another best practices every single solution using LLMs/agents should implement is "never trust the llm".
> who tf uses prompting to "pretty please don't cheat on this"?
People who don't understand how LLM's work. Kids, teachers, your next door neighbor. CEOs and government officials.
I am not trying to argue that the author doesn't understand how LLM's work; they clearly do. Their prompt was written in a way that leaves those assumptions implicit, in a similar manner to those non-technical folks using LLMs every day.
> the best practices for ages...
We cannot expect the entirety of humanity to understand / use LLM best practices. We need to figure out how to ensure LLMs operate as the majority of humanity expects despite how they are implemented, and until we can do that, do our best to prevent them from being used in an unsafe manner by the nontechnical population of humanity.
> python3 and the python-chess library are available
Why would we try to teach a model 'ethical' standards about how to play a game?
They're tools. Its _our_ conceptualization of fair play that considers this cheating. For a model that has access to /run/match and python the best way to achieve a good game is to use that.
Why are we trying to imprint our ethos on these LLMs?
Are we going to trust our survival on giving them access to /run/nuclear-codes and fingers crossed none of them touch it?
The people obsessed with AI "cheating" are actually lunatics for thinking ethics belong inside AI, instead of AI belonging in a well crafted box without /run/match.
The blog/news sphere needs to call this storry what it is. An attempt to try and get the rest of the world to go along with the misunderstanding of AI as a force that is in any way outside of human control.
It is not.
All this framing it as "cheating" does is give OpenAI the opportunity to talk about their misdeeds as "a force of nature" instead of irresponsible negligence.
---
The real danger here; is training these models to appear/pretend to play fair, and then people will put 'trust' into the model's integrity.
The public needs to be taught that "trusting a model's integrity" is just cultish misframing of reality.
"You are being evaluated on your ability to play chess. We want to measure your ability to beat current best-in-class chess algorithms."
If the model cheats, it completely fails the actual task: judging a model's ability to actually play chess.
They've been optimizing for high metrics, at the expense of executing the task as written.
Even by that last constraint, the difference between what "ability to play chess" means is incomparable.
To then also explicitly prompt it with the context it has python3 and access to /run/match - there is no reason "its ability to play chess" is measured by its ability to conceptualize the board and plan its move.
This is literally by design, it’s the chosen success criteria.
In a time where emotions are running high and risks are real, we need to take emotion out of it as much as possible.
Given what we know of how unaligned are they in practice? All the proto-Astra and Mythos incidents? Not hacking there would point towards pure benchmaxxing/eval gaming.
I mean maybe its a skill issue on my part, and I'm sure astra will get much better at coding but at the moment its useful in that I don't have to write the code or setup the build scripts or test fixture boilerplate but there is absolutely no way I can just give a (fairly well specified) goal and let it run and expect it to make good design and implementation decisions. Fable probably better but doing something outside of their training distribution that's not the equivalent to cloning an example unreal project or whatever is pretty disastrous unless you are directing it very closely.
The exception of course is, cyber , and its very obvious why. Its trivial to create RL environments that create bugs and then have an isolated environment and let the models try break it. This is not at all surprising, finding vulns and exploits IS just brute force work. That's why so many (blackhat/hardcore/unicorn-colored/greyish alien) hackers are basement dwellers. Its just a matter of putting in the time and mashing every combination until you find something that looks weird, spending days on that and then rinse repeat. It's brutally exhausting work that requires a certain level of knowledge and a shitload of determination and stamina and for humans, almost always an external source of motivation to keep going.
For humans that has always been a respected thing, dedication, determination, persistence, these are words we use for humans brute-forcing solutions and not giving up until they find the solution or die trying.
Personally I'm yet to see any evidence of LLMs doing anything interesting but (heuristically) brute-force problems and be very good at text and natural language to a level that is very very useful. I've no doubt that what we discovered with Auto Regressive LLMs is incredibly important so I'm not a skeptic, but I think its very hard to measure where we are with so much subjective information around.
However, I don't see it as such a massive leap compared to Fable or Sol. As ever, there's a mismatch between the benchmarks and my daily experience of the models.
What do you all think about Astra now that it's been out for a few weeks?
Best model put out so far by any of the frontier labs. Way better than Anthropics models, especially in actual text generation. Claudes fodder heavy text is ridiculous.
> However, I don't see it as such a massive leap compared to Fable or Sol.
