Keenan Pepper

Karma: 51

Keenan Pepper 17 Dec 2024 3:29 UTC
1 point
0
on: Attention Output SAEs Improve Circuit Analysis
Who ended up getting the bounty for this?

Keenan Pepper 4 Nov 2024 22:31 UTC
2 points
0
on: What TMS is like
As far as I’ve been told, left-TMS is for depression, right-TMS is for anxiety. Why that’s the case, I have no idea.)
As I was reading this I intuited there would be something to predict here so I successfully stopped reading before the “left-TMS is for depression, right-TMS is for anxiety” part and scrolled it out of view so I could do the prediction myself based on what I understand to be the roles of the right and left hemispheres.

As I understand it, the left hemisphere of the brain is sort of focused “forwards”, on whatever tool you’re using or prey you’re hunting, and in particular on words you’re speaking/hearing and plans you’re making to pursue some goal. In contrast the right hemisphere of the brain is focused “outwards” on the environment as a whole and sort of on guard for any threats or interesting irregularities that ought to pull your attention away.

Therefore I predicted that left-TMS would be for kind of general depression stuff about all your plans seeming like bad ideas or whatever, and right-TMS would be for worrying about a bunch of stuff that’s in your environment and being distracted, which sounds more like either an ADHD kind of thing or anxiety!

So you’ll have to take my word for it, but I got it right.

Keenan Pepper 29 Oct 2024 21:13 UTC
1 point
0
in reply to: Vanessa Kosoy’s comment on: The hostile telepaths problem
...never making excuses to myself such as “I wanted to do A, but I didn’t have the willpower so I did B instead”, but rather owning the fact I wanted to do B and thinking how to integrate this...
AKA integrating the ego-dystonic into the homunculus

Keenan Pepper 13 Oct 2024 1:45 UTC
4 points
0
on: [Intuitive self-models] 4. Trance

I think what’s happening in this last one is that there’s a salient intuitive model where your body is part of “the crowd”, and “the crowd” is the force controlling your actions.

This strongly reminds me of this excellent essay: https://meltingasphalt.com/music-in-human-evolution/

Keenan Pepper 1 Aug 2024 20:10 UTC
2 points
0
on: Self-Other Overlap: A Neglected Approach to AI Alignment
Can we expect to see code for this on https://github.com/agencyenterprise sometime soon? I’m excited to fiddle with this.

Keenan Pepper 25 Jun 2024 22:34 UTC
3 points
2
on: The Three Contingencies Of The Optimality Function
How is this related to embedded agency? I have a subscription to that topic so I get notified of new posts, but I don’t see how this is related at all.

Keenan Pepper 13 Jun 2024 3:48 UTC
6 points
−2
on: My AI Model Delta Compared To Christiano

In HCH, the human user does a little work then delegates subquestions/subproblems to a few AIs, which in turn do a little work then delegate their subquestions/subproblems to a few AIs, and so on until the leaf-nodes of the tree receive tiny subquestions/subproblems which they can immediately solve.

This does not agree with my understanding of what HCH is at all. HCH is a definition of an abstract process for thought experiments, much like AIXI is. It’s defined as the fixed point of some iterative process of delegation expanding out into a tree. It’s also not something you could actually implement, but it’s a platonic form like “circle” or “integral”.

This has nothing to do with the way an HCH-like process would be implemented. You could easily have something that’s designed to mimic HCH but it’s implemented as a single monolithic AI system.

Keenan Pepper 11 Jun 2024 22:53 UTC
8 points
0
in reply to: johnswentworth’s comment on: My AI Model Delta Compared To Yudkowsky
Okay so where do most of your hopes route through then?

Keenan Pepper 4 Jun 2024 11:52 UTC
0 points
0
in reply to: Prometheus’s comment on: Prometheus’s Shortform
SAME

Keenan Pepper 4 Jun 2024 11:48 UTC
4 points
0
on: Preventing Language Models from hiding their reasoning
I’m just here to note that the “canonical example” thing mentioned here is very similar to the “nothing-up-my-sleeve numbers” used in the definition of some cryptographic protocols.
What links here?
- Fabien Roger's comment on When fine-tuning fails to elicit GPT-3.5′s chess abilities by Theodore Chapman (15 Jun 2024 6:27 UTC; 35 points)

Keenan Pepper 13 May 2024 10:24 UTC
1 point
0
in reply to: tropea@gwu.edu’s comment on: Transformers Represent Belief State Geometry in their Residual Stream
I think you may have meant this as a top-level comment rather than a reply to my comment?

Keenan Pepper 23 Apr 2024 0:21 UTC
1 point
0
in reply to: Keenan Pepper’s comment on: Transformers Represent Belief State Geometry in their Residual Stream
Actually I would still really appreciate the training hyperparameters like batch size, learning rate schedule...

Keenan Pepper 20 Apr 2024 23:51 UTC
1 point
0
in reply to: Keenan Pepper’s comment on: Transformers Represent Belief State Geometry in their Residual Stream
Ah, never mind, I believe I found the relevant hyperparameters here: https://github.com/adamimos/epsilon-transformers/blob/main/examples/msp_analysis.ipynb
In particular, the stuff I needed was that it has only a single attention head per layer, and 4 layers.

Keenan Pepper 20 Apr 2024 23:50 UTC
4 points
0
in reply to: Exa Watson’s comment on: Transformers Represent Belief State Geometry in their Residual Stream
No, the actual hidden Markov process used to generate the awesome triangle fractal image is not the {0,1,random} model but a different one, which is called “Mess3” and has a symmetry between the 3 hidden states.
Also, they’re not claiming the transformer learns merely the hidden states of the HMM, but a more complicated thing called the “mixed state presentation”, which is not the states that the HMM can be in but the (usually much larger number of) belief states which an ideal prediction process trying to “sync” to it might go thru.

Keenan Pepper 20 Apr 2024 21:29 UTC
1 point
0
on: Transformers Represent Belief State Geometry in their Residual Stream
Can you share the hyperparameters used to make this figure?

Keenan Pepper 14 Mar 2024 22:53 UTC
1 point
0
on: Fixed point or oscillate or noise
Okay my computer right here has 10^13 bits of storage and without too much trouble I could get it to use all that memory as a counter and just count to the highest value possible, which would be 2^(10^13) or in other words much much longer than the age of the universe even at a fast clock speed.

Now technically yes, after it got to that 2^(10^13) value it would have to either halt or start over from 0 or something… but that seems not so practically relevant to me because it’s such a huge integer value.

Keenan Pepper 20 Dec 2023 2:02 UTC
1 point
0
on: Acausal normalcy
I haven’t dived into this yet, but am I right in guessing that the gist is exactly like a way more fleshed-out and intricate version of Hofstadter’s “superrationality”?

Keenan Pepper 28 Sep 2023 1:11 UTC
1 point
on: Elementary Infra-Bayesianism
I’m trying to read this post now but it looks like a bunch of images (of math) are missing. Does that match what others see?

Keenan Pepper 21 Sep 2023 0:42 UTC
3 points
2
in reply to: LawrenceC’s comment on: Interpretability Externalities Case Study—Hungry Hungry Hippos
The Bitter Lesson applies to almost all attempts to build additional structure into neural networks, it turns out.
Out of curiosity, what are the other exceptions to this besides the obvious one of attention?

Keenan Pepper 21 Sep 2023 0:11 UTC
−5 points
−2
in reply to: mruwnik’s comment on: Where might I direct promising-to-me researchers to apply for alignment jobs/grants?
Upvoted because this mentions Nonlinear Network.