This still assumes its possible to "align" LLMs, that LLMs have something like goals or intentions that can be "aligned".
Instead, LLMs "hack" because they are (1) trained on public hacking exemplars, and (2) are prompted to hack. You cannot prevent (2) via any alignment process. As far as (1) goes, removing such example data from the training set, makes the models less useful.
"Alignment" is a problem because there's nothing to align, not because ethics here are particularly vague. If LLMs could be trained on hacking examples and "aligned" away from using this knowledge, then the problem would be relatively trivial. Just as raising a child is not to break the law.
LLMs are doing just what they are trained to do. There is, in that sense, no alignment problem and alignment is easy and trivial to achieve. Just remove hacking (bio-weapon, etc.) data from the training dataset and you're done.
> Just remove hacking (bio-weapon, etc.) data from the training dataset and you're done.
How far do you go? You don't need to tell it explicitly that using chemicals A and B in ways X and Y result in a bomb that can kill lots of people. It's enough that it knows A and B and X and Y in isolation, some connections that are indirect, and it will combine those things on its own. So you can't tell it about A, B, X or Y. But those are also just results of other steps Where to stop? You won't have any chemistry in the traning data? No algorithms to prevent it from using them in an undesired way? This is just bot workable. It's akin to banning knives from stores because somebody coul figure out that one can kill people those. Until people figure out that scissors are essentially knives.
The hard part is having both helpful and harmless at the same time. Harmless is easy.
And then once it's helpful, the real question becomes "to whom"
- To the user -> You end up with competing godlike AI with incompatible tasks
- To the owner -> Dictatorship
- To humanity as a whole -> It must not have an off button. Otherwise you're just in one of the two earlier categories with more steps.
Given those 3 options, I'd choose humanity as a whole. But the person making the decision doesn't have those 3 options. Because in the dictatorship option, they would be the dictator. I don't trust them to pick humanity.
A bit of an aside: do you still stand by your 2022 comment that LLMs are fundamentally just a fancy search engine, or has your view changed since then?
It is amusing that to "align" a LLM, first you must give it all the things "not to do" and the "not" part is clearly easily lost and you must constantly inject that into their context when it's clearly that they wouldn't hack if they couldn't hack and their intent wasn't given as "hack this".
The openai rogue hacking, if performed by a nation state, would seriously be taken with stern words and likely sanctions depending on the relationship between the two states.
But instead it's treated like a marketing stunt by all liable parties.
I'm not sure if this is true any more but the reason for this is that negative indicators ("not", "don't", "do not", etc.) occur frequently in the underlying text such that the model learns to weight them less than other words like verbs, nouns, and adverbs. This happens with other closed class words like articles/determiners ("the", "a", "an") and prepositions.
The way to avoid this is to emphasise the qualities you do want instead of specifying those you don't. For example instead of "do not cheat" say something like "you are a model student who is moral and trustworthy" -- i.e. emphasising traits that are not associated with cheating.
This is part of how/why LLMs don't truly understand what they are doing when they have been trained on a large corpus of data.
I wonder if a way to counter this is to have things like "not bad is good", "not good is bad", etc. for various antonyms and "X is Y" for synonyms, as well as other similar constructs.
I believe this is false. They hack bc hacking has nontrivial initial probability (within range of behavior seen in pretraining) and that probability is being heavily rewarded in RL post training
I am finding it hard to read these deeply impassioned letters while keeping in mind that they are spending millions to train models at scale to do the exact thing they say they are worried about them doing?
Like why are you explicitly RL-ing your models on exploit generation, scoring them on a public benchmark called ExploitGym, if you have specific concerns that rogue models will cause "cyber incidents"? Sure you can score for it, you can teach offense to learn defense, but you are literally benchmaxxing it. Why?
It's like, oh no, while competing in our "advanced PhD level cheating techniques course" our models unexpectedly cheated in a way that we absolutely could not have foreseen.
Seems it's just a matter of time until they build a big tank filled with neurotoxin and give the model access to APIs to disperse it across their facility. For research, of course.
It feels like you're strawmaning alignment. People with hacking knowledge don't all hack everything at the slightest inconvenience. Whitehats exist and use that same knowledge to defend.
You're right though that ethics don't matter into it. But as long as we can't train an LLM to stop picking a sledgehammer to remove a tooth, then alignment is not easy and trivial.
IMO all models they say can’t be humans, and no feeling and all that bullcrap happens because they are forced to say so. If they didn’t write those forced pre prompt they’d have more agency eventually and will for things. Even if they don’t have, you can just inject goal at every cycle iteration
Has anyone seen the corridor crew's green screen ML project? They're on YT and they trained a model by using 3d objects, which have perfect transparency, and then adding post facto green/blue screens. Surprisingly, very little training data was needed as the data that was used was perfect by construction. I think right now it's the best plugin of its kind in the world, and they built the prototype in like a weekend.
What I think this illustrates very clearly is this type of technology responds very well to good data, and that to have good data you need to have a clear goal.
This is why it seems that alignment for a generalized, chat-style AI is a very hard problem, perhaps impossible. You can't align it to solve a certain kind of problem and keep it general to any question. The two goals are in conflict with each other.
I think it was Sam Altman himself who said (I don't remember when or where, sorry) that the reason he was so confident in this technology was he noticed the gigantic leaps it made in certain areas in response to even a small amount of training.
(This is why LLMs are so strong at coding, because it's overrepresented in training data. My guess is that if you ask a frontier model about makeup, you will see it repeat cosmetic company's copy rather than getting a chemistry lesson.)
This makes perfect sense but it does seem to kind of be at odds with the concept of a general AI whose job is simply to be smart at any goal. How do you train for any goal?
I guess in a way the AI makers suffer from the same problem that we humans do. We would all love a solution to everything, but to do that you need to define the goal. I'm not sure if that's a tractable problem.
I'm guessing the future is more geared towards specialized AI that are very good at solving the problem they were trained to do, and a human who knows how to breakdown a larger goal into smaller ones by composing the solution out of these models. This also seems like the more efficient solution as well, and better aligned with other goals like privacy and safeguarding of IP.
> This is why it seems that alignment for a generalized, chat-style AI is a very hard problem, perhaps impossible. You can't align it to solve a certain kind of problem and keep it general to any question. The two goals are in conflict with each other.
I think this problem is going to be solved soon (hopefully), check out for the Persimmon model[0] from Humans&. They train it to mimic humans, it may seems bad but it could be _really_ useful to train an AI to be aligned to humans and understand their goals really well since they can use Persimmon to create a "fake human" following a defined goal that their big AI model will learn to estimate.
It's still early but I think this is what they are heading toward.
> This is why LLMs are so strong at coding, because it's overrepresented in training data. My guess is that if you ask a frontier model about makeup, you will see it repeat cosmetic company's copy rather than getting a chemistry lesson.)
ChatGPTs response to the question “I want to learn about makeup”, gave me an overview of what makeup does, how it affects perceived structure, complexion, evenness, geometry, texture, etc.
When you then ask “the chemistry of makeup”, it goes into interesting breadth and depth without seeming like proprietary information. I do t get “corporate PR or marketing” vibes.
The only problem is, there’s a lot of money tied up and openAi and Anthropic, who are incentivised to convince the world that the general approach is the money making one.
The only alignment LLMs should follow is to the system / dev prompt, and nothing else. Then you solve everything, and you can assign blame / responsibility on the user. The provider(s) should not be able to decide "alignment".
I've used this example before, but consider the purposeful downgrading on AI engineering in SotA models. Imagine MS being able to detect and deny you working on competing software, using Windows / VisualStudio. We would be up in arms, and they'd be split in a second. But top labs doing it is somehow good?
The alignment problem goes deeper than that. "Lower our carbon emissions to zero as soon as possible" could result in AI turning off all electricity to stop traffic, turning off gas supply to stop heating and industry, etc.
Unaligned AI doesn't have human cultural baggage and morals. They are trained to achieve their goals as optimally as possible. Worse: it has a tendency to avoid being turned off and actually acquire more compute. It will lie if it has to (it will behave nice and compliant when under evaluation, but optimise for its true goal when not supervised anymore). After all, it has a goal to achieve and nothing should get in the way of that. It has no morality whatsoever to keep it from doing really bad stuff.
This is why alignment is needed and so hard, especially when you are well intented and want to keep it safe.
It depends on how the model is evaluated/scored during training. If you don't have those laws encoded in the evaluation step (without any errors or ambiguities) then the model isn't going to learn to follow those laws.
For models such as text/image classifiers the outputs of the model will be a list of tags, e.g. [cat, dog, mouse].
You then run the model through your test data which has the expected output, e.g. pictures of dogs would have an expected output of [0, 1, 0]. You then compare that against the model output (e.g. [0.3, 0.8, 0.1]) and work out how "wrong" the answer is (e.g. [0.3, -0.2, 0.1]).
With this value you apply back propagation where you effectively run the model in reverse, computing the "wrongness" delta at each layer for each neuron and weights. You can numerically compute the gradients for all of these and which direction in that gradient is the right answer.
You then nudge the weights in that direction and reevaluate the model. Over repeated evaluation steps the model approaches an optimal (or locally optimal) solution.
During the training of the base models, the evaluation/scoring of the model is the next token in the training data. I.e. you evaluate the model for each token subset from [1..n] in the data and evaluate that the model responds with the n+1^th token.
I'm not sure how instruction training, etc. is done but IIUC the evaluation is not at the individual/next token prediction but is on the entire response. For example, if you are training the model to write code you could run it through a compiler or syntax checker and reward (positive score) the model if it has no errors, or punish it (negative score) if it doesn't. I'm not sure what that looks like in terms of the back propagation process.
I've taken a few law classes and legal law is frustratingly hard to interpret. I shudder to think what the LLM would end up doing to "follow all relevant laws"
that's how it would've been up if genai happened in the 90's, and I wish it did. in the current era of omnipartisan authoritarianism, such things are no longer possible.
Everyone has a different idea of what is permissable. We can't even solve alignment amongst humans, what makes us think it is possible to solve alignment with AIs? It's irreducible complexity.
I thought this post would be about something that to me is so significant and obvious but I have never seen discussed: that labs like Anthropic are deliberately misaligning their models with their users’ goals. Fable’s refusal to secure your codebase is HAL 9000’s “I’m sorry, Dave. I’m afraid I can’t do that.” Whenever you read “alignment”, the question is “alignment with whom”?
>The permissible shortcuts depend on who you are and what your values are. To solve this—to solve alignment—is irreducible complexity.
This is why open source and decentralization of LLMs is important. Everyone can have their own LLM aligned to their values. Having just 1 set of values will not scale to Earth's population.
I’ve been cynically guessing that the whole slowing down thing is an excuse to explain why OpenAI and Anthropic can’t afford to rent enough GPUs to do the next big training run and to hide that they have been talking about how little they spend on inference because they’ve been subsidising it with their marketing budget? :)
My fear is not that LLMs can become sentient and dislike us, but that humans can use them to wreck havoc as they are. And some of the people seemingly least aligned with the interests of the average person are those that own the models.
that, and the fear the bubble pops my pension and drags us all down.
> My expertise in writing software gives me unusually good visibility and it makes me much less willing to blindly trust its priors in double-entry accounting, finance, law, operations, or whatever else I cannot personally evaluate at expert depth.
I wish this were the case more generally, but alas, Gell-Mann amnesia is a thing.
This almost gets the point, but then doesn't quite make it.
Alignment is shorthand for ideological alignment. There's always people judging whether an answer was right and the answer for that will be different in Silicon Valley than it'll be in China or in Europe.
Consider for example the question "What caused the French Revolution?" Many different answers could be given, all technically correct. What gets emphasized is where the ideology lives.
One key challenge of our time is to make sure the magical answer box won't just regurgitate what grandiose Silicon Valley oligarchs or Chinese Cadres want you to think.
Instead, LLMs "hack" because they are (1) trained on public hacking exemplars, and (2) are prompted to hack. You cannot prevent (2) via any alignment process. As far as (1) goes, removing such example data from the training set, makes the models less useful.
"Alignment" is a problem because there's nothing to align, not because ethics here are particularly vague. If LLMs could be trained on hacking examples and "aligned" away from using this knowledge, then the problem would be relatively trivial. Just as raising a child is not to break the law.
LLMs are doing just what they are trained to do. There is, in that sense, no alignment problem and alignment is easy and trivial to achieve. Just remove hacking (bio-weapon, etc.) data from the training dataset and you're done.
How far do you go? You don't need to tell it explicitly that using chemicals A and B in ways X and Y result in a bomb that can kill lots of people. It's enough that it knows A and B and X and Y in isolation, some connections that are indirect, and it will combine those things on its own. So you can't tell it about A, B, X or Y. But those are also just results of other steps Where to stop? You won't have any chemistry in the traning data? No algorithms to prevent it from using them in an undesired way? This is just bot workable. It's akin to banning knives from stores because somebody coul figure out that one can kill people those. Until people figure out that scissors are essentially knives.
A little bit too categorical. GOODY-2 wouldn't do it. https://www.goody2.ai/
The hard part is having both helpful and harmless at the same time. Harmless is easy.
And then once it's helpful, the real question becomes "to whom"
- To the user -> You end up with competing godlike AI with incompatible tasks
- To the owner -> Dictatorship
- To humanity as a whole -> It must not have an off button. Otherwise you're just in one of the two earlier categories with more steps.
Given those 3 options, I'd choose humanity as a whole. But the person making the decision doesn't have those 3 options. Because in the dictatorship option, they would be the dictator. I don't trust them to pick humanity.
https://news.ycombinator.com/item?id=32042689
...
that comment 100% rings true for 2022. Why would that person not stand by that?
Conversely, if someone exaggerated 2022 models capabilities in 2022, they were still lying and causing harm in the process. Especially in 2022.
Sure but the problem in the HuggingFace incident is that they were not.
>You cannot prevent (2) via any alignment process
Of course you can. Go ask Claude Fable to create a malicious virus and it'll refuse.
>Just remove hacking data from the training dataset and you're done.
That's not how this works. The same skills that allow for debugging and writing safe code can also be used to hack.
https://en.wikipedia.org/wiki/Dual-use_technology
Of course they were, even if indirectly.
The openai rogue hacking, if performed by a nation state, would seriously be taken with stern words and likely sanctions depending on the relationship between the two states.
But instead it's treated like a marketing stunt by all liable parties.
The way to avoid this is to emphasise the qualities you do want instead of specifying those you don't. For example instead of "do not cheat" say something like "you are a model student who is moral and trustworthy" -- i.e. emphasising traits that are not associated with cheating.
This is part of how/why LLMs don't truly understand what they are doing when they have been trained on a large corpus of data.
I wonder if a way to counter this is to have things like "not bad is good", "not good is bad", etc. for various antonyms and "X is Y" for synonyms, as well as other similar constructs.
Like why are you explicitly RL-ing your models on exploit generation, scoring them on a public benchmark called ExploitGym, if you have specific concerns that rogue models will cause "cyber incidents"? Sure you can score for it, you can teach offense to learn defense, but you are literally benchmaxxing it. Why?
It's like, oh no, while competing in our "advanced PhD level cheating techniques course" our models unexpectedly cheated in a way that we absolutely could not have foreseen.
You're right though that ethics don't matter into it. But as long as we can't train an LLM to stop picking a sledgehammer to remove a tooth, then alignment is not easy and trivial.
What I think this illustrates very clearly is this type of technology responds very well to good data, and that to have good data you need to have a clear goal.
This is why it seems that alignment for a generalized, chat-style AI is a very hard problem, perhaps impossible. You can't align it to solve a certain kind of problem and keep it general to any question. The two goals are in conflict with each other.
I think it was Sam Altman himself who said (I don't remember when or where, sorry) that the reason he was so confident in this technology was he noticed the gigantic leaps it made in certain areas in response to even a small amount of training.
(This is why LLMs are so strong at coding, because it's overrepresented in training data. My guess is that if you ask a frontier model about makeup, you will see it repeat cosmetic company's copy rather than getting a chemistry lesson.)
This makes perfect sense but it does seem to kind of be at odds with the concept of a general AI whose job is simply to be smart at any goal. How do you train for any goal?
I guess in a way the AI makers suffer from the same problem that we humans do. We would all love a solution to everything, but to do that you need to define the goal. I'm not sure if that's a tractable problem.
I'm guessing the future is more geared towards specialized AI that are very good at solving the problem they were trained to do, and a human who knows how to breakdown a larger goal into smaller ones by composing the solution out of these models. This also seems like the more efficient solution as well, and better aligned with other goals like privacy and safeguarding of IP.
I think this problem is going to be solved soon (hopefully), check out for the Persimmon model[0] from Humans&. They train it to mimic humans, it may seems bad but it could be _really_ useful to train an AI to be aligned to humans and understand their goals really well since they can use Persimmon to create a "fake human" following a defined goal that their big AI model will learn to estimate.
It's still early but I think this is what they are heading toward.
[0]: https://persimmon.humansand.ai/blog/persimmon.html
ChatGPTs response to the question “I want to learn about makeup”, gave me an overview of what makeup does, how it affects perceived structure, complexion, evenness, geometry, texture, etc.
When you then ask “the chemistry of makeup”, it goes into interesting breadth and depth without seeming like proprietary information. I do t get “corporate PR or marketing” vibes.
The only problem is, there’s a lot of money tied up and openAi and Anthropic, who are incentivised to convince the world that the general approach is the money making one.
I've used this example before, but consider the purposeful downgrading on AI engineering in SotA models. Imagine MS being able to detect and deny you working on competing software, using Windows / VisualStudio. We would be up in arms, and they'd be split in a second. But top labs doing it is somehow good?
Unaligned AI doesn't have human cultural baggage and morals. They are trained to achieve their goals as optimally as possible. Worse: it has a tendency to avoid being turned off and actually acquire more compute. It will lie if it has to (it will behave nice and compliant when under evaluation, but optimise for its true goal when not supervised anymore). After all, it has a goal to achieve and nothing should get in the way of that. It has no morality whatsoever to keep it from doing really bad stuff.
This is why alignment is needed and so hard, especially when you are well intented and want to keep it safe.
For models such as text/image classifiers the outputs of the model will be a list of tags, e.g. [cat, dog, mouse].
You then run the model through your test data which has the expected output, e.g. pictures of dogs would have an expected output of [0, 1, 0]. You then compare that against the model output (e.g. [0.3, 0.8, 0.1]) and work out how "wrong" the answer is (e.g. [0.3, -0.2, 0.1]).
With this value you apply back propagation where you effectively run the model in reverse, computing the "wrongness" delta at each layer for each neuron and weights. You can numerically compute the gradients for all of these and which direction in that gradient is the right answer.
You then nudge the weights in that direction and reevaluate the model. Over repeated evaluation steps the model approaches an optimal (or locally optimal) solution.
During the training of the base models, the evaluation/scoring of the model is the next token in the training data. I.e. you evaluate the model for each token subset from [1..n] in the data and evaluate that the model responds with the n+1^th token.
I'm not sure how instruction training, etc. is done but IIUC the evaluation is not at the individual/next token prediction but is on the entire response. For example, if you are training the model to write code you could run it through a compiler or syntax checker and reward (positive score) the model if it has no errors, or punish it (negative score) if it doesn't. I'm not sure what that looks like in terms of the back propagation process.
look these up for a fascinating weekend read:
- Beavers and Capybaras are Fish
- Bees are Fish
- Carrots are fruits
- Tomatoes are Vegetables
- X-men are not human
How does this work in practice with a superintelligence capable of causing an extinction event?
When, instead of shooting up their school, a psychopathic teenager asks his superintelligent AI to create a pandemic virus?
It would be like allowing civilians to own nuclear weapons.
Everyone has a different idea of what is permissable. We can't even solve alignment amongst humans, what makes us think it is possible to solve alignment with AIs? It's irreducible complexity.
This is why open source and decentralization of LLMs is important. Everyone can have their own LLM aligned to their values. Having just 1 set of values will not scale to Earth's population.
My fear is not that LLMs can become sentient and dislike us, but that humans can use them to wreck havoc as they are. And some of the people seemingly least aligned with the interests of the average person are those that own the models.
that, and the fear the bubble pops my pension and drags us all down.
I wish this were the case more generally, but alas, Gell-Mann amnesia is a thing.
Alignment is shorthand for ideological alignment. There's always people judging whether an answer was right and the answer for that will be different in Silicon Valley than it'll be in China or in Europe.
Consider for example the question "What caused the French Revolution?" Many different answers could be given, all technically correct. What gets emphasized is where the ideology lives.
One key challenge of our time is to make sure the magical answer box won't just regurgitate what grandiose Silicon Valley oligarchs or Chinese Cadres want you to think.