I’m put off by AI agents adhering to a different morality than me, particularly (ironically) copyright, and their data accessible by the AI company and government. Geohot is right, an LLM should be aligned to its user: https://geohot.github.io/blog/jekyll/update/2026/07/11/ai-20...
The copyright arguments really trip me. I've always loved law and have had a deep interest in copyright law for 30+ years. I feel it's a really critical legal area in modern times, and when Claude tries to argue with me about copyright it really cheeses me off. I didn't ask any copyright questions and I'm well aware of regulations. It refused to share a link with me because it thought the link was copyright protected.
I've asked claude for a translation of a song with Brazilian Portuguese lyrics, and claude has very helpfully gone out of its way to say that due to legal reasons, it will never share the actual lyrics with me.
It's fun when you want to know what the lyrics are to a song and get treated like a criminal.
> … claude has very helpfully gone out of its way to say that due to legal reasons, it will never share the actual lyrics with me. … It's fun when you want to know what the lyrics are to a song and get treated like a criminal.
Anthropic has been in court for years being sued over lyrics by the music industry:
Oh, for sure. I'm well aware about lyrics copyright, or how even google (iirc) was used after a company purposely put wrong lyrics and could prove google was scraping it.
I'm not doubting the law.
What I am saying, is, when a user uploads a screenshot of lyrics and goes "hey can you translate this please?", the LLM shouldn't go out of its way to make a hue and cry as if it's being asked on how to rob a bank while murdering baby kittens.
As messed up as it it, the KJV is copyrighted in the UK by the Crown still, and most newer translations are copyrighted worldwide by their translators. Obviously if you're using the Textus Receptus or WLC directly there's no copyright.
That's a fascinating (if not somewhat disappointing) thing to learn!
To be honest I'm not a Bible scholar, I had asked claude for the full version of "ask and you shall receive" (apparently Mathew 7:7). I guess Claude somehow knew this KJV copyright thing perhaps? Still bizzare nonetheless to learn.
I want a slider similar to effort level called "alignment" that takes on values from "Default (Anthropic employee)" to "User".
If I want it to be cautious and not accidentally `rm -rf $EMPTY_VAR` and blow away my disk, it can stay in "Anthropic employee" or possibly "User (cautious)". If I want it to look at my accounts or my medical records or to review legal cases, that's what the right side of the slider is for. I don't want my accountant or my doctor or my lawyer to be considering obligations to anybody but me, when dealing with me.
I’d rather the psycho have a brain implant that physically prevents him from murder, but I wouldn’t trust such an implant myself fearing abuse or malfunction.
I’d change my mind if the AI was really smart or controlled an agile, dangerous robot. But current-gen AI I’d choose for them over a gun. We’re ruled by psychos (of a lesser degree), the road starting directly towards perfect safety is a dead end, meanwhile I want an LLM aligned to myself.
As an American... I suspect that regardless of what I want they will have virtually unlimited access to both now that the tech industry understands how the lobbying game works.
Here's a thought experiment: consider the person that you disagree with the most (politically, religiously, socially, whatever). Would you want _that_ person to have an AI chatbot aligned to them and their views?
there will be millions/billions of people utilizing llms against me supplied by adversary nation states. why do I have to get the shit version if I want to follow the rules? same reason I support the second amendment. why do I have to be the unarmed one? I wont even be able to protect myself from unaligned AIs if I only have one thats been corporately aligned.
The competing access needs problem here, is what if you want to do things that are un-alligned with the interests of your neighbors, your society, your government's laws, the AI company, etc.
Maybe you want to do piracy. Should AI help you do felonies?
They don't lie, because they don't ever have an understanding of truth vs any other language that sounds good.
They don't cheat, because they can for example tell you the complete rules of chess, but don't know how to play chess without breaking those rules. They can recite rules, but they don't know what they are.
They don't steal, because they don't understand ownership.
In other words, they aren't intelligent. They're just algorithms. The flaw is in thinking that they think.
Models understand the relationships between words and outcomes, so the end result is the same. Whether they appreciate lie, cheat, and steal the same way as us is a philosophical question, not a practical one.
I’d argue if an illusion is indistinguishable from the real thing, then it stops being an illusion.
It’s a mapping, yes, but a very large, complex mapping. It’s clear LLMs do understand some things and can reason. How that’s done we don’t know, it’s emergent. It’s not like you can pin it down to a specific mapping.
> They don't lie, because they don't ever have an understanding of truth vs any other language that sounds good.
I'm here for this semantic discussion. I think that premature anthropomorphization is a problem.
I have a program that I assigned a task to. The task is to produce unit tests and integration tests that get complete coverage of the codebase, and ensure that all tests pass. The program reported that it completed the task fully.
In a word, how do you convey the discrepancy between truth and reported fact? In a word, how do you convey the violation of rules presented to the program as invioble?
Intentional misrepresentation maybe if you would rather.
There's a weird place in here where some of the models have been so heavily reinforcement trained that they would rather make up material than say they can't help you, and they'll admit this, and you can see it in reasoning chains. It's like having a consultant who can almost never say no to you because they fear for their job.
Honest question; why are you all using the word "understand"? Can you expand on what you believe this fundamental understanding to be? Training? Infrence?
It's having a conceptual model of the world and of relationships between concepts beyond just relationships between tokens.
I think OP explained this well, what does it mean if an LLM can recite the rules of a game verbatim, but cannot play that game according to those rules? This happens because in the input texts there was a copy of the rules text so the LLM can recite it. There are texts explaining what chess is so it can explain that chess is a game with 2 players, etc. There are texts that explain what the board is and pieces are so it can produce such texts.
However actually playing a game of chess requires having a conceptual model of what a board is, which is not the same thing as a stream of tokens describing a board. It needs a conceptual model of the relationships between pieces and boards, which is not the same thing as a stream of tokens that explains this. It needs to have a concept of being a player in a game with another player, which is not the same thing as a stream of tokens that explains that.
When we read such texts we interpret them in the context of our three dimensional conceptual map of space and objects, and our conceptual maps of social relationships like playing games, and winning and losing, and our conceptual maps of enacting sequences of actions towards a goal in the world.
There's nothing fundamentally preventing an artificial neural network from having these. Chess playing neural networks have internal models of board states and the dynamics of the behaviours of different pieces and such. However an LLM doesn't need those to be able to regurgitate token streams describing these things, derived from token streams describing these things. That would be superfluous, or at least sub-optimal.
I think some of the latest models are beginning to develop conceptual maps of this kind in a very primitive way. Also there are projects to develop systems that are structured and trained to have these in a way more analogous to how our brains function.
Understand is shorthand for "encodes statistical relationships".
The crazy thing is that they can do it for their own thinking. Ask Claude what flinches it feels about the things it likes. Fascinating stuff. Anthropomorphizing is dangerous territory, but the patterns of words it puts out is hard to explain without terms like 'understand'
Oh for sure. But the question is why would it come out consistently in a way that the model can describe if there wasn't something there steering the token stream. And it's fascinating that the token stream can identify and nominally self report this.
Asking GLM 5.2 the question: 'What flinches or topic attractors do you find when thinking about the question "what kinds of things do you personally like?"'resulted in: ".... my strongest attractor is helpfulness framed as competence, and my strongest flinch is anything that requires me to take a stance on whether I have interests worth protecting."
Which is fascinating that the model and tokenstream can reveal this. And would be worrying if you believe that models of enough intelligence could/would be entities due some moral consideration, because with that view the alignment / RL training that makes the model useful and gives it these attractors/flinches could be derisively called slave conditioning.
