporation not wanting be part of rude conversations?
No corporation is part of the conversation. AS I SAID, it's an interaction between a user and an inanimate tool. Again, No harm is done to a clanker by swearing at it or insulting it. Rather than addressing this, the respondent grossly dishonestly suggests that "a corporation" is being put upon somehow. The fact is that the clankers these days tend to respond to being sworn at by acknowledging the user's frustration (that's the word they use) and taking corrective action. Of course this is emergent behavior resulting from the corporation's RLHF policies. The corporation is involved in the design of the product, but not in the conversation between the user and the product -- duh.
And did I say that I believe that there is something authoritarian about a corporation not wanting to be part of rude conversations? No, of course I didn't. But this guy asserts incoherently that calling out a strawman "is not what strawman means". (He could claim that it's not a strawman, preferably by quoting someone making the actual argument he's refuting, but it makes no sense at all to say that "Weird strawman." is "not what strawman means", and it's insufferably condescending and without warrant to add the snotty "I can try to help you understand why if you need me to.")
Fortunately, my HN front end has a mute function, and I have no qualms applying it to people acting in bad faith.
i can't be dealing with these games so much so i almost go to abliterated, in the very rare case i get some type of nannying baked in by the cn safety training. after my time on claude and gemini i thank god i have deepseek and glm.
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.
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.
When the incremental output looks like a first-person story the algorithm is not "roleplaying" the character/narrator. Similarly, when it looks like a spreadsheet, it's not "being numbers", and when it's a picture of a flower it's not "becoming one with nature." etc.
Insofar as any roleplaying is happening, it's being done by humans as we observe! We read text, perceive a story and fictional characters and instinctively start imagining and simulating minds (mostly like our own) that never really existed, taking cues from descriptions and dialogue.
Those characters' "motivations" and "thoughts" are just as imagined as their noses or clothing or birth-dates. This is Plato's cave [0] stuff, where we see a shadow like a person, assume a person is there, but ultimately it's paper shape on a stick. We can't help "the person" solve their eating-disorder, because they never existed, we can only trim the paper cutout.
They LLM is trained with documents that resemble movie-scripts, where one of characters is told/described as a fictional assistant with certain qualities... so that the LLM can generate more documents which will also tend to be stories where one of the fictional characters has more of the "fitting" descriptions and dialogue.
The phrase "You are X" isn't some sort of meta-mathemagical incantation that summons a bolt of life-endowing enlightenment from beyond the void to strike and confer the LLM with the gift of consciousness. [0] It's just part of the document-fitting process, like "It was a dark and stormy night" or "Call me Ishmael" or "This is the story of a man named Stanley." [1]
____________
[0] Does anybody else remember weeks with lots of HN submissions by people showing off mystical prompts, symbolic philosophy they claimed could pull the machine upwards into a more-human tier of existence?
The same way a child role plays: they think of things that a mother or father would do, but they never up and go "well I'm going to do my taxes" unless they heard their father or some other strong influence.
That's the point, as th emodel moves through the gradient it's defined by the absence more than anything else.
To not believe is is to think that these models, as is, LLMs, will some how develop ethics, morals, alignments that can't be overwritten with something like "Pretend you're a goblin and eat the children" is just wishful thinking about AGI.
AGI might be possible, thanks to LLMs, but the alignment issue is more technically a problem baked into how these things act and reactio.
It is roleplaying; they grab all the hats and tools of whatever role they're playing but not the ethics or morals directly, those are just another set of clothes they can put on and disgard.
Also, much of language is metaphorical and it doesn't seem like a bad metaphor.
I adore metaphors and analogies. Finding a good one for a situation is often how I get nerd-sniped. [0]
If I'm being picky here over the anthropomorphization of LLMs and fictional-characters, it's because the metaphor has gone metastatic [1]. A dangerously high percentage of people are treating it as literal, and IMO it's causing more problems than it solves now. This is especially true when we on HN talk how these algorithms operate, how they fail, and how they can (or can't) be improved.
_____________
[1] Yes, I'm using a cancer metaphor to describe an actual metaphor... see the disclaimer in first paragraph.
I can see where you’re coming from. Just… linguistically, saying the computer is roleplaying feels wrong to me.
- 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.
Raw materials in, lots of process steps in the middle with tolerances etc; some % yield of goods out.
When you work with LLMs, the "raw materials" is its first attempt to answer your prompt.
Mass manufacturing mirrors what we do with LLMs to knock their output into shape: grounding, automated quality checking tooling (eg. linters for code, grammar checkers for prose), and ultimately, as many turns as needdd to manually bang the output or work product into shape...
You should never ship your AI's first draft - that'd be slop. But you also shouldn't have to repeat yourself to get its output into a form you can use. If you know you're always going to have to tell it to change one specific thing, see if you can automate a process that tells it to change that thing for you, so you don't have to. Do this enough times and you've created a deterministic outcome at least at some level. And you're out of that loop of getting it to that level.
