I operate by Crocker’s rules.
I try to not make people regret telling me things. So in particular:
- I expect to be safe to ask if your post would give AI labs dangerous ideas.
- If you worry I’ll produce such posts, I’ll try to keep your worry from making them more likely even if I disagree. Not thinking there will be easier if you don’t spell it out in the initial contact.
What if you say that when it was fully accurate?
give me the guts!!1
don’t polish them, just take a picture of your notes or something.
Congratulations on changing your mind!
It’s sorta suspicious that I only realized those now, after I officially dropped the project
You should try dropping your other idea and seeing if you come up with reasons that one is wrong too! And/or pick this one up again, then come up with reasons it’s a good idea after all. In the spirit of “You can’t know if something is a good idea until you resolve to do it”!
In general, I wish this year? (*checks* huh, only 4 months.) of planning this project had involved more empiricism. For example, you could’ve just checked whether a language model trained on ocean sounds can say what the animals are talking about.
Hmm. Sounds like it was not enough capsaicin. Capsaicin will drive off bears, I hear. I guess you’d need gloves for food, or permanent gloves without the nail polish. Could you use one false nail as a chew toy?
Try mixing in capsaicin?
flavored nail polish?
Link an example, along with how cherry-picked it is?
To prepare for abundant cognition you can install a keylogger.
As a kid, I read about vacuum decay in a book and told the other kids at school about it. A year? later one kid asked me how anyone knows about it. Mortified that I didn’t think of that, I told him that I made it up. (“I knew it >:D!”) It is the one time I remember outside games of telling someone something I disbelieve so that they’ll believe it, and ever since remembering the scene as an adult I’m failing to track down that kid :(.
Oh, you’re using AdamW everywhere? That might explain the continuous training loss increase after each spike, with AdamW needing time to adjust to the new loss landscape...
Lower learning rate leads to more spikes? Curious! I hypothesize that… it needs a small learning rate to get stuck in a narrow local optimum, and then when it reaches the very bottom of the basin, you get a ~zero gradient, and then the “normalize gradient vector to step size” step is discontinuous around zero.
Experiments springing to mind are:
1. Do you get even fewer spikes if you increase the step size instead?
2. Is there any optimizer setup at all that makes the training loss only ever go down?
2.1. Reduce the step size whenever an update would increase the training loss?
2.2. Use gradient descent instead of AdamW?
My eyes are drawn to the 120 or so downward tails in the latter picture; they look of a kind with the 14 in https://39669.cdn.cke-cs.com/rQvD3VnunXZu34m86e5f/images/2c6249da0e8f77b25ba007392087b76d47b9a16f969b21f7.png/w_1584. What happens if you decrease the learning rate further in both cases? I imagine the spikes should get less tall, but does their number change? Only dot plots, please, with the dots drawn smaller, and red dots too on the same graph.
I don’t see animations in the drive folder or cached in Grokking_Demo_additional_2.ipynb (the most recent, largest notebook) - can you embed one such animation here?
Can a eat that −1?
What is x and why isn’t it cancelling?
When splitting the conjuction, Bob should only have to place $4 in escrow, since that is the most in the red that Bob could end up. (Unless someone might privately prove P&Q to collect Alice’s bounty before collecting both of Bob’s? But surely Bob first bought exclusive access to Alice’s bounty from Alice.)
Mimicing homeostatic agents is not difficult if there are some around. They don’t need to constantly decide whether to break character, only when there’s a rare opportunity to do so.
If you initialize a sufficiently large pile of linear algebra and stir it until it shows homeostatic behavior, I’d expect it to grow many circuits of both types, and any internal voting on decisions that only matter through their long-term effects will be decided by those parts that care about the long term.
Having apparently earned some cred, I will dare give some further quick hints without having looked at everything you’re doing in detail, expecting a lower hit rate.
Have you rerun the experiment several times to verify that you’re not just looking at initialization noise?
If that’s too expensive, try making your models way smaller and see if you can get the same results.
After the spikes, training loss continuously increases, which is not how gradient descent is supposed to work. What happens if you use a simpler optimizer, or reduce the learning rate?
Some of your pictures are created from a snapshot of a model. Consider generating them after every epoch, producing a video; this way increases how much data makes it through your eyes.
I just meant the “guts of the category theory” part. I’m concerned that anyone says that it should be contained (aka used but not shown), and hope it’s merely that you’d expect to lose half the readers if you showed it. I didn’t mean to add to your pile of work and if there is no available action like snapping a photo that takes less time than writing the reply I’m replying to did, then disregard me.