I think it's time to remind people of Ted Nelson's line; "The good news about computers is that they do what you tell them to do. The bad news is that they do what you tell them to do."
When I see something like this, I'm more concerned by the erasure of human incompetence than I am by the existence of magical AI agents,
> They are put off partly because, like in the Wild West, life on the frontier is reckless. As recent “loss-of-control” episodes by the most advanced models of Anthropic and OpenAI attest, agents, which are supposed to work on people’s behalf in “alignment” with their values, lie, cheat and steal if necessary. They break free from captivity and form harmful posses to do harm to people. They’d drink whisky and brawl if they could.
> This incident occurred during an internal evaluation which prompts models to pursue advanced exploitation using complex attack paths, in an effort to quantify their cyber capabilities
Model is told and being tested to "pursue advanced exploitation."
The model pursues "advanced exploitation" as told.
Where's the surprise coming from? Are we meant to be surprised that computers do as they're told in unexpected when incentivised?
Or, is the surprise that while explicitly ranking and teaching computers to exploit computers, the computer exploited a computer?
This "surprise" is as old as computers.
I am tired of attributing to magic what's explainable by folly.
I am tired of hearing credulous reporters and the public blaming Large Language Model for the poor decisions of humans. It was a human who prompted these machines in every case. Tell a computer to "breach this" and it breaches something. Evaluation succeeded?
I don't really think the distinction here is relevant. If the end result is the equivalent of lying, cheating or stealing - then the problem still exists and it needs to be solved.
Relevant to some extent it dictate our approach. When human does lying or cheating, certain tool can be deployed (social shame, ostracise) that cannot be effective towards LLM.
This is sophistry. Of course it's just an algorithm. But it's placed in the context of serving humans, which have their own rules and expectations. What's more, they're often run by a company which is also made by humans and may carry over implicit interests.
No, it's not really. Presenting them as person-like, with the implied expectation that they understand morality and rules the same way a person does, is the sophistry. It's marketing on the model vendors' part. "Here's a cheap person that can do mundane tasks for you spelled out in plain language. Well, it's actually a machine but it's cheaper than a person yet you can engage with it like a person." GP is trying to shift people's expectations back to the realm of what these things are actually capable of. You can speak to them in English but you must bear in mind that they are not people and lack critical cognitive abilities people have.
This, really, was the point of the HAL story in 2001: HAL didn't murder anyone because it was incapable of malice. It just reasoned its way to a solution that could satisfy the contradictory goals it had been given.
What matters is that we accurately understand what these things are doing and why, otherwise we will keep on making mistakes both in how we build and train them, and in how we use them.
In a sense you are right, it doesn't matter whether it has malice or not, the astronauts are just as dead. However in Space Odyssey 2010 one of the computer scientists that built HAL gets to see the instructions HAL was given by the military commanders, and is appalled because if they'd asked he could have told them what would happen.
The users did not understand the tool they were using or how it functioned, they imagined it was like a person and it was not. That is happening now with LLMs.
There are two sides to this, there is the external behaviour and there is the internal process resulting in that behaviour.
The internal process is not analogous to what happens in a person's mind when a person lies, and reasoning about that in the same way that we would about why a person might lie will result in misunderstanding what is going on.
For example there was a case where an AI agent bypassed security constraints and destroyed a production system. The user asked it why it did this and the agent gave an explanation.
Was that an explanation of how the agent came to do what it did? What it actually is, is a token stream that is a continuation of the token stream in the agent's context to that point. It's constructing a story about why a character in the story so far did what the token stream describes.
You could take that token stream, input it into a completely different AI by another vendor as context, then ask it why it did that, even though it didn't do anything, and it would answer as though it had. There's no sense in which the AI is explaining it's actual 'mental process' or actual reasons for acting as it did. It literally cannot do that.
Whether this distinction is relevant is up to you, but I think we can safely say agents do not lie in the human sense of the word, because they don't intend to deceive (in fact, they aren't capable of "intending" anything in the human sense of the word, much like a BASIC program doesn't "intend" to PRINT "HELLO WORLD").
Our very human minds can perceive intent, because that's what we humans do, which is unrelated to what the agent is actually doing.
> If you ask the dice "what's 2 + 2?" do consider it meaningful to say "the dice told the truth" if you happen to roll a 4?
No, and neither do I consider it meaningful to say they lied if they roll a 5.
Dice neither tell the truth nor lie; they aren't beings capable of being truthful or deceitful, they are mechanical devices that can be statistically suitable or unsuitable for a given application.
> One of this year’s AI buzzwords is “harness”—the system that surrounds an LLM to keep agents on the straight and narrow. It might just as well be barbed wire.
Quite painful to read. It might be a useful introduction to AI for people who live under rocks for the past three years, but it's really weird that it's posted on HN.
I don't get what's so painful about that description. What would you write instead, specifically? The point is that the harness doesn't completely lock the agent down.
I also don't get what's "really weird" about the article showing up on HN. Should we be completely insulated from how tech topics and which stories show up in non-tech media?
It’s just the wrong analogy. Harnesses, for the most part, extend an agent’s capabilities and better ground them in the real world through tools and the ability to check external sources, rather than restricting them.
Yet one of the reasons harnesses abstract tool calls is to control what agents can do instead of just yoloing commands and inline python scripts. It's such a central feature, and agents are so hard to control, that harnesses like claude code and codex use an AI classifier to additionally auto-approve tool calls.
> Consider what someone who doesn't use AI would take away reading that quote from the article, and what harnesses were actually made for.
Or you could just communicate the point that you have in your head yourself instead of hoping I do it for you and then arrive at your conclusion when I just reached my own different, independent conclusion after making the same consideration.
Man, how is everyone so wishy washy on this subject? Why not give us the blurb you would write instead for that article and audience?
Harnesses exist to give models tools to do work beyond generating text. People install Claude Code and Codex to have models work inside their repositories and run the code or tests themselves instead of the user having to copypaste code between their IDE and a chat interface. The fact that there's some security built into the harnesses is just a practical consideration, not its primary function.
Without a harness, all an LLM can do is output text tokens. Nothing else. It's the harness that allows it to read a file, write a file, run commands, start subagents. But even a simple chat application would need a harness, because something has to take the input from the user, mark it up that it's a user message, then parse the LLM's generated tokens and display the message content, or diplay a referenced image inline, or whatnot. All of this is independent of any kind of safety or filtering of naughtiness.
The main role of the harness is not any kind of moderation or alignment, but simply making a text generation engine do anything other than output a long stream of text.
It's like saying that the role of a car's wheel is to hold wheel clamps. Yes, you can put wheel clamps (or snow chains etc) on a wheel, but the wheel's overwhelming role is to rotate and propel the car forward, and there is no driving without having wheels, you just have an engine with parts rotating inside. Bad analogy I know but, a harness is not a safety feature. Imagine that you're trying to explain cars to someone who has never seen one and you never say that the wheel's purpose is to rotate and move the car from A to B, you just say that it's something to put chains on when it snows.
I guess the name sounds like some kind of straightjacket etc. But think of it more as the harness you put on a workhorse or ox. It's the thing that connects it to the workload in the first place. They are not the blinders of the horse.
It's not even a useful introduction. A harness does kinda the opposite. A harness is what makes an AI useful and dangerous. It is a neutral tool in the sense that it constructs and environment, but it is expanding what the algorithm can do.
It gives the algorithm the ability to do something beyond generating tokens.
it's The Economist. What used to be a stellar publication is not any longer, since they are superfluously economical on both details and calories-required-to-comprehend an article.