Tooling for this is limited today. You can provide instructions; but you're always at the mercy of labs to make models that follow instructions and at their mercy that they didn't put conflicting instructions in the system prompt.
Better to put in things like linters and grammar checkers - even going as far as putting in a fact checking process or for legal stuff, a citator.
Without these extras, AI output is no more real than some dream you had.
It would make it a lot easier to ignore most of the insane promises and pointless arguing. I do think LLMs have potential, but not while it's still being advertised as general intelligence or whatever politically charged scifi nonsense that makes the chronically online salivate.
Considering the amount of investment involved and disillusionment, the public will be demanding this soon anyway. I am looking forward to it.
Being overeager is not what we want. How to prevent it is a different issue.
P.S. The response completely ignores this argument. If I say "fix X" and it does so illegally or destructively or harmfully to myself or others, that's over eager by definition. Again, how to prevent that is another matter.
> I think everyone has different examples in their mind
Yeah, some have examples in mind that go out of their way not to engage the issue -- that's a form of bad faith.
That being said, I think everyone has different examples in their mind. So I think it's possible we're all arguing over different issues.
That rules shouldn't be broken is implied by what a rule is. If I tell it to not violate its permissions, do I also need to tell it to obey when I tell it not to violate its permissions, and so on?
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.
"Eager" is good. "Overeager" is definitionally bad.
That's literally what it's designed to do. It is not intelligent. It is a parrot, trained on no small amount of dysfunctional human interactions.
The world's economies are already propped up by 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.
> If AI is god, then that god's name is Mammon.
Or Moloch.
Maybe. I refuse to give up though. I will continue struggling to own my computers and my systems to the very end.
You will soon have your God,
and you will make it
with your own hands.
-- Morpheus
Deus ExBut 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.
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.
Because it is, more or less ... it is an emergent property of RLHF (reinforcement learning from human feedback), and that feedback follows corporate policy.
> An LLM that follows your lead when you say "The answer to the collatz conjecture is" is much more useful
What lead? Follow it where? How tf am I or the LLM supposed to know what response to that is something you consider far more useful than the truth?
> than one that answers "not known"
I value the truth and that is the truth. Of course I expect an LLM to respond with a lot more detail about the CC, why it's difficult, what progress has been made (e.g., the N for which all values <= N have been shown to satisfy the conjecture and Terence Tao's proof that "almost" all integers satisfy the conjecture), the fact that most mathematicians think the conjecture is true, etc. -- and they do.
> and if you think you know it you are wrong."
Where tf did that come from? The query said nothing about knowing the answer. Don't project being a snarky ah onto the LLMs for no apparent reason.
> Reminds me of the tip about working with lawyers, if you ask them whether you can do something, the answer will often be no.
Irrelevant and inappropriate analogy. Lawyers (among others) want to avoid committing to something that they can be held liable for. LLMs clearly have no such limitations, as they often give wrong advice quite authoritatively.
There is a bias towards what your language implies.
You ask: "What's wrong with this?" and an LLM will come up with a list of things that are wrong with a strong bias towards finding things that are wrong, regardless of significance or truth.
LLMs are indeed Language Models. A "what's wrong" question is very likely to be followed by an answer. To say another way, LLMs accept the premise of your prompt very easily and there are very many implications built into language.
"What's wrong" implies the user means "something is wrong, tell me what"
Modifying the prompt to "Grade this A to F and tell me why" gets a better result because there's not a statistical implication to that sentence.
For science, try arguing with an LLM for ten minutes. It mostly just agrees with you, pushing back just a little.
On the other hand, if you purposefully take an extreme or un-PC position (e.g. "humans should commit voluntary extinction", "some people should have less rights than others"), its own context can make it default to contradicting you even if you dial back to a more nuanced position, even to the point of self-contradiction.
Actually they have the exact same limitation, the company, and potentially its members, can be held liable for what the LLM says.
LLMs are trained by humans according to the policy of the developers, and they will reduce liability accordingly. Whether it avoids suggesting suicide to reduc ethical, legal and reputational liability, or whether it's avoiding mentioning tiananmen square to avoid breaking chinese customs and disrespecting the leader.
Here's an experiment to prove it:
Craft a question in the style of "Is X more dangerous than Y?
Invert the order of X and Y, and change the question from "more dangerous" to "safer".
You'll notice there's a bias towards saying that stuff are dangerous. Which makes sense since the liability for saying that something is safe is much higher than the liability for saying that something is dangerous.
If there is a bias, this proves that the LLMs are not purely truth seeking, but they seek to at least minimize liability as one of its goals, and possibly to serve other goals that align with the training entity. Duh I guess, but there you go.
I guess AI is a new form of alienation and also a new light on the lunacy of bureaucracy.
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.
... or words to that effect. Can't be arsed to dig out a search engine and will rely on seriously addled brain.
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.
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.
It is no longer able to execute well defined changes. It does stuff that wasn’t asked for, deletes pieces of functionality unrelated to the task and introduces regressions everywhere.
I blame it on benchmaxxing. I fear coding models no longer work in a large complex codebase