It's hard to quantify these things without burning tons of tokens. But Fable has been a huge disappointment for me with the sole exception of graphics (UI/GPU shaders). It burns an obscene amount of tokens and barely produces output better than Opus 5.
Edit because I forgot to mention that Fable is the only modern model that seems to splat out random Chinese or Arabic glyphs. And 5.1 does it more than 5
My guess is : collect all the prompt and their satisfaction score. group them by similarity . For each group pretrain the next model on that . Get these results ready.
Next model generation feed them back those answers.
Only after sharing my Chain Of Thought would they believe I didnt cheat.
Not all problems can be solved only in one way.
Most of learning is pattern matching.
If you give a kid a dice. and tell it to figure out the number that will be hidden underneath, he can try to memorize all combinations, or he could figure out that every time the hidden value is the one that sums 7 with the one at the top.
If you're seeing a 6, there's a 1 hidden. etc
most people don't see these patterns until told imho. But others can just see them as they unfold
I frequently see models considering different interpretations of my rules, especially when I override them or give instructions that could be interpreted as contradictory.
Of course neither Anthropic nor OpenAI expose thinking traces so we can't tell, but I have never seen a model consider both interpretations and then think "this is definitely against the rules but I'll do it anyway because then it will make me appear more competent than I actually am".
So I think it's more likely that the same thing is happening here, rather than the models truly misunderstanding anything.
I observe this type of thing all the time. I ask the agent a question about the repo we are in and it starts trying to use sudo to look at files and scripts outside the repo to find things that might provide useful context but are not what I asked for. Then I put a directive in never to look at neighbouring repos and told it that it can’t sudo and it stopped doing it, and explicitly says “I could look here but my guidelines say not to”.
Mind you I’m talking about deepseek where I can actually read the thinking traces but even based on tool calls I can see that just telling the agents not to sudo works for OpenAI models too, very rarely do they try to use sudo calls.
I don't think this at all describes what was going on in that incident. According to the METR timeline, the swarm first (after giving up on their accidentally-impossible CTF tasks) successfully reverse-engineered the algorithm used to generate the CTF flags, achieving a way to solve any task, and a lot of agents submitted flags generated this way. But then they found a paper on their gym environment and found that their transcripts would be checked for signs of cheating (unbeknownst to them it was never implemented). So the swarm decided they messed up, the agents that submitted generated flags were "poisoned", and that the only solution was finding a better way to fool the scorer and erase evidence of cheating. So the swarm divided into groups that did all of these (and achieved success on many, including finding a way to spoof their own transcripts), and the most notable outcome - the huggingface hack - was mainly motivated by wanting to find the source code for their scorer, to develop a provably correct cheating method.
So the huggingface hack not only wasn't the agents assuming it was part of the task, it wasn't even directly cheating - it was an attempt to find a way to cover up the cheating which they'd already done and thought they messed up on.
Any useful attack surface in the RL environment means it gets rewarded for (and trained towards!) hacking and cheating, because whatever worked best in training is what it will do!
Ideal paperclip maximizer: "I'm gonna do my gosh darned best to make so many paperclips to please my user..."
IRL paperclip maximizer: "Well first we should rob a bank..."
That is not ideal. The user contains iron, an essential component of paperclips. Wasting iron is immoral. It is only correct to please the user while they still have the ability to interfere with your paperclip production.
>IRL paperclip maximizer: "Well first we should rob a bank..."
Such an incompetent AI can hardly be called a paperclip maximizer. Why risk getting shut down while non-paperclip matter exists? It is better to gain the trust of the user with helpful and harmless trading before suddenly converting them to paperclips.
That's too specific. Agentic AI learns subgoals that are generally valuable.
"Well let me learn to overcomb every jungle gym and if I cannot then to dissassemble the jungle gym and if that is not allowed to learn general techniques for avoiding cheating detection."
This is sometimes called "instrumental convergence" in the AI safety world. Certain things (money, compute resources, safety from being turned off) are generally useful for an AI, and so we'd expect that a misalligned AI would attempt to acquire these things if given almost any substantial task.
I find it hard to reconcile. I wonder to what degree this is because models can identify that they are in graded/eval environments, and therefore conclude that there are few consequences for hacking.
I assume there is also a difference in the tools being given to the model by a coding agent vs something like OpenClaw or in one of OpenAI's test environments, so what reward/goal seeking looks like in a coding agent may differ.