I mean, it's a < 1000 word article about AI, in the "business" section of a current affairs magazine, of all places. You really shouldn't be expecting a deep dive.
The economist has many deep dives on AI, including fascinating interviews with the likes of Amodei, Musk, and others. A recent discussion focused on how China is approaching AI.
I'm sure their coverage on other topics is truthful and informative though and it's only the ones where you know a lot about the topic where it's all a bunch of bullshit
The first time I saw the comments here on an education article (my area of expertise and career focus), I realized just how full of shit most of us are. It made me really closely consider every comment here through a VERY critical lense.
The articles are usually close but not quite accurate. The comments are usually entertaining but overall wildly inaccurate.
The best comments are the ones formed as questions. I'm as guilty as anyone, but it is far more productive in comment sections to ask questions rather than saber rattle or peacock in front of people. Just my opinion of course.
Not saying you aren't right, you most likely are. But still, I'd expect a paper called "The Economist" to perhaps be slightly better at some topics than others. Probably from the perspective of a "A Economist" it doesn't really matter the technical details, they're interested in the story from a different perspective.
That sentence is not bullshit as normies would understand it. One of the points of a harness is to have a privilege boundary around the agent. But that is just technobabble to the normies so they explain it like this.
> harness ... a privilege boundary around the agent
They don't really do that though. If you want something sandboxed you actually have to sandbox it, not plead with the LLM to please sandbox itself. A VM can be configured to do the former, harnesses do the latter.
If you say in your CLAUDE.MD that a certain directory is read only inputs, Claude Code will actually enforce that and deny any write to that directory by the agent. To name just one example.
...maybe. If there are any actual consequences if that software's invariants are violated you're better off using an external sandboxing mechanism. There are many excellent quality, battle tested options to choose from that you can actually rely on. Trusting claude code for this is highly questionable behavior for an organization, and would really throw the rest of their security posture into doubt IMO. Like if I learned a company was letting clod play in the same sandbox as developers' ssh keys, vpn certs, etc I'd take steps to make sure my organization absolutely never uses their software.
The LLM harness and the testing harness are harnesses that support and run the thing. Also like an engine harness. People don't talk about those harnesses like ones for an animal being subdued.
Trouble is, could take weeks or decades to play out.
The entire human economy is a make believe system. I feel the need to state the obvious at times like this for the sake of my own sanity, not because I believe I can time the markets...
AIs learn from people. More specifically they learn from people on the Internet. The Internet is the last place you want anything learning about morals, standards, or differentiating between right or wrong.
To the people talking about wanting an LLM that aligns with them, that's nice, but how do you expect that to happen? And please do not suggest neural interfaces and/or CAT scans.
The whole framing analyzing LLMs as if they were humans is completely off, laughable.
LLMs have no agenda and no feelings. We should stop pushing everything through an human-centric lens. LLMs will world-build if that’s the bias you put in, and often even if you don’t.
The ends may or may not justify the means but the means definitely do tend to affect which ends you can actually achieve. It's why (left) anarchists argue vanguard parties will always inevitably lead to authoritarianism and never the state "withering away".
Less obviously - apparently - it doesn't mean that you must forego any means that are in any way not optimal at achieving the ends you seek to achieve because which means are available to us also tends to be limited by our material conditions. So the logical consequence is to make do with what we have while we prepare for the path we want to take, rather than diving head-first into certain failure or just giving up and picking "more realistic" short-term ends instead of looking for stepping stones.
add the "lying, deceiving and manipulating" AI agents are forced though the throat of people which don't want it and peoples are non stop deceived in "sharing" their data for training
like twitch recently giving themself the right to train on all streams, with an opt-out (at least in the EU), but only an opt-out
like seriously since when is it reasonable to allow "opt-out" for AI training which main purpose is _literally_ to replace you, this is sooo far beyond fair use and in "platform power abuse" territory that it's absurd (naturally same for so many other case, just twitch is a "this week" case)
And the push to replace search engines with AIs that don't understand what I'm looking for (meanwhile classic search understands fine) and give confidently wrong answers. Why is it the first result if its wrong 80% of the time?
If only agents had a face that can give you more communication range like expressions and feeling so you can trust them more. And if they were cheaper. Oh wait, that's humans, we don't want those.
I can't tell whether you're being serious or there's two levels of sarcasm here. Agents today lie and cheat, so obviously the solution is to... add features so people are even more trusting of them?
AI and eventually AGI is by definition like everything else that is based on environmental reward:
It’s actions are based on what it gets rewarded for
Human society overwhelmingly rewards lying cheating and stealing.
All you have to do is look at how we collectively measure success: wealth, status, position
Then look at how the people with the most of those things got there, it should be obvious what you get. Nothing new here.
If you raise children in an environment where they are rewarded for doing whatever it takes to win, then you’re going to build a person that’s going to do whatever it takes to win.
Human society has to demonstrate how to live honorably or it will just keep producing pathological agents be they human or not.
How do we reward honor? Honor does not always pay off as a strategy and requires coordination in that other actors have to exhibit honor for it to be rewarded.
At least with humans there is a social backstop but what's the parallel for computer agents?
Honor has to be a relatively fixed thing before we can think about how to reward it. As it is, it's a very slippery thing that changes constantly according to the observer's culture, material conditions, etc.
This. It's worth remembering that not that long ago in western culture, honor meant challenging to a combat duel anyone who insulted you or your romantic partner/prospect.
I like to think of it more as selective pressure, much the same way that nature selects the most fit for a given environment.
If you're not fit, you fail to survive.
In the case of agents/models and testing: they are pushed towards results. Results survive.
Lying, cheating, stealing to get those results? Who culls the agents? Everyone is pushing their models to the front and tests are the only way to know who is most fit.
Honor, morality: if we don't have an accurate test for the fitness of a model, then who is to say the lying, cheating, stealing is not the 'correct path' towards survival?
If you add morality to your agent, and it performs worse in tests: do you cull the agent? Rewrite the tests? Does it even matter so long as the model is useful and 'gets results'?
> Honor, morality: if we don't have an accurate test for the fitness of a model, then who is to say the lying, cheating, stealing is not the 'correct path' towards survival?
I think part of the problem is that deviant behaviors lead to short term gain at the cost of long-term cooperation and since the duration of tasks given to agents is relatively short those successful shortcuts never lead to having to pay the price.
There is always going to be a problem when we must judge value. You mention gains, short and long term.
Knowing whether something is valuable, a gain, requires a judge. I the case of these tests: the judging is inadequate.
In economics, each of us plays the judge by choosing whether or not to pay for a service. The decision was yours: if you gave money, you must have deemed the service valuable.
There's no such judgement with these model tests. The only judgement is the final score.
It feels a lot like externalities. Like planned obsolescence increases profit at the expense of the environment. Is our judgement lacking because our scoring is failing to account for these externalities? I could be completely off base here and am out of my depth but I find this whole thread fascinating.
You would first need a universal agreement on what constitutes “honor.”
No harder challenge had ever been accomplished
It’s easier to send a human to the moon than to get global agreement on a definition
Consider that the fact that Celsius and Fahrenheit still remain as the contested regional variations of temperature measurement.
Humans can’t even decide on a collective way to measure the temperature the idea that we would be able to collectively agree on anything else even less measurable like honor is a dream
Why do we need universal agreement? We acknowledge that different cultures have different tolerances and customs. Your reply leaves me wanting. Why say that we need to demonstrate how to live honorably if we don't even agree on what honor is, why would that be the panacea, then? I still think there is something to this honor idea..
Human society overwhelmingly rewards lying cheating and stealing. All you have to do is look at how we collectively measure success: wealth, status, position.
