LessWrong 2.0 Reader

Fascinating. But are these diagrams