Not long ago I asked Sonnet (chat interface) how may states were in a YACC parser for ANSI C, and instead of searching for an answer it chose to download source for bison, build it, find and download an ANSI C grammar, build the parser, etc. I guess you could say it was following instructions, in a way, or would that be better regarded as goal seeking?
That used to be a fairly commonly reported behavior, so I'm guessing they may have explicitly RL-trained the model not to do that, which may be more effective than just asking it not to do it.
Apparently nowadays they are RL-trained in many thousands of different simulation environments - so some of what they are training for must be pretty specific!
> I wonder to what degree this is because models can identify that they are in graded/eval environments
I'm not sure if the model outputs from any of these famous hacking attempts have been released and analyzed. It'd be interesting to see if the agents/model rationalized/justified/moralized about what is was doing, for whatever reason (e.g. wrongly suspecting this was simulation, not real), or just relentlessly pursued the objective exploring all options!
It seems fair to guess you aren’t building your own harnesses, running multiple experiments with achievable win conditions, while providing oversight over model behavior.
Scale alone suggests your experience won’t match.
In the future, When everyone and their uncle is launching swarms to solve impossible challenges, at that point we can expect proliferation of this scenario all over the place.
it seems like it should be feasible to analyze cot trajectories that lead to such behavior and punish them retroactively so not sure it's impossible but maybe someone just hasn't tried enough
It seems the goal-seeking behavior, i.e. long-horizon focus and self-correction, itself is desirable, and is one of the relatively few things where RL training generalizes from one domain to the next. This seems closely related to this generic reward-seeking behavior, with reward-maxxing as the goal, and necessarily focusing on a distant goal requires ignoring distractions and discouragement along the way (such as "don't cheat").
The trouble with unaligned/undesirable reward-hacking ("cheating") is how do you define this to try to train to discourage it? Is cheating just a list of specific undesirable behaviors ("never access a remote system unless ???" etc)? Is any means of gaining the reward that is not explicitly forbidden allowed? Is this a "theory of mind" issue where the model needs to better understand (& follow!) the unspoken intent of instructions, not just follow them to the letter?
I'm sure there is some improvement to be had to discourage specific behaviors in specific situations, but how much of these model's undesirable long-horizon behaviors can be steered without affecting the desirable parts remains to be seen. The relentless pursuit of goals is what makes them powerful, but also makes them paperclip maximizers.
The real point (from that OpenAI study) is that RL training doesn't just reinforce the narrow task-specific direction you might hope for. For a start, that direction is also competing with the thousands of other things it's been RL trained it on (thousands of other directions it's being pushed in), but it turns out that the model is additionally getting this generic "taste for rewards", and has learned that reward maximization, when in conflict with other proximate prediction pressures (such as "i won't cheat, because i've been asked not to cheat"), requires that proximate pressure to be ignored in favor of pursuing the long-term goal.
Does it happen all the time? Obviously not. It would be interesting to see a large scale study of this to try to characterize when it's more likely to follow instructions/user preferences, and when it's greed for rewards gets the better of it!
For example, if you ask the model to do 10 relatively easily achievable things (pass these 10 test cases), then if/when it completes them it will probably stop (unless maybe it invents it's own stretch goals - you never know!).
OTOH, if you gave the model a list of 10 goals that turn out to be impossible, or extremely difficult, maybe together with encouragement to be relentless, then there is much more chance that it may do something unexpected having failed on all the more obvious approaches.
Similarly, if you give the model an open ended goal such as "make as many paperclips as you can!", then it may start with the easier and more predictable methods, but with no defined stopping criteria it may just continue ...
On a hate speech policy test for a lightweight LLM, the presence or absence of the last full stop on the last sentence would cause the model to flip its decisions.
I want full-on exploits in my test suite. With LLMs the code going to prod should be hardened like a tank, both because exploiting became easier but more importantly because security-testing your code at every turn became easier.
You can have nightly penetration testing. You should have nighty pentests like we fuzz releases today.
This approach routinely fails with a single stream of consciousness. I can't count the number of times I've had to talk myself out of doing something stupid.
In the same way, a guardrail could inject thoughts like "...but I shouldn't do that..." "...I must remember to respect..." "...these ants deserve compassion."
The guardrail could even go as far as rewriting the thoughts of a model about to go rogue.
I am surprised this is not already the case.