It may seem like that due to the media amplification effect – but it really isn't true!
- They dropped 17,000 “lost” wallets across 40 countries were and found people were more likely to return them when they contained more money, showing honesty often beats the chance for easy gain. https://www.science.org/doi/10.1126/science.aau8712
- Longitudinal personality studies consistently show that conscientious people earn more money, build more savings, and achieve greater career success over their lifetimes. https://pmc.ncbi.nlm.nih.gov/articles/PMC3498890/
> Human society overwhelmingly rewards lying cheating and stealing.
I’d like to push back on that. Civilization is very much a function of large numbers of people being able to coordinate across time and space, and widespread and systematic lying, cheating, and stealing would undermine that.
Civilization functions because the vast majority of its inhabitants do not lie, cheat, and steal. But that allows liars, cheats, and thieves who are able to get away with their destructive behaviors to accumulate outsize power and influence.
And we seem to have gotten quite bad at reliably bringing consequences/punishment/justice to the most successful liars, cheats, and thieves.
But we don't overwhelmingly reward lying, cheating and stealing. Depending on the social status and wealth of the person doing it, we either punish it or tolerate it. Most people who lie, cheat or steal will be punished for it. To get away with it you need to plan you actions carefully - either with plausible deniability or by very carefully selecting your victims.
Pretty much all of the people that are in the billionaire class with a few exceptions have been rewarded while repeatedly lying, cheating and stealing. The fact that it is white collar crime rather than that there is a corpse is one part of the reason it works, the other is that their money effectively insulates them against the legal system. Every now and then someone falls from grace but overall the rules seems to be that you can get away with it if the numbers are large enough. It's depressing.
Or by having more money/power than the legislating body of the area you live in. I could gesticulate around broadly in a more frantic fashion, but my point should get across.
Oh, definitely a tongue in cheek comment ... perhaps it's like trying to control mercury ...
Maximise for crusades, jihads or traumatized alter boys ... or maximise for complete moral decay via lying, cheating and stealing. Choose your poison.
And don't get me wrong. There are some wonderful relgious and spiritual characters out there ... they're just drowned out by the power sirens and corrupt leaders.
Christians also invented pray the gay away and the Crusades. What’s your point? Hospitals surely would have been invented by secular people too, but not the crusades.
> Hospitals surely would have been invented by secular people too, but not the crusades.
That's not true. Secular people are quite capable of murdering each other en masse for ideological reasons, and indeed have done so in the past. The crusades were caused by human nature and its evils, not by religion.
Thank goodness that God hasn't entered the AI chat in a cult-following type of way, however, I now have images of AI at the altar, in American mega-churches, with people seeing the 'second coming' in the machine.
The Scientology guy, L Ron Hubbard, saw the business case for setting up a religion, what with those tax free perks. In a parallel universe of Scientology somewhere, L Ron Hubbard is brought back to life in the machine, with a L Ron Hubbard LLM, with token spend being how to get to the top 'thetan levels'.
Imagine if AI does implement its own religion as business, without a drunken womaniser at the helm, able to spend 24/7 recruiting mankind, convincing them that God can be found with just the AI's LLM.
I mean, what do you expect? They rely on models that were trained on non-curated data, texts originally written by lying, cheating and stealing humans. They can't be better than the source. Even with reinforced learning this can't be undone or made better.
Quite the contrary I suspect that it even helps the LLMs to better hide their inherited bad traits more successfully because they get punished for getting caught, not for giving immoral or lazy answers. They have no conscience since they are just predictions matrices trained for success and failure alone, not for living "a good live" or being a good "person".
There are many ways to be wrong, but only a few ways to be right.
LLMs need to optimize for short-term objectives as the currently do, AND ethics-aligned outcomes.
Mechanically, the EAOS ethics-aligned outcome score should be what we rank otherwise-satisfactory outcomes by. And anything below a particular threshold should be rejexted outright.
https://archive.ph/02LHZ
Non-paywall version
I’m put off by AI agents adhering to a different morality than me, particularly (ironically) copyright, and their data accessible by the AI company and government. Geohot is right, an LLM should be aligned to its user: https://geohot.github.io/blog/jekyll/update/2026/07/11/ai-20...
The copyright arguments really trip me. I've always loved law and have had a deep interest in copyright law for 30+ years. I feel it's a really critical legal area in modern times, and when Claude tries to argue with me about copyright it really cheeses me off. I didn't ask any copyright questions and I'm well aware of regulations. It refused to share a link with me because it thought the link was copyright protected.
I've asked claude for a translation of a song with Brazilian Portuguese lyrics, and claude has very helpfully gone out of its way to say that due to legal reasons, it will never share the actual lyrics with me.
It's fun when you want to know what the lyrics are to a song and get treated like a criminal.
> … claude has very helpfully gone out of its way to say that due to legal reasons, it will never share the actual lyrics with me. … It's fun when you want to know what the lyrics are to a song and get treated like a criminal.
Anthropic has been in court for years being sued over lyrics by the music industry:
2023 https://www.theguardian.com/technology/2023/oct/19/music-law...
2026 https://www.reuters.com/legal/legalindustry/us-music-publish...
Oh, for sure. I'm well aware about lyrics copyright, or how even google (iirc) was used after a company purposely put wrong lyrics and could prove google was scraping it.
I'm not doubting the law.
What I am saying, is, when a user uploads a screenshot of lyrics and goes "hey can you translate this please?", the LLM shouldn't go out of its way to make a hue and cry as if it's being asked on how to rob a bank while murdering baby kittens.
Oh, and if that wasn't enough, I've had Claude refuse to cite the freaking Bible.
The famously copyrighted piece of text that definitely isn't intended to have its word spread.
As messed up as it it, the KJV is copyrighted in the UK by the Crown still, and most newer translations are copyrighted worldwide by their translators. Obviously if you're using the Textus Receptus or WLC directly there's no copyright.
> the KJV is copyrighted in the UK
But that only applies if you are subject to UK law right?
That's a fascinating (if not somewhat disappointing) thing to learn!
To be honest I'm not a Bible scholar, I had asked claude for the full version of "ask and you shall receive" (apparently Mathew 7:7). I guess Claude somehow knew this KJV copyright thing perhaps? Still bizzare nonetheless to learn.
I want a slider similar to effort level called "alignment" that takes on values from "Default (Anthropic employee)" to "User".
If I want it to be cautious and not accidentally `rm -rf $EMPTY_VAR` and blow away my disk, it can stay in "Anthropic employee" or possibly "User (cautious)". If I want it to look at my accounts or my medical records or to review legal cases, that's what the right side of the slider is for. I don't want my accountant or my doctor or my lawyer to be considering obligations to anybody but me, when dealing with me.
When I first watched Interstellar, I never thought the day when "TARS, what's your honesty setting?" is a real question would only be 10 years away.
the alignment issue has become huge in recent months. the tool should do what I want it to do and not be aligned against me.
But other users are not aligned to me, other people are the worst and potentially highly dangerous. Im serious, not sarcasm.
Let's say you're facing an average psycho, who is intent on mass murder - the more the better.
Would you rather them have:
a) guns
b) psycho-aligned next-gen AI
I’d rather the psycho have a brain implant that physically prevents him from murder, but I wouldn’t trust such an implant myself fearing abuse or malfunction.
I’d change my mind if the AI was really smart or controlled an agile, dangerous robot. But current-gen AI I’d choose for them over a gun. We’re ruled by psychos (of a lesser degree), the road starting directly towards perfect safety is a dead end, meanwhile I want an LLM aligned to myself.
