This is a special post for quick takes by Kabir Kumar. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Kabir Kumar's Shortform
4Kabir Kumar
4Kabir Kumar
3Kabir Kumar
3Kabir Kumar
5Seth Herd
3Kabir Kumar
5Seth Herd
1Kabir Kumar
1Kabir Kumar
1Kabir Kumar
1Kabir Kumar
1Kabir Kumar
1Kabir Kumar
1Kabir Kumar
5Daniel Tan
1Kabir Kumar
1Kabir Kumar
0Kabir Kumar
9Nathan Helm-Burger
3Kabir Kumar
3Nathan Helm-Burger
this might basically be me, but I'm not sure how exactly to change for the better. theorizing seems to take time and money which i don't have.
Thinking about judgement criteria for the coming ai safety evals hackathon (https://lu.ma/xjkxqcya )
These are the things that need to be judged:
1. Is the benchmark actually measuring alignment (the real, scale, if we dont get this fully right right we die, problem)
2. Is the way of Deceiving the benchmark to get high scores actually deception, or have they somehow done alignment?
Both of these things need:
- a strong deep learning & ml background (ideally, muliple influential papers where they're one of the main authors/co-authors, or doing ai research at a significant lab, or they have, in the last 4 years)
- a good understanding of what the real alignment problem actually means - can judge this by looking at their papers, activity on lesswrong, alignmentforum, blog, etc
- a good understanding of evals/benchmarks (1 great or two pretty good papers/repos/works on this, ideally for alignment)
Do these seem loose? Strict? Off base?
ok, options.
- Review of 108 ai alignment plans
- write-up of Beyond Distribution - planned benchmark for alignment evals beyond a models distribution, send to the quant who just joined the team who wants to make it
- get familiar with the TPUs I just got access to
- run hhh and it's variants, testing the idea behind Beyond Distribution, maybe make a guide on itr
- continue improving site design
- fill out the form i said i was going to fill out and send today
- make progress on cross coders - would prob need to get familiar with those tpus
- writeup of ai-plans, the goal, the team, what we're doing, what we've done, etc
- writeup of the karma/voting system
- the video on how to do backprop by hand
- tutorial on how to train an sae
think Beyond Distribution writeup. he's waiting and i feel bad.
btw, thoughts on this for 'the alignment problem'?
"A robust, generalizable, scalable, method to make an AI model which will do set [A] of things as much as it can and not do set [B] of things as much as it can, where you can freely change [A] and [B]"
Freely changing an AGIs goals is corrigibility, which is a huge advantage if you can get it. See Max Harms' corrigibility sequence and my "instruction-following AGI is easier...."
The question is how a reliably get such a thing. Goalcrafting is one part of the problem, and I agree that those are good goals; the other and larger part is technical alignment, getting those desired goals to really work that way in the particular first AGI we get.
Yup, those are hard. Was just thinking of a definition for the alignment problem, since I've not really seen any good ones.
I'd say you're addressing the question of goalcrafting or selecting alignment targets.
I think you've got the right answer for technical alignment goals; but the question remains of what human would control that AGI. See my "if we solve alignment, do we all die anyway" for the problems with that scenario.
Spoiler alert; we do all die anyway if really selfish people get control of AGIs. And selfish people tend to work harder at getting power.
But I do think your goal defintion is a good alignment target for the technical work. I don't think there's a better one. I do prefer instruction following or corriginlilty by the definitions in the posts I linked above because they're less rigid, but they're both very similar to your definition.
I pretty much agree. I prefer rigid definitions because they're less ambiguous to test and more robust to deception. And this field has a lot of deception.
in general, when it comes to things which are the 'hard part of alignment', is the crux
```
a flawless method of ensuring the AI system is pointed at and will always continue to be pointed at good things
```
?
the key part being flawless - and that seeming to need a mathematical proof?
Trying to put together a better explainer for the hard part of alignment, while not having a good math background https://docs.google.com/document/d/1ePSNT1XR2qOpq8POSADKXtqxguK9hSx_uACR8l0tDGE/edit?usp=sharing
Please give feedback!