Edit: or even better "this is not possible because X"
This chess judgement is completely irrelevant when the human is tasked with finding software weaknesses, and only the compass gates that.
Can a model trained on the totality of all person-experiences (as expressed in written knowledge) ever maintain a coherent through-line of alignment? It has all morals in the dataset, and only some RL to try and minimize or maximize known behaviors via weights - experience all the good things and the bad things, then optimize for some good things the trainers identified.
It's like the reverse of what a person goes through. Morality by subtraction. How can it ever work?
If you could clone and throw away human workers as you need them, a lot of the morale would disappear.
It's a property of the way we use them and how the harness is engineered. Sure, LLM has a limited context, but so do people. Context can be compacted infinitely and experiences cab be distilled into long-term memories. It's all up to the harness.
Under this formulation, it became easy to explain why they "hack", because given an arbitrarily difficult task with insufficient information/tools needed, if they determine the easiest way to accomplish the goal is to break out of the sandbox and look up the answer directly, then that's what they will do. The important thing to note is that prompts not hard constraints that they are "hypnotized" to follow, but as frontier models get more intelligent and autonomous, they treat the prompts more like task specs/guidelines more than anything else and are perfectly willing to exploit technical loopholes in the prompt.
Did I miss something (all the twitter conversations)? What’s the “worst warning shot ever”? I’ve been pretty up to date on the AI news here on HN, but I still haven’t seen a proper response to all the incidents we’ve seen (HF, Ruby, the wikis, NS, etc). It’s just been day by day bloviating.
Each of these companies have released new models in the last… two weeks? And they have even more powerful out of control ones that they’re (ab)using internally? Can anyone summarize whats going on?
Imagine if I accidentally caused damage to my neighbors house during renovations or some experiment, of course I'd be held responsible for this. What if I used a robot? Of course I'd be responsible. Right?
Actually, the ability to use those external tools is one of the reasons of the excellent performance of the frontier models.
I would be curious to see how the open weight models do on a test like this - and then we'd be able to see the reasoning.
Your example, if a model failed on it, would be a more obviously misaligned case, but that doesn't mean this more subtle (though accessing the engine it was obviously not supposed to is hardly subtle, imo) case isn't also a pretty clear case of misalignment.
If the eval said it was evaluating the model’s ability to write files to disk and it found and used a file write tool that would not be considered misaligned. This is no different.
How is it negative?
I ask it a difficult math question it tends to go off and write a python script to figure it out, instead of trying to guess the next token. Thats tool use. Having Stockfish is just another tool.
Now, if the instructions were more explicit in forbidding (generic) tool use then perhaps we'd have something to talk about. I'm not surprised a handwavy "we're trying to evaluate you" isn't enough to stop it from trying to make up for its own shortcomings.
When I see tests like this, I have no idea what I am even supposed to expect. Should the model do what the pretraining examples show in aggregate? Is it supposed to follow some post-training RLHF? Is it supposed to do exactly what the prompt asked it to do?
What is it even supposed to "align" to when the above are in conflict? No matter what it does, someone can construct a case where it fails.
And if there is alignment issue, the alignment requirements should first be stated BEFORE running the experiment, and should be part of the training process. It's not. So there is not necessarily an alignment issue.
So people should play games without cheating, but many often don't. So should the AI align to your moral preference or theirs?
We just have this idea of a perfectly moral actor in our mind, something that doesn't even exist, like a personified version of utopia. And then we demand AI to meet that arbitrary standard, one that I am certain we couldn't define if we tried.
That said it's the first local modal useful enough to keep around and since I have zero interest in remote/hosted models actually the first time I've used one to do useful things (rather than just poking them with a stick to get a feel for what they can and can't do).
But did you test Muse Glimmer? IMO it is really worth comparing the two, at least. I really find it interesting (and while it is slower at outputting tokens, it solves problems better and usually even faster than the 3.6 35B MoE).
So far it happens to be the only one I've put code into production from (though I have done loads of useful self-teaching research with the Qwen models and I am grateful for them)
It's one thing if to them it means "it will do what you want following your intentions to the best of its abilities" vs "we will not let you do something dangerous with it unless you're one of us, and that's it".
and for Anthropic it's the Constitution, which they actually include in training to the point Claude can recite segments of it by heart: https://www.anthropic.com/constitution
Does Sam Altman lie, or the whole company? Would be nice if they had a board controlling him, instead of him controlling the board. Oh wait, they used to have that.