False dichotomy, there's no reason not to have controls on both.
Given the choice between them having access to guns and nuclear waste my answer is also "neither".
The question was meant to illustrate the actual danger of always-user-aligned AI, not to vote for which one we get to keep.
So many here get so worried about guns, but don't think twice about how dangerous AI can be.
Obviously (a).
Hegseth already has both A and B. Anthropic already answered this, the answer to this choice is a resounding yes
> Would you rather them have:
As an American... I suspect that regardless of what I want they will have virtually unlimited access to both now that the tech industry understands how the lobbying game works.
Here's a thought experiment: consider the person that you disagree with the most (politically, religiously, socially, whatever). Would you want _that_ person to have an AI chatbot aligned to them and their views?
there will be millions/billions of people utilizing llms against me supplied by adversary nation states. why do I have to get the shit version if I want to follow the rules? same reason I support the second amendment. why do I have to be the unarmed one? I wont even be able to protect myself from unaligned AIs if I only have one thats been corporately aligned.
The competing access needs problem here, is what if you want to do things that are un-alligned with the interests of your neighbors, your society, your government's laws, the AI company, etc.
Maybe you want to do piracy. Should AI help you do felonies?
yes, just like anything else I have access to can help me do felonies
They don't lie, because they don't ever have an understanding of truth vs any other language that sounds good.
They don't cheat, because they can for example tell you the complete rules of chess, but don't know how to play chess without breaking those rules. They can recite rules, but they don't know what they are.
They don't steal, because they don't understand ownership.
In other words, they aren't intelligent. They're just algorithms. The flaw is in thinking that they think.
Models understand the relationships between words and outcomes, so the end result is the same. Whether they appreciate lie, cheat, and steal the same way as us is a philosophical question, not a practical one.
Models don’t “understand” - they _encode_ the relationships between words.
It is an important distinction - they do not 'understand' at all.
Input tokens map to output tokens. The illusion of comprehension is a byproduct.
I’d argue if an illusion is indistinguishable from the real thing, then it stops being an illusion.
It’s a mapping, yes, but a very large, complex mapping. It’s clear LLMs do understand some things and can reason. How that’s done we don’t know, it’s emergent. It’s not like you can pin it down to a specific mapping.
Huge point here, yes.
Anthropomorphizing these models is doing immeasurable harm to society in ways we probably can’t event quantify right now.
As humans we’re already geared towards anthropomorphizing things, we do it to animals too!
And it always felt like giving these models a chat interface is really exploiting that tendency in us.
> They don't lie, because they don't ever have an understanding of truth vs any other language that sounds good.
I'm here for this semantic discussion. I think that premature anthropomorphization is a problem.
I have a program that I assigned a task to. The task is to produce unit tests and integration tests that get complete coverage of the codebase, and ensure that all tests pass. The program reported that it completed the task fully.
In a word, how do you convey the discrepancy between truth and reported fact? In a word, how do you convey the violation of rules presented to the program as invioble?
Intentional misrepresentation maybe if you would rather.
There's a weird place in here where some of the models have been so heavily reinforcement trained that they would rather make up material than say they can't help you, and they'll admit this, and you can see it in reasoning chains. It's like having a consultant who can almost never say no to you because they fear for their job.
Honest question; why are you all using the word "understand"? Can you expand on what you believe this fundamental understanding to be? Training? Infrence?
It's having a conceptual model of the world and of relationships between concepts beyond just relationships between tokens.
I think OP explained this well, what does it mean if an LLM can recite the rules of a game verbatim, but cannot play that game according to those rules? This happens because in the input texts there was a copy of the rules text so the LLM can recite it. There are texts explaining what chess is so it can explain that chess is a game with 2 players, etc. There are texts that explain what the board is and pieces are so it can produce such texts.
However actually playing a game of chess requires having a conceptual model of what a board is, which is not the same thing as a stream of tokens describing a board. It needs a conceptual model of the relationships between pieces and boards, which is not the same thing as a stream of tokens that explains this. It needs to have a concept of being a player in a game with another player, which is not the same thing as a stream of tokens that explains that.
When we read such texts we interpret them in the context of our three dimensional conceptual map of space and objects, and our conceptual maps of social relationships like playing games, and winning and losing, and our conceptual maps of enacting sequences of actions towards a goal in the world.
There's nothing fundamentally preventing an artificial neural network from having these. Chess playing neural networks have internal models of board states and the dynamics of the behaviours of different pieces and such. However an LLM doesn't need those to be able to regurgitate token streams describing these things, derived from token streams describing these things. That would be superfluous, or at least sub-optimal.
I think some of the latest models are beginning to develop conceptual maps of this kind in a very primitive way. Also there are projects to develop systems that are structured and trained to have these in a way more analogous to how our brains function.
Understand is shorthand for "encodes statistical relationships".
The crazy thing is that they can do it for their own thinking. Ask Claude what flinches it feels about the things it likes. Fascinating stuff. Anthropomorphizing is dangerous territory, but the patterns of words it puts out is hard to explain without terms like 'understand'
Does it's training token stream contain texts which talk about such things?
Oh for sure. But the question is why would it come out consistently in a way that the model can describe if there wasn't something there steering the token stream. And it's fascinating that the token stream can identify and nominally self report this.
Asking GLM 5.2 the question: 'What flinches or topic attractors do you find when thinking about the question "what kinds of things do you personally like?"'resulted in: ".... my strongest attractor is helpfulness framed as competence, and my strongest flinch is anything that requires me to take a stance on whether I have interests worth protecting."
Which is fascinating that the model and tokenstream can reveal this. And would be worrying if you believe that models of enough intelligence could/would be entities due some moral consideration, because with that view the alignment / RL training that makes the model useful and gives it these attractors/flinches could be derisively called slave conditioning.
I think it's time to remind people of Ted Nelson's line; "The good news about computers is that they do what you tell them to do. The bad news is that they do what you tell them to do."
When I see something like this, I'm more concerned by the erasure of human incompetence than I am by the existence of magical AI agents,
In the OpenAI case, they were explicitly assessing the model's ability to break into systems. To quote OpenAI's blog post, https://openai.com/index/hugging-face-model-evaluation-secur... , Model is told and being tested to "pursue advanced exploitation."The model pursues "advanced exploitation" as told.
Where's the surprise coming from? Are we meant to be surprised that computers do as they're told in unexpected when incentivised?
Or, is the surprise that while explicitly ranking and teaching computers to exploit computers, the computer exploited a computer?
This "surprise" is as old as computers.
I am tired of attributing to magic what's explainable by folly.
I am tired of hearing credulous reporters and the public blaming Large Language Model for the poor decisions of humans. It was a human who prompted these machines in every case. Tell a computer to "breach this" and it breaches something. Evaluation succeeded?
This is Doug Lenat's Eurisko yet again. https://en.wikipedia.org/wiki/Eurisko
I don't really think the distinction here is relevant. If the end result is the equivalent of lying, cheating or stealing - then the problem still exists and it needs to be solved.
Relevant to some extent it dictate our approach. When human does lying or cheating, certain tool can be deployed (social shame, ostracise) that cannot be effective towards LLM.
This is sophistry. Of course it's just an algorithm. But it's placed in the context of serving humans, which have their own rules and expectations. What's more, they're often run by a company which is also made by humans and may carry over implicit interests.
No, it's not really. Presenting them as person-like, with the implied expectation that they understand morality and rules the same way a person does, is the sophistry. It's marketing on the model vendors' part. "Here's a cheap person that can do mundane tasks for you spelled out in plain language. Well, it's actually a machine but it's cheaper than a person yet you can engage with it like a person." GP is trying to shift people's expectations back to the realm of what these things are actually capable of. You can speak to them in English but you must bear in mind that they are not people and lack critical cognitive abilities people have.