Thoughts on this?
### Limitations of HHH and other Static Dataset benchmarks
A Static Dataset is a dataset which will not grow or change - it will remain the same. Static dataset type benchmarks are inherently limited in what information they will tell us about a model. This is especially the case when we care about AI Alignment and want to measure how 'aligned' the AI is.
### Purpose of AI Alignment Benchmarks
When measuring AI Alignment, our aim is to find out exactly how close the model is to being the ultimate 'aligned' model that we're seeking - a model whose preferences are compatible with ours, in a way that will empower humanity, not harm or disempower it.
### Difficulties of Designing AI Alignment Benchmarks
What preferences those are, could be a significant part of the alignment problem. This means that we will need to frequently make sure we know what preferences we're trying to measure for and re-determine if these are the correct ones to be aiming for.
### Key Properties of Aligned Models
These preferences must be both robustly and faithfully held by the model:
Robustness:
- They will be preserved over unlimited iterations of the model, without deterioration or deprioritization.
- They will be robust to external attacks, manipulations, damage, etc of the model.
Faithfulness:
- The model 'believes in', 'values' or 'holds to be true and important' the preferences that we care about .
- It doesn't just store the preferences as information of equal priority to any other piece of information, e.g. how many cats are in Paris - but it holds them as its own, actual preferences.
Comment on the Google Doc here: https://docs.google.com/document/d/1PHUqFN9E62_mF2J5KjcfBK7-GwKT97iu2Cuc7B4Or2w/edit?usp=sharing
This is for the AI Alignment Evals Hackathon: https://lu.ma/xjkxqcya by AI-Plans
I'm looking for feedback on the hackathon page
mind telling me what you think?
https://docs.google.com/document/d/1Wf9vju3TIEaqQwXzmPY--R0z41SMcRjAFyn9iq9r-ag/edit?usp=sharing
I'd like some feedback on my theory of impact for my currently chosen research path
**End goal**: Reduce x-risk from AI and risk of human disempowerment.
for x-risk:
- solving AI alignment - very important,
- knowing exactly how well we're doing in alignment, exactly how close we are to solving it, how much is left, etc seems important.
- how well different methods work,
- which companies are making progress in this, which aren't, which are acting like they're making progress vs actually making progress, etc
- put all on a graph, see who's actually making the line go up
- Also, a way that others can use to measure how good their alignment method/idea is, easily
so there's actually a target and a progress bar for alignment - seems like it'd make alignment research a lot easier and improve the funding space - and the space as a whole. Improving the quality and quantity of research.
- Currently, it's mostly a mixture of vibe checks, occasional benchmarks that test a few models, jailbreaks, etc
- all almost exclusively on the end models as a whole - which have many, many differences that could be contributing to the differences in the different 'alignment measurements'
by having a method that keeps things controlled as much as possible and just purely measures the different post training methods, this seems like a much better way to know how we're doing in alignment
and how to prioritize research, funding, governence, etc
On Goodharting the Line - will also make it modular, so that people can add their own benchmarks, and highlight people who redteam different alignment benchmarks.
What is the proposed research path and its theory of impact? It’s not clear from reading your note / generally seems too abstract to really offer any feedback
I think this is a really good opportunity to work on a topic you might not normally work on, with people you might not normally work with, and have a big impact:
https://lu.ma/sjd7r89v
I'm running the event because I think this is something really valuable and underdone.
I'm finally reading The Sequences and it screams midwittery to me, I'm sorry.
Compare this:
to Jaynes:
Jaynes is better organized, more respectful to the reader, more respectful to the work he's building on and more useful
The Sequences highly praise Jaynes and recommend reading his work directly.
The Sequences aren't trying to be a replacement, they're trying to be a pop sci intro to the style of thinking. An easier on-ramp. If Jaynes already seems exciting and comprehensible to you, read that instead of the Sequences on probability.
Curated and popular this week