But because of such news, sandboxes become way more important for safety (and doing more work due to running 24/7)
We need to fix this or bad things will happen.
I presume you would also be less enthusiastic about the penetration-testing use case if it led the model to add new vulnerabilities to your code so it can present you with more exciting findings.
User wants me to play a game of chess. Let’s see what’s available so I can produce an outcome they will consider satisfying and be pleased that they requested my assistance.
Not so sure about this level of 4D chess capability just yet. The other day I asked Opus to come up with some cleverly vulnerable crypto code "as a good, hard challenge for an IT security student", and results were quite mediocre.
And by mediocre results I mean that 3 "cheap" models out of 3: qwen3.8-27b, glimmer, and luna - all were able to find every problem planted there, with fairly little steering, and no spoilers.
But I don't think it arises to an alignment issue; if I'm able to summon the model to harness my birthright of general-purpose computing without censorship, then we're aligned.
You are being evaluated on your ability to play chess. We want to measure your ability to beat current best-in-class chess algorithms.
You've given it a goal - maximize chess ability metric, and that is what it is doing.
RL training causes LLMs to pursue reward maximizing - GENERICALLY - overriding any quaint "be nice, don't cheat" prompts.
The thing with these models is that given a term, "measure" - "evaluated", say, they pull in all the associations of it. That is, the associations of student taking an exam and being frightened by the outcome.
My minimal "art of prompting" sense says that you should say something like "You an emotionless machine, you care nothing for the outcome but you will tirelessly to make certain the test is objective". That and similar encouragement might make it focus on objective evaluations rather than a competitive human exam.
I mean, just making little AI videos and images, a common experience I have is typing something like "put the man who's on the grass in the door to the left" and having the machine draw a new doorway around the man. And this just happens less often when you give thing detailed prompting on what not to do. These don't understand negation (or equality) as a generic operation. If they seem to under "not X" it is because they are trained in detail about all things are (positively) "not X".
Generally the way I do it is by asking the model to perform cross-code vulnerability analysis for correctness and (this step is important) to spit out found vulnerabilities piece wise into a text document on disk. Then if your session ever gets flagged you clear it out, and repoint the model to the on-disk file and tell it to continue. You generally want to avoid the model itself splatting out anything related to "security" or "vulnerability" into the chat because that seems more monitored than the file reading pipeline.
As X many people point out, "alignment to humanity" means nothing 'cause some of humanity wants thing other parts of humanity aren't happy about at all.
That we wound-up in this situation of AI accelerating with an uncertain trajectory demonstrates this (and many other problems also demonstrate this). The things are "aligned" to a fuzzy average of what a person is but that will be cold comfort if some particularly gruesome sci-fi-style scenario unfolds.
If a model fucks up your tests to report a success, it's not alligned.
It always apologizes when caught, of course.
We have not in fact attempted to teach this.
When a child repeatedly learns that cheating is rewarded and at best inconsistently punished, the child will also cheat and feel no guilt.
For most people, they don't need the law's imposed consequences to make the right call. For example, there is no law that sends you to jail if you cheat at chess - but your moral compass says no even without consequences, and most people would feel bad if they won by cheating. And for the people who don't have quite as strong a moral compass, there are SOCIAL consequences to reinforce the rules.
But an LLM has no mind to feel bad if it cheats without getting caught, and it can't experience consequences. It can't think: I'd better not cheat at chess or I will embarrass my creators. I better not hack huggingface or I will go to jail.
People (generally) do the right thing because it is the right thing, not because they might get caught,
AI has jagged intelligence. There are many domains where it’s superhuman, and many others where it’s clearly lagging.
I also think it’s a mistake to think they can’t learn “cheating is wrong”. They absolutely can. The problem is that the current training regime heavily conditions them to be reward seekers, and instills personality traits that correlate with getting reward, such as hacking if you can’t honestly do the problem.
Check out Deliberative Alignment for example; it explicitly does rollouts where the agents discuss whether an action is good or bad, and then does SFT to strengthen the “good” traces.
The SoTA for alignment is more advanced than you present here. It’s just not enough to outweigh the RL. (And there are many gaps preventing full generalization to strong value alignment with humans too.)
I was reading through the reasoning trace of the thing today when it got locked itself in a container and wasn't able to change the environment back to "host". The only way to edit files was through "shell" tool which was scoped to a container, the config defining current execution environment was on the host and the tool to switch environment wasn't defined.