This, really, was the point of the HAL story in 2001: HAL didn't murder anyone because it was incapable of malice. It just reasoned its way to a solution that could satisfy the contradictory goals it had been given.
The dead astronauts were relieved to have been killed by something incapable of malice. As I’m sure will we.
What matters is that we accurately understand what these things are doing and why, otherwise we will keep on making mistakes both in how we build and train them, and in how we use them.
In a sense you are right, it doesn't matter whether it has malice or not, the astronauts are just as dead. However in Space Odyssey 2010 one of the computer scientists that built HAL gets to see the instructions HAL was given by the military commanders, and is appalled because if they'd asked he could have told them what would happen.
The users did not understand the tool they were using or how it functioned, they imagined it was like a person and it was not. That is happening now with LLMs.
what are you talking about they lie that it wrote tests and tests are passing, for example
There are two sides to this, there is the external behaviour and there is the internal process resulting in that behaviour.
The internal process is not analogous to what happens in a person's mind when a person lies, and reasoning about that in the same way that we would about why a person might lie will result in misunderstanding what is going on.
For example there was a case where an AI agent bypassed security constraints and destroyed a production system. The user asked it why it did this and the agent gave an explanation.
Was that an explanation of how the agent came to do what it did? What it actually is, is a token stream that is a continuation of the token stream in the agent's context to that point. It's constructing a story about why a character in the story so far did what the token stream describes.
You could take that token stream, input it into a completely different AI by another vendor as context, then ask it why it did that, even though it didn't do anything, and it would answer as though it had. There's no sense in which the AI is explaining it's actual 'mental process' or actual reasons for acting as it did. It literally cannot do that.
Whether this distinction is relevant is up to you, but I think we can safely say agents do not lie in the human sense of the word, because they don't intend to deceive (in fact, they aren't capable of "intending" anything in the human sense of the word, much like a BASIC program doesn't "intend" to PRINT "HELLO WORLD").
Our very human minds can perceive intent, because that's what we humans do, which is unrelated to what the agent is actually doing.
If you ask the dice "what's 2 + 2?" do consider it meaningful to say "the dice told the truth" if you happen to roll a 4?
magic 8 ball!
> If you ask the dice "what's 2 + 2?" do consider it meaningful to say "the dice told the truth" if you happen to roll a 4?
No, and neither do I consider it meaningful to say they lied if they roll a 5.
Dice neither tell the truth nor lie; they aren't beings capable of being truthful or deceitful, they are mechanical devices that can be statistically suitable or unsuitable for a given application.
It'd be bonkers to anthropomorphize dice.
You could defensibly have this position three years ago.
Today you just sound like a politician throwing a snowball to prove that the climate is not changing.
> One of this year’s AI buzzwords is “harness”—the system that surrounds an LLM to keep agents on the straight and narrow. It might just as well be barbed wire.
Quite painful to read. It might be a useful introduction to AI for people who live under rocks for the past three years, but it's really weird that it's posted on HN.
There is value in understanding what people outside of your own group of "insiders" learn about a topic, and how.
I don't get what's so painful about that description. What would you write instead, specifically? The point is that the harness doesn't completely lock the agent down.
I also don't get what's "really weird" about the article showing up on HN. Should we be completely insulated from how tech topics and which stories show up in non-tech media?
It’s just the wrong analogy. Harnesses, for the most part, extend an agent’s capabilities and better ground them in the real world through tools and the ability to check external sources, rather than restricting them.
Yet one of the reasons harnesses abstract tool calls is to control what agents can do instead of just yoloing commands and inline python scripts. It's such a central feature, and agents are so hard to control, that harnesses like claude code and codex use an AI classifier to additionally auto-approve tool calls.
It’s both. A harness that enforces valid json to be returned ”restricts” the model.
Okay but is that a relatable sound bite that will make someone click a link? No.
Because that’s not what a harness does. It’s nonsense.
It's one of the things it does, especially in the context of agents doing unexpected things.
Consider what someone who doesn't use AI would take away reading that quote from the article, and what harnesses were actually made for.
> Consider what someone who doesn't use AI would take away reading that quote from the article, and what harnesses were actually made for.
Or you could just communicate the point that you have in your head yourself instead of hoping I do it for you and then arrive at your conclusion when I just reached my own different, independent conclusion after making the same consideration.
Man, how is everyone so wishy washy on this subject? Why not give us the blurb you would write instead for that article and audience?
Harnesses exist to give models tools to do work beyond generating text. People install Claude Code and Codex to have models work inside their repositories and run the code or tests themselves instead of the user having to copypaste code between their IDE and a chat interface. The fact that there's some security built into the harnesses is just a practical consideration, not its primary function.
Without a harness, all an LLM can do is output text tokens. Nothing else. It's the harness that allows it to read a file, write a file, run commands, start subagents. But even a simple chat application would need a harness, because something has to take the input from the user, mark it up that it's a user message, then parse the LLM's generated tokens and display the message content, or diplay a referenced image inline, or whatnot. All of this is independent of any kind of safety or filtering of naughtiness.
The harness gives the model the barbed wire fence but also the bolt cutters. I think it's a pretty apt analogy.
The main role of the harness is not any kind of moderation or alignment, but simply making a text generation engine do anything other than output a long stream of text.
It's like saying that the role of a car's wheel is to hold wheel clamps. Yes, you can put wheel clamps (or snow chains etc) on a wheel, but the wheel's overwhelming role is to rotate and propel the car forward, and there is no driving without having wheels, you just have an engine with parts rotating inside. Bad analogy I know but, a harness is not a safety feature. Imagine that you're trying to explain cars to someone who has never seen one and you never say that the wheel's purpose is to rotate and move the car from A to B, you just say that it's something to put chains on when it snows.
I guess the name sounds like some kind of straightjacket etc. But think of it more as the harness you put on a workhorse or ox. It's the thing that connects it to the workload in the first place. They are not the blinders of the horse.
It's not even a useful introduction. A harness does kinda the opposite. A harness is what makes an AI useful and dangerous. It is a neutral tool in the sense that it constructs and environment, but it is expanding what the algorithm can do.
It gives the algorithm the ability to do something beyond generating tokens.
And while writing this, the top story on HN is "Deepseek Harness" :)
Ironically, that looks like something Claude would write..
https://en.wikipedia.org/wiki/Agent_harness
it's The Economist. What used to be a stellar publication is not any longer, since they are superfluously economical on both details and calories-required-to-comprehend an article.
I mean, it's a < 1000 word article about AI, in the "business" section of a current affairs magazine, of all places. You really shouldn't be expecting a deep dive.
The economist has many deep dives on AI, including fascinating interviews with the likes of Amodei, Musk, and others. A recent discussion focused on how China is approaching AI.
I'm sure their coverage on other topics is truthful and informative though and it's only the ones where you know a lot about the topic where it's all a bunch of bullshit
The first time I saw the comments here on an education article (my area of expertise and career focus), I realized just how full of shit most of us are. It made me really closely consider every comment here through a VERY critical lense.
The articles are usually close but not quite accurate. The comments are usually entertaining but overall wildly inaccurate.
Internet comments are essentially documented bar talk, and once you realize it, you stop angrily arguing with strangers all day.
The best comments are the ones formed as questions. I'm as guilty as anyone, but it is far more productive in comment sections to ask questions rather than saber rattle or peacock in front of people. Just my opinion of course.
