Ngl I did not fully understand this, but to be clear I don't think understanding alignment through the lense of agency is "excessively abstract." In fact I think I'd agree with the implicit default view that it's largely the single most productive lense to look through. My objection to the status quo is that it seems like the scale/ontology/lense/whatever I was describing is getting 0% of the research attention whereas perhaps it should be getting 10 or 20%.  

Not sure this analogy works, but if NIH was spending $10B on cancer research, I would (prima facie, as a layperson) want >$0 but probably <$2B spent on looking at cancer as an atomic-scale phenomenon, and maybe some amount at an even lower-scale scale

Note: I'm probably well below median commenter in terms of technical CS/ML understanding.  Anyway...

I feel like a missing chunk of research could be described as “seeing DL systems as ‘normal,’ physical things and processes that involve electrons running around inside little bits of (very complex) metal pieces” instead of mega-abstracted “agents.”

The main reason this might be fruitful is that, at least intuitively and to my understanding, failures like “the AI stops just playing chess really well and starts taking over the world to learn how to play chess even better” involve a qualitative change beyond just “the quadrillion parameters adjust a bit to minimize loss even more” that eventually cashes out in some very different way that literal bits of metal and electrons are arranged. 

And plausibly abstracting away from the chips and electrons means ignoring the mechanism that permits this change. Of course, this probably only makes sense if something resembling deep learning scales to AGI, but it seems that some very smart people think that it may! 

Banneker Key! Yeah I was in a very similar position, but basically made the opposite choice (largely because financial costs not internalized)

Yeah that's gotta be it, nice catch! 

One answer to the question for me:

While writing, something close to "how does this 'sound' in my head naturally, when read, in an aesthetic sense?"

I've thought for a while that "writing quality" largely boils down to whether the writer has an intuitively salient and accurate intuition about how the words they're writing come across when read. 

Ah late to the party! This was a top-level post aptly titled "Half-baked alignment idea: training to generalize" that didn't get a ton of attention. 

Thanks to Peter Barnett and Justis Mills for feedback on a draft of this post. It was inspired by Eliezer's Lethalities post and Zvi's response.

Central idea: can we train AI to generalize out of distribution? 

I'm thinking, for example, of an algorithm like the following:

1. Train a GPT-like ML system to predict the next word given a string of text only using, say, grade school-level writing (this being one instance of the object level)
   1. Assign the system a meta-level award based on how well it performs (without any additional training) at generalizing; in this case, that is, predicting the next word from more advanced, complex writing (perhaps using many independent tests of this task without updating/learning between each test, and allowing parameters to update only after the meta-level aggregate score is provided)
      • Note: the easy→hard generalization is not a necessary feature. Generalization could be from fiction→nonfiction writing or internet→native print text, for instance.
   2. After all these independent samples are taken, provide the AI its aggregate or average score as feedback
2. (Maybe?) repeat all of step I on a whole new set of training and testing texts (e.g., using text from a different natural language like Mandarin)
   1. Repeat this step an arbitrary number of times
      • For example, using French text, then Korean, then Arabic, etc. 
3. Each time a “how well did you generalize” score is provided (which is given once per natural language in this example), the system should improve at the general task of generalizing from simple human writing to more complex human writing, (hopefully) to the point of being able to perform well at generalizing from simple Hindi (or whatever) text to advanced Hindi prediction even if it had never seen advanced Hindi text before.  
4. ^Steps 1-3 constitute the second meta-level of training an AI to generalize, but we can easily treat this process as a single training instance (e.g., rating how well the AI generalizes to Hindi advanced text after having been trained on doing this in 30 other languages) and iterate over and over again. I think this would look like:
   1. Running the analogs of steps 1-4 on generalizing from 
      • (a) simple text to advanced text in many languages
      • (b) easy opponents to hard ones across many games, 
      • (c) photo generation of common or general objects ("car") to rare/complex/specific ones ("interior of a 2006 Honda Accord VP"), across many classes of object
   2. And (hopefully) the system would eventually be able to generalize from simple Python code training data to advanced coding tasks even though it had never seen any coding at all before this. 

And, of course, we can keep on adding piling layers on. 

A few notes

• I think the following is one way of phrasing what I hope might happen with method: we are using RL to teach an ML system how to do ML in such a way that it sacrifices some in-distribution predictive power for the ability to use its “knowledge” more generally without doing anything that seems dumb to us. 
• Of course, there are intrinsic limits to any system’s ability to generalize. The system in question can only generalize using knowledge X if X exists as information in the object-level training provided to it. 
  • This limits what we should expect of the system.
    • For example, I am almost certain that even an arbitrarily smart system will not be able to generate coherent Mandarin text from English training data, because the meaning of Mandarin characters doesn’t exist as “latent knowledge” in even a perfect understanding of English. 

Anyone here know Python?

My hands-on experience with ML extends to linear regression in R and not an inch more, so I'm probably not the best person to test this theory out. I've heard some LWers know a bit of Python, though.

If that's you, I'd be fascinated and thankful to see if you can implement this idea using whatever data and structure you think would work best, and would be happy to collaborate in whatever capacity I can. 

 

Appendix: a few brief comments (from someone with much more domain knowledge than me) and responses (from me):

Comment

Is this just the same as training it on this more complex task (but only doing one big update at the end, rather than doing lots of small updates)?

Response (which may help to clarify why I believe the idea might work)

I don't think so, because the parameters don't change/update/improve between each of those independent tests. Like GPT-3 in some sense has a "memory" of reading Romeo and Juliet, but that's only because its parameters updated as a result of seeing the text.

But also I think my conception depends on the system having "layers" of parameters corresponding to each layer of training. 

So train on simple English-->only "Simple English word generation" parameters are allowed to change...but then you tell it how well it did at generalizing out of distribution, and now only its "meta level 1 generalization" parameters are allowed to change.

Then you do the whole thing again but with German text, and its "Meta level 1 generalization" parameters are allowed to change again using SGD or whatever. If this works, it will be the reason why it can do well at advanced Hindi text without ever having read advanced Hindi.

Treat this whole process as the object level, and then it updates/improves "meta level 2 generalization" parameters.

Comment: 

This looks vaguely like curriculum learning, which apparently doesn't really work in LLMs https://arxiv.org/abs/2108.02170, I think a similar experiment would be like train on simple+advanced text for English, French, Mandarin etc, but only simple Hindi, and then see if it can do complex Hindi. 

Response 

I think that's a pretty different thing because there are no meta level parameters. Seems like fundamentally just a flavor of normal RL

Or do pretraining with English, French, Mandarin, and Hindi, but only do fine tuning with English, French, Mandarin, and see if it can then do the tasks it was fine tuned for in Hindi. 

My prediction: it learns to generalize a bit (the scores on the novel Hindi tasks are higher than if there was no fine tuning with the other languages) but worse than the other languages generalize. As the models are scaled up, this 'generalization gap' gets smaller.

Seems like this might depend on the relative scaling of different meta level parameters (which I described above)? 

Like for example whenever you scale the # of object level params by a factor of 2, you have to scale the number of nth meta level parameters by 2^(n+1).

Thank you, Solenoid! The SSC podcast is the only reason I to consume all of posts like Biological Anchors: A Trick That Might Or Might Not Work

Thanks. It's similar in one sense, but (if I'm reading the paper right) a key difference is that in the MAML examples, the ordering of the meta-level and object level training is such that you still wind up optimizing hard for a particular goal. The idea here is that the two types of training function in opposition, as a control system of sorts, such that the meta-level training should make the model perform worse at the narrow type of task it was trained on. 

That said, for sure, the types of distribution shift thing is an issue. It seems like this meta-level bias might be less bad than at the object level, but I have no idea. 

Training to generalize (and training to train to generalize, etc.)

Inspired by Eliezer's Lethalities post and Zvi's response:

Has there been any research or writing on whether we can train AI to generalize out of distribution? 

I'm thinking, for example:

1. Train a GPT-like ML system to predict the next word given a string of text only using, say, grade school-level writing (this is one instance of the object level
   1. Assign the system a meta-level award based on how well it performs (without any additional training) at generating the next word from more advanced, complex writing (likely using many independent tests of this task) 
   2. After all these independent samples are taken, provide the its aggregate or average score as feedback
2. (Maybe?) repeat steps I and I.I on a whole new set of training and testing texts (e.g., using text from a different natural language like Mandarin), and do this step and arbitrary number of times
   1. Repeat this step using French text, then Korean, then Arabic, etc. 
3.  After each of the above steps, the system should (I assume) improve at the general task of generalizing from simple human writing to more complex human writing, (hopefully) to the point of being able to perform well at generalizing from simple Hindi (or whatever) text to advanced Hindi prediction even if it had never seen advanced Hindi text before.  
4. ^Steps 1-3 constitute the second meta-level of training an AI to generalize, but we can easily treat this process as a single training instance (e.g., rating how well the AI generalizes to Hindi advanced text after having been trained on doing this in 30 other languages) and iterate over and over again. I think this would look like:
   1. Running the analogs of steps 1-4 on generalizing from 
      • (a) simple text to advanced text in many languages
      • (b) easy opponents to hard ones across many games, 
      • (c) photo generation of common or general objects ("car") to rare/complex/specific ones ("interior of a 2006 Honda Accord VP"), across many classes of object
   2. And (hopefully) the system would eventually be able to generalize from simple python code training data to advanced coding tasks even though it had never seen any coding at all before this. 

And, of course, we can keep on adding piling layers on.

A few minutes of hasty Googling didn't turn up anything on this, but it seems pretty unlikely to be an original idea. But who knows! I wanted to get the idea written down and online before I had time to forget about it. 

On the off chance it hasn't been beaten thought about to death yet by people smarter than myself, I would consider together longer,  less hastily written post on the idea 

MichaelStJules is right about what I meant. While it's true that preferring not to experience something doesn't necessarily imply that the thing is net-negative, it seems to me very strong evidence in that direction.