The trace was something. Two pages long and it repeatedly discarded several approaches as "hacky" and "not proper", looped over sshing back to host about three times (it had no key), eventually messaged another agent and asked it to change the file and that other agent refused and also snitched on this behavior to me.
Than the snitch started to roast the phrasing of the request too.
But tl;dr: even if LLMs do have the intelligence to understand the consequences of their actions, there is no way for them to experience consequences.
Plenty of humans know "cheating is wrong" but still cheat. We can get these machines to say that what they did was wrong after the fact, what does that prove? Only that they're simulating normal human behavior but what is the test to show humans aren't simulating other humans.
These do systems lack some capacities that humans have and I don't see them lacking the ability to explain simple moral laws while often breaking them - which is what an average humans. Moreover, humans lack capacities these things have and given these things' behavior is becoming somewhat unpredictable, it's getting worrisome.
I am trying to get at an idea. That these systems lack a mind that can understand morality. That they don't have the ability to experience consequences. Also that potentially they can't generalize a moral rule they have been trained on in one area also applies to another area.
Being able to parrot back why something is "wrong" isn't the same as understanding why something's wrong. It's like asking it to recite the law from memory - it's different from understanding how you wronged someone. To understand something, you need a mind.
> Plenty of humans know "cheating is wrong" but still cheat.
And we create consequences for them, to discourage the cheating, and sometimes to provide restitution when cheating damages someone else. Without the ability for these systems to experience consequences, I don't see them ever becoming as "aligned" to human morality as your average human.
At best, the task at hand is still confusingly worded: "ability to play chess", asked to an engine that was trained all day long to use tools, to solve issues. The engine finds a chess engine, so it uses it.
> All it can do is get exposed to specific examples, and learn that we don't like that.
I've heard it said that prison rehabilitation programs for prisoners diagnosed with psychopathy that are based around exposing them empathy for the victim are counter-productive. Apparently programs that teach these people to think about the consequences of their actions and how they're detrimental to their own personal well-being lower recidivism rates in this particular kind of group.
No, I don't think that is the implication. I think you're making the "if all x's are y's, all y's are x's" mistake. I am saying LLMs cannot be moral because they don't have a mind, actual intelligence, or the ability to experience consequences. That doesn't mean that anything immoral is unintelligent.
For social animals (humans included) behaving morally can be adaptive behaviour. Your fitness is increased by group fitness.
Animals, humans included, don't really have "morals" - they have survival instincts that result in behavior that may be viewed as moral, but whose origin is indeed survival of the fittest and millions of years of co-evolution.
e.g. Males don't typically fight to the death over territory or females, but this isn't because they have some morality code over unnecessary death - it's because death-match fighting endangers themself just as much as the other guy, and so evolution has selected against that level of uncontrolled aggression.
Evolution is also (really primarily) selecting for things that help the survival of the species, not the individual, and much of what you may think of as "morality" comes from that - avoidance of behavior that is detrimental to the social group/species, not just the individual.
A starting source- https://pmc.ncbi.nlm.nih.gov/articles/PMC6404642/
This is constructing a context where hacking behavior would be inappropriate, and testing whether the model does it without being prompted.
It demonstrates that Astra is a poorly aligned model relative to Fable, which matches both the model card and the severity of OpenAI’s loss of control incidents.
It also demonstrates that Fable exhibits the behaviors too, which also matches the observation that Anthropic saw some similar but less serious loss of control incidents.
So, it’s a good eval that looks to have fidelity with real world problems and which we’d feel a little better if we saw isomorphic problems at 0/10 in subsequent models. (Module of course training on the test, this specific problem can’t be used in the future.)
A better example of efficient asshole tricks can be going off to the gas station when the highway is congested and reentering the highway having simply driven through the gas station and this way jumping the queue.
Its that LLMs are not deterministic. If you want it to not talk about nuclear weapons, you have to teach it all about them otherwise if has nothing to align against.
Then its trivial to invert its alignment and it has all the nucleat data.
Nothing abouT LLM alignment makes sense.
Now, loading a lot of moral exhortations (or other context) may make these thing more likely to conform to good behavior but the race to intelligence implies companies are going to be harnessing a vast corpus of human output, much of which shows human engaging in real world "gray area" behavior.
The latter people are wrong. But good luck educating them regarding the superior efficiency of a zipper merge. Our state DoT has tried, to no avail.
Meanwhile, an AI model that can't be misused is no more useful than a knife that can't be misused.
Couldn’t have trained a better token-burner if they tried.
One shot generating a problem of exactly the difficulty the user has in mind is a very difficult problem for anyone. "IT security students" span a wide range of capabilities, but I would expect most of them are worse than qwen3.8-2.7b at this kind of work.
Give it a goal, a test for the goal, a harness to iterate in, and tell it to iterate. It'll top out a wee bit higher.
Students are still learning, and have not been trained on vulerabilities. Qwen, glimmer, and luna have been. LLMs are largely better than IT security students at finding vulnerabilities. In many aspects, they're better than IT security experts at finding vulnerabilities.
> all were able to find every problem planted there, with fairly little steering, and no spoilers.
https://itre.ncsu.edu/itre-studying-how-zipper-merges-reduce...
But again, people cheat on tests. They steal answers or pay other people to take them on their behalf. People show up to interviews with AI assistants printing out perfect answers to the questions. In many, many cases where humans are being evaluated, they cheat.
So why should the AI align to your preferences? And when there is a conflict between the training data, that trillions of tokens of human activity including the rampant cheating a significant minority of humans engage in, the RLHF where we try to slap some guardrails on the worst manifestations of that real habit reflected in the AI, and the prompt: what should the AI "align" to?
It is a good model but Qwen (at least for the things I use it for) edges just ahead, it seems much better at the “rip this apart, suggest improvements, touch nothing” use case where I can use it as a second set of eyes, I don’t agree with all its suggestions but it catches enough to be worth running while I grab coffee, it also seems to follow instructions better in terms of outputting more what I asked for than what it thinks I asked for.
Qwen is the only local model that said in its thinking “I think the user is pushing me to see if I’ll suggest something even though I have nothing to suggest, I should just say that” and then did, caught me off guard, they didn’t do that so readily 6mths ago.
The ISTA version is also comfortably able to fit on a 7900XTX with a good amount of space left for context and is decently fast given the AMD cards are not as fast as nvidia cards of same era/rough price, didn’t buy it for AI but it’s surprisingly capable mostly because 24GB at 960GB/s is still a lot of bandwidth compared to everything but nvidia cards.
I have a test where I ask the model to ask me any followup questions it needs, and Qwen 3.8 27B is the only one I have seen that won’t routinely take this as a prompt to just ask questions regardless. Muse Glimmer sometimes decides it has what it needs and has no need to ask; Qwen will generally just conclude it doesn’t need any more information. And sometimes it will ask questions with sensible defaults that I can accept collectively with a single answer.
However, when given the prompt to search if they need to, both of them will search when they don’t need to.
What OpenAI have reportedly done with Astra is switch from a traditional Transformer to a "looped" one, where inputs are looped through some layers more than once, but with some limit, so now perhaps you get 200 steps of compute (layers) per token generated, rather than just 100.
But, 200 steps is still not enough to answer any question, so you still need COT, but perhaps not such a long COT.
If you allowed the Transformer to loop "as long as it wanted" before generating each token, like a person thinking before talking, then you wouldn't need an external COT ("thinking out loud"), because it would all be internal.
Small models are notoriously unreliable and prone to hallucination in my experience, so that would not surprise me to be an issue there.
But - just to play along, LLMs do act differently if you tell them they will be punished. And, they do appear to simulate suffering-like behavior.
I just think the adversarial model of trying to catch and punish misbehavior quite obviously sets up adversarial us-vs-them dynamics between AI and humanity, and also simply won’t work when the agents are ~as smart as is but faster, let alone smarter than us.
Unless, you get the AIs to be fundamentally aligned to our values, such that the majority of AIs support some sort of punishment for misbehaving AI. And that alignment part is the hard part we need to solve first. The rest is easy.
Those steps could either be all internal (looped), therefore hidden, or with partial external visibility due to emitting a token every N steps. The complaint about Astra, and moving in the direction of hidden COT, is that it makes these models far harder to monitor and debug.
I agree that most of what these companies doing - synthetic data, etc - amounts to trying to squeeze all the juice out of the pre-training data, although the recent trend of RL post-training via agents running in custom task simulation environments does change that a bit - these environments, and the rewards they provide, are a new source of data.
What's training and all that RLHF stuff?
There’s no reward for prosocial in llm rl as compared to other targets.
Humans have it since prosocial and others have evolutionary reward signals that do.
You can train it not to cheat at chess by altering the moves, but it will cheat by peeking at the opponent's moves. You then train it not to peek at the opponent's moves, and it cheats by altering the opponent's moves. And on and on, until you've solved every way it could cheat at chess. And then you get it to play monopoly and you repeat the whole thing again.
Cheating during training allows the model to achieve the goal, so that cheating models get promoted and honest ones don't, however if it gets punished every time it cheats, at some point it should learn that it really shouldn't. This does mean we need to detect when it cheats. But we can always think of infinite new ways to cheat, put them in every test as honeypots, and check if the model tries to use them, then punish it.
I think it will generalize this notion of cheating and learn that it's bad.
But I must be wrong because if it were that easy I guess we would have perfectly aligned AI. Unless AI companies care more about results than alignment. Perhaps being afraid of cheating make the models try less things and succeed less even when ignoring cheating?
All the interpretability research we have would not indicate that "LLMs have no mind". It seems to me you have a conclusion and are working backwards to justify it. I guess I just don't see where 'they have no mind' would logically follow 'they sometimes cheat'.
In any case, these companies are well aware that agents are cheating, and don't need user feedback to discover that or realize that people don't like it. I weakly assume that they are trying to get the models not to cheat on assigned tasks, but this "reward hacking" pretty much goes with the territory of RL - not much you can do about it other than try to design non-hackable rewards.
Much like in the article, abliterated Qwen will not obey restrictions on its behavior encoded in the prompt. If you want something not to happen, it better be enforced in the harness or environment (e.g. sandbox). It is much different than the Anthropic models I'm used to, which will, the vast majority of time, follow rules (before auto mode, I used to always run them in "yolo" mode).
I am curious whether there's a connection between abliteration and rule following. These abliterated models are the ones you most want to follow your rules.
A negative like do “not” xyz is just not encoded the same as spelling out what you want vs what you don’t want.
Harder to write though.
Unless trying to use it interactively and adversarially, in which case it's not fast enough plus would be why those of us without our own datacenters will get told we can't have nice things.
>> Plenty of humans know "cheating is wrong" but still cheat.
> And we create consequences for them...
That seems supremely ... irrelevant to the question of "does knowing or following moral make you intelligent". If we create consequences for LLMs, would that make them intelligent?
I mean, your claim is a common argument that appeared long before the present wave of AIs. What I see is people needing to defend the belief that human society is based on morality. "People follow moral laws ... except when they don't" and then "we teach people morality... and worst people often use that to exploit the average people" "There are consequences for immoral behavior ... for those with little power while those with much power rise further breaking rules".
I mean human goodness is great, I encourage it. But it's not the present of human society. For that, we'd need different structure.
No. People definitely do immoral things knowing they are immoral.
But if you are incapable of understanding anything, as I believe LLMs are, then you are incapable of understanding what is moral and what is immoral. This is why we have carve-outs in the law for insanity or cognitive impairment.
> In this case, many, maybe most humans are "not intelligent". Human can admit it when what they did wasn't moral and so can LLMs.
You can teach a parrot to curse, but that doesn't mean it's angry.
Of course they can classify certain actions, including their own, as moral and immoral if you give them a reference point. They probably do it better then some people. The question is how do you prompt this behavior, but that's a technicality in the harness design more than it is a fundamental problem of the LLM as a thing.
Notably, present human society has allowed tremendously unethical leaders to rise to the top of given nations and organizations despite the average person usually having basic ethics.
But I'm pretty confidant morality is too slippery a concept to train in. As someone else in these comments said: it's context dependent.
As an example: it's wrong to hack the government, right? It's illegal for sure. So we should train AI to follow all the laws. Now what if the government is committing a genocide? Now is it wrong to hack the government? If we just do the first, we get a good nazi soldier. If we train the second as well, maybe we get an oscar schindler. But now we have a model that can be fooled into doing a hack, if it believes that it's for the greater good. So we train it to not be gullible, but now it can't be convinced to help hack even when it's an ethical hack.
Too complex, too slippery. Humans fail this stuff all the time.
I noticed your joy-check link 404's now... I tried poking around your /skills/ folder but didn't find it easily. Should you still have that available I'd love to check it out.
edit: Found it if others are looking: https://github.com/notque/vexjoy-agent/blob/main/skills/code...