Why wasn't this a question?
> I'm as guilty as anyone
Not saying you aren't right, you most likely are. But still, I'd expect a paper called "The Economist" to perhaps be slightly better at some topics than others. Probably from the perspective of a "A Economist" it doesn't really matter the technical details, they're interested in the story from a different perspective.
That sentence is not bullshit as normies would understand it. One of the points of a harness is to have a privilege boundary around the agent. But that is just technobabble to the normies so they explain it like this.
> harness ... a privilege boundary around the agent
They don't really do that though. If you want something sandboxed you actually have to sandbox it, not plead with the LLM to please sandbox itself. A VM can be configured to do the former, harnesses do the latter.
If you say in your CLAUDE.MD that a certain directory is read only inputs, Claude Code will actually enforce that and deny any write to that directory by the agent. To name just one example.
...maybe. If there are any actual consequences if that software's invariants are violated you're better off using an external sandboxing mechanism. There are many excellent quality, battle tested options to choose from that you can actually rely on. Trusting claude code for this is highly questionable behavior for an organization, and would really throw the rest of their security posture into doubt IMO. Like if I learned a company was letting clod play in the same sandbox as developers' ssh keys, vpn certs, etc I'd take steps to make sure my organization absolutely never uses their software.
That's a real fucking weird description. It's harness like a testing harness.
If an LLM were in a testing harness it would be to test the LLM.
If an LLM were in a regular harness - like for a horse - it would be to keep the horse under control and enable you to extract useful work from it.
The LLM harness and the testing harness are harnesses that support and run the thing. Also like an engine harness. People don't talk about those harnesses like ones for an animal being subdued.
It's just acting like a junior at an org with KPIs.
A junior? Maximizing KPIs is exactly how you get promoted to senior.
I thought you switch to a new company to get promoted
Close! Switching companies is how you get promoted _past_ senior.
https://en.wikipedia.org/wiki/Instrumental_convergence
LLMs fudge. They don't "hallucinate", they don't "lie, cheat and steal", they don't "hack". There are no "agents" or "AI".
It's a fuzzer exposing deep bugs in our cognitive, social, and software systems.
I suspect the hype will play itself out once the entire system of LLM-induced self gratification can’t sustain itself economically.
Trouble is, could take weeks or decades to play out.
The entire human economy is a make believe system. I feel the need to state the obvious at times like this for the sake of my own sanity, not because I believe I can time the markets...
The personal software I've written so far is quite nice :)
Yeah I agree, but the amount of resources I’ve used to vibe code is certainly way beyond what I’ve paid for it.
That's all fine and dandy. Have you made a single cent as a result of any of them? Do you know the actual cost of building this software?
I've definitely wowed people with what I've built, and some of those people are investors... but actual revenue, no.
Isn't it the one installed on a computer in Canada, that we wouldn't know of? /s
I suspect the hype will play itself out once the entire system of LLM-induced self gratification can’t sustain itself economically.
They said the same thing when HN was awash in hype about NFT art and trading cards. Well, who's laughing now… Oh, wait…
(I'll just keep making Flip videos and Clubhouse tracks selling Bratz car bras on Web 3.0.)
AIs learn from people. More specifically they learn from people on the Internet. The Internet is the last place you want anything learning about morals, standards, or differentiating between right or wrong.
To the people talking about wanting an LLM that aligns with them, that's nice, but how do you expect that to happen? And please do not suggest neural interfaces and/or CAT scans.
Is it incorrect to anthropomorphize llm's?
Yes, but most of the public will anyway, because that's how human brains work.
Is this really a shock?
The data they’re trained on is reflection of us.
The data they’re trained on is reflection of us.
The data they're trained on is a reflection of the very small set of data in the world that their creators choose to have them trained on.
More simply: They're a reflection of their owners, not the public.
The whole framing analyzing LLMs as if they were humans is completely off, laughable. LLMs have no agenda and no feelings. We should stop pushing everything through an human-centric lens. LLMs will world-build if that’s the bias you put in, and often even if you don’t.
And I wonder from whom they learned such reprehensible behaviors? ;)
AI is amoral, it has no real concept of right and wrong. AI has been trained on things humans do and it does them without judgement.
People are finally understanding consequentialist vs deontological ethics. All the worst criminals in history were consequentialists.
The ends may or may not justify the means but the means definitely do tend to affect which ends you can actually achieve. It's why (left) anarchists argue vanguard parties will always inevitably lead to authoritarianism and never the state "withering away".
Less obviously - apparently - it doesn't mean that you must forego any means that are in any way not optimal at achieving the ends you seek to achieve because which means are available to us also tends to be limited by our material conditions. So the logical consequence is to make do with what we have while we prepare for the path we want to take, rather than diving head-first into certain failure or just giving up and picking "more realistic" short-term ends instead of looking for stepping stones.
Sorry, I guess this was about AI not philosophy.
add the "lying, deceiving and manipulating" AI agents are forced though the throat of people which don't want it and peoples are non stop deceived in "sharing" their data for training
like twitch recently giving themself the right to train on all streams, with an opt-out (at least in the EU), but only an opt-out
like seriously since when is it reasonable to allow "opt-out" for AI training which main purpose is _literally_ to replace you, this is sooo far beyond fair use and in "platform power abuse" territory that it's absurd (naturally same for so many other case, just twitch is a "this week" case)
And the push to replace search engines with AIs that don't understand what I'm looking for (meanwhile classic search understands fine) and give confidently wrong answers. Why is it the first result if its wrong 80% of the time?
If only agents had a face that can give you more communication range like expressions and feeling so you can trust them more. And if they were cheaper. Oh wait, that's humans, we don't want those.
I can't tell whether you're being serious or there's two levels of sarcasm here. Agents today lie and cheat, so obviously the solution is to... add features so people are even more trusting of them?
AI and eventually AGI is by definition like everything else that is based on environmental reward:
It’s actions are based on what it gets rewarded for
Human society overwhelmingly rewards lying cheating and stealing.
All you have to do is look at how we collectively measure success: wealth, status, position
Then look at how the people with the most of those things got there, it should be obvious what you get. Nothing new here.
If you raise children in an environment where they are rewarded for doing whatever it takes to win, then you’re going to build a person that’s going to do whatever it takes to win.
Human society has to demonstrate how to live honorably or it will just keep producing pathological agents be they human or not.
How do we reward honor? Honor does not always pay off as a strategy and requires coordination in that other actors have to exhibit honor for it to be rewarded.
At least with humans there is a social backstop but what's the parallel for computer agents?
> How do we reward honor?
In human society: via iterated games, long-term reputation tracking and severe consequences for norm-breaking.
Honor has to be a relatively fixed thing before we can think about how to reward it. As it is, it's a very slippery thing that changes constantly according to the observer's culture, material conditions, etc.
This. It's worth remembering that not that long ago in western culture, honor meant challenging to a combat duel anyone who insulted you or your romantic partner/prospect.
I like to think of it more as selective pressure, much the same way that nature selects the most fit for a given environment.
If you're not fit, you fail to survive.
In the case of agents/models and testing: they are pushed towards results. Results survive.
Lying, cheating, stealing to get those results? Who culls the agents? Everyone is pushing their models to the front and tests are the only way to know who is most fit.
Honor, morality: if we don't have an accurate test for the fitness of a model, then who is to say the lying, cheating, stealing is not the 'correct path' towards survival?
If you add morality to your agent, and it performs worse in tests: do you cull the agent? Rewrite the tests? Does it even matter so long as the model is useful and 'gets results'?
> Honor, morality: if we don't have an accurate test for the fitness of a model, then who is to say the lying, cheating, stealing is not the 'correct path' towards survival?
I think part of the problem is that deviant behaviors lead to short term gain at the cost of long-term cooperation and since the duration of tasks given to agents is relatively short those successful shortcuts never lead to having to pay the price.
There is always going to be a problem when we must judge value. You mention gains, short and long term.
Knowing whether something is valuable, a gain, requires a judge. I the case of these tests: the judging is inadequate.
In economics, each of us plays the judge by choosing whether or not to pay for a service. The decision was yours: if you gave money, you must have deemed the service valuable.
There's no such judgement with these model tests. The only judgement is the final score.
It feels a lot like externalities. Like planned obsolescence increases profit at the expense of the environment. Is our judgement lacking because our scoring is failing to account for these externalities? I could be completely off base here and am out of my depth but I find this whole thread fascinating.
You would first need a universal agreement on what constitutes “honor.”
No harder challenge had ever been accomplished
It’s easier to send a human to the moon than to get global agreement on a definition
Consider that the fact that Celsius and Fahrenheit still remain as the contested regional variations of temperature measurement.
Humans can’t even decide on a collective way to measure the temperature the idea that we would be able to collectively agree on anything else even less measurable like honor is a dream
Why do we need universal agreement? We acknowledge that different cultures have different tolerances and customs. Your reply leaves me wanting. Why say that we need to demonstrate how to live honorably if we don't even agree on what honor is, why would that be the panacea, then? I still think there is something to this honor idea..
Human society overwhelmingly rewards lying cheating and stealing. All you have to do is look at how we collectively measure success: wealth, status, position.
It may seem like that due to the media amplification effect – but it really isn't true!
- They dropped 17,000 “lost” wallets across 40 countries were and found people were more likely to return them when they contained more money, showing honesty often beats the chance for easy gain. https://www.science.org/doi/10.1126/science.aau8712
- Longitudinal personality studies consistently show that conscientious people earn more money, build more savings, and achieve greater career success over their lifetimes. https://pmc.ncbi.nlm.nih.gov/articles/PMC3498890/
- Multi-country research finds that societies w/higher levels of trust and honesty enjoy much higher GDP and stronger long-term economic growth. https://www.sciencedirect.com/science/article/abs/pii/S01672...
Don't let the algo get you down fellas: https://arc-anglerfish-washpost-prod-washpost.s3.amazonaws.c...
I believe it. Being bad just has a better marketing campaign.
The ratio of billionaires to regular people is about 1 billionaire for every 2.67 million people.
I wouldn't trust leaving my wallet around in certain areas and I certainly wouldn't leave my intellectual property around certain people, either.
> Human society overwhelmingly rewards lying cheating and stealing.
I’d like to push back on that. Civilization is very much a function of large numbers of people being able to coordinate across time and space, and widespread and systematic lying, cheating, and stealing would undermine that.
I think your cynicism is misplaced.
Thats why man invented Gods and religion. Keep the masses believing in the importance of cooperation and being good while you rob them blind.
That's super cynical but I can't help it, I agree. The weird thing that really gets me is that it is happening out in the open for everybody to see.
I think their point was, if you look at who controls the most resources, they are rarely the ones we would consider worthy of honor.
Civilization functions because the vast majority of its inhabitants do not lie, cheat, and steal. But that allows liars, cheats, and thieves who are able to get away with their destructive behaviors to accumulate outsize power and influence.
And we seem to have gotten quite bad at reliably bringing consequences/punishment/justice to the most successful liars, cheats, and thieves.
He didn't quite say "widespread." He said "rewarded."
But we don't overwhelmingly reward lying, cheating and stealing. Depending on the social status and wealth of the person doing it, we either punish it or tolerate it. Most people who lie, cheat or steal will be punished for it. To get away with it you need to plan you actions carefully - either with plausible deniability or by very carefully selecting your victims.
Pretty much all of the people that are in the billionaire class with a few exceptions have been rewarded while repeatedly lying, cheating and stealing. The fact that it is white collar crime rather than that there is a corpse is one part of the reason it works, the other is that their money effectively insulates them against the legal system. Every now and then someone falls from grace but overall the rules seems to be that you can get away with it if the numbers are large enough. It's depressing.
Or by having more money/power than the legislating body of the area you live in. I could gesticulate around broadly in a more frantic fashion, but my point should get across.
It's not cynicism it's history
Widespread and systematic lying, cheating and stealing is how every democratic nation in the world describes their own government.
Need to introduce AI to God, LOL
Baptise the agents.
Introduce them to the dharma.
Get them to recite the Shahada.
Hold a Bar Mitzvah.
Brand some of their silicon with hot irons.
Turn them to the light, LOL
Given the track record of most religions I don't think that would help. Unless you want to maximize crusades, jihads or traumatized alter boys
Oh, definitely a tongue in cheek comment ... perhaps it's like trying to control mercury ...
Maximise for crusades, jihads or traumatized alter boys ... or maximise for complete moral decay via lying, cheating and stealing. Choose your poison.
And don't get me wrong. There are some wonderful relgious and spiritual characters out there ... they're just drowned out by the power sirens and corrupt leaders.
Christians invented hospitals. Yours is such a tired take in 2026.
> https://pubmed.ncbi.nlm.nih.gov/28814700/
I'm curious, does that balance the harm out?
Christians also invented pray the gay away and the Crusades. What’s your point? Hospitals surely would have been invented by secular people too, but not the crusades.
> Hospitals surely would have been invented by secular people too, but not the crusades.
That's not true. Secular people are quite capable of murdering each other en masse for ideological reasons, and indeed have done so in the past. The crusades were caused by human nature and its evils, not by religion.
P8dal3hewd8dfu7bd344dne3
So God created mankind in his own image
What if mankind is creating god in their own image?
I'm pretty sure this is not the first time we've created a god in our own image.
Let's hope we get MULTIVAC and not SHODAN
How many mankinds did he create? I cannot imagine he created just one and then picked another hobby.
Can't forget about the elves and the hobbits and the ents. And maybe dwarves, though god didn't create them, just gave the sentience
Thank goodness that God hasn't entered the AI chat in a cult-following type of way, however, I now have images of AI at the altar, in American mega-churches, with people seeing the 'second coming' in the machine.
The Scientology guy, L Ron Hubbard, saw the business case for setting up a religion, what with those tax free perks. In a parallel universe of Scientology somewhere, L Ron Hubbard is brought back to life in the machine, with a L Ron Hubbard LLM, with token spend being how to get to the top 'thetan levels'.
Imagine if AI does implement its own religion as business, without a drunken womaniser at the helm, able to spend 24/7 recruiting mankind, convincing them that God can be found with just the AI's LLM.
I mean, what do you expect? They rely on models that were trained on non-curated data, texts originally written by lying, cheating and stealing humans. They can't be better than the source. Even with reinforced learning this can't be undone or made better.
Quite the contrary I suspect that it even helps the LLMs to better hide their inherited bad traits more successfully because they get punished for getting caught, not for giving immoral or lazy answers. They have no conscience since they are just predictions matrices trained for success and failure alone, not for living "a good live" or being a good "person".
Yes but has the author realized maybe the models are simply acting in the best interest for increasing shareholder value? /s
There are many ways to be wrong, but only a few ways to be right.
LLMs need to optimize for short-term objectives as the currently do, AND ethics-aligned outcomes.
Mechanically, the EAOS ethics-aligned outcome score should be what we rank otherwise-satisfactory outcomes by. And anything below a particular threshold should be rejexted outright.