LessWrong 2.0 Reader

Enjoyed this post! Quick question about obtaining the steering vectors:

Do you train them one at a time, possibly adding an additional orthogonality constraint between each train?

Under the current version of the interactive model, its median prediction is just two decades earlier than that from Cotra’s forecast

Just?

Question about the "rules of the game" you present.  Are you allowed to simply look at layer 0 transcoder features for the final 10 tokens - you could probably roughly estimate the input string from these features' top activators.  From you case study, it seems that you effectively look at layer 0 transcoder features for a few of the final tokens through a backwards search, but wonder if you can skip the search and simply look at transcoder features.  Thank you.

By "gag order" do you mean just as a matter of private agreement, or something heavier-handed, with e.g. potential criminal consequences?

I have trouble understanding the absolute silence we seem to be having. There seem to be very few leaks, and all of them are very mild-mannered and are failing to build any consensus narrative that challenges OA's press in the public sphere.

Are people not able to share info over Signal or otherwise tolerate some risk here? It doesn't add up to me if the risk is just some chance of OA trying to then sue you to bankruptcy, especially since I think a lot of us would offer support in that case, and the media wouldn't paint OA in a good light for it.

I am confused. (And I grateful to William for at least saying this much, given the climate!)

Maybe GPT-5 will be extremely good at interpretability, such that it can recursively self improve by rewriting its own weights.

I am by no means an expert on machine learning, but this sentence reads weird to me. 

I mean, it seems possible that a part of a NN develops some self-reinforcing feature which uses the gradient descent (or whatever is used in training) to go into a particular direction and take over the NN, like a human adrift on a raft in the ocean might decide to build a sail to make the raft go into a particular direction. 

Or is that sentence meant to indicate that an instance running after training might figure out how to hack the computer running it so it can actually change it's own weights?

Personally, I think that if GPT-5 is the point of no return, it is more likely that it is because it would be smart enough to actually help advance AI after it is trained. While improving semiconductors seems hard and would require a lot of work in the real world done with human cooperation, finding better NN architectures and training algorithms seems like something well in the realm of the possible, if not exactly plausible.

So if I had to guess how GPT-5 might doom humanity, I would say that in a few million instance-hours it figures out how to train LLMs of its own power for 1/100th of the cost, and this information becomes public. 

The budgets of institutions which might train NN probably follows some power law, so if training cutting edge LLMs becomes a hundred times cheaper, the number of institutions which could build cutting edge LLMs becomes many orders of magnitude higher -- unless the big players go full steam ahead towards a paperclip maximizer, of course. This likely mean that voluntary coordination (if that was ever on the table) becomes impossible. And setting up a worldwide authoritarian system to impose limits would also be both distasteful and difficult. 

I have been contemplating Connor Leahy's Cyborgism [LW · GW] and what it would mean for us to improve human workflows enough that aligning AGI looks less like:

Sisyphus attempting to roll a 20 tonne version of The One Ring To Rule Them All into the caldera of Mordor while blindfolded and occasionally having to bypass vertical slopes made out of impossibility proofs that have been discussed by only 3 total mathematicians ever in the history of our species - all before Sauron destroys the world after waking up from a restless nap of an unknown length.

I think this is what you meant by "make the ultimate <programming language/text editor/window manager/file system/virtual collaborative environment/interface to GPT/...>"

Intuitively, the level I'm picturing is:

A suite of tools that can be booted up from a single icon on the home screen of a computer which then allows anyone who has decent taste in software to create essentially any program they can imagine up to a level of polish that people can't poke holes in even if you give a million reviewers 10 years of free time.

Can something at this level be accomplished?

Well, what does coding look like currently? It seems to look like a bunch of people with dark circles under their eyes reading long strings of characters in something basically the equivalent to an advanced text editor, with a bunch of additional little windows of libraries and graphics and tools.

This is not as domain where human intelligence performs with as much ease as in other domains like spearfishing or bushcraft.

If you want to build Cyborgs, I am pretty sure where you start is by focusing on building software that isn't god-machines, throwing out the old book of tacit knowledge, and starting over with something that makes each step as intuitive as possible. You probably also focus way more on quality over quantity/speed.

So, plaintext instructions on what kind of software you want to build, or a code repository and a plaintext list of modifications? Like, start with an elevator pitch, see the raw/AI generated material, critique in a step-by-step organized fashion where debugging/feature analysis checklists are generated and scored on whether they included everything you would have thought of/stuff that is valid which you didn't think of.

I think the point in this post is valid, though a bit more in the realm of angsty shower thought rather than a ready-to-import heuristic for analysing the gap between competence-in-craft and unleased-power-levels.

There is a bit of a bootstrapping problem with Cyborgism. I don't think you get the One Programming Suite To Rule Them All by plugging in a bunch of different LLMs fine tuned to do one part of the process really well - then packaging the whole thing up and running it on 6 gaming GPUs. That is the level of super-software that seems in reach, and it just seems doomed to be full of really hard-to-perceive holes like a super high-dimensional block of swiss cheese.

Does that even get better if we squeeze the weights of LLMs to get LLeMon juice:

Python script that does useful parts of the cool smart stuff LLMs do without all the black box fuzzy embedding-spaces/vectors + filter the juice for pulp/seeds (flaws in the specific decoded algorithm that could cause errors via accidental or deliberate adversarial example) + sweeten it (make the results make sense/be understandable to humans)

... then plug a bunch of that LLeMonade into a programming suite such that the whole thing works with decently competent human programmer(s) to reliably make stuff that actually just works & alert people ahead of time of the exhaustive set of actual issues/edge cases/gaps in capability of a program?

This problem does seem difficult - and probably the whole endeavor just actually won't work well enough IRL, but it seems worth trying?

Like, what does it look like to throw everything and the kitchen sink at Alignment? It probably looks at least a little like the Apollo program, and if you're doing NASA stuff properly, then you end up making some revolutionary products for everyone else.

I think those products - the random externalities of a healthy Alignment field - look more like tools that work simply and reliably, rather than the giant messy LLM hairballs AI labs keep coughing up and dumping on people.

Maybe all of this helps flesh out and make more useful the flinchy heuristic of "consumer software works terribly -> ... -> AI destroys the future."

Alignment as a field goes out ahead of the giant rolling gooey hairball of spaghetti-code Doom - untangles it and weaves it into a beautiful textile - or we are all dead.

Agreed, but it's not just software.  It's every complex system, anything which requires detailed coordination of more than a few dozen humans and has efficiency pressure put upon it.  Software is the clearest example, because there's so much of it and it feels like it should be easy.

Primarily people come to this on the discord, so I just have this on lw for visibility

Something I think humanity is going to have to grapple with soon is the ethics of self-modification / self-improvement, and the perils of value-shift due to rapid internal and external changes. How do we stay true to ourselves while changing fundamental aspects of what it means to be human?

Huh, that's indeed somewhat surprising if the SAE features are capturing the things that matter to CLIP (in that they reduce loss) and only those things, as opposed to "salient directions of variation in the data". I'm curious exactly what "failing to work" means -- here I think the negative result (and the exact details of said result) are argubaly more interesting than a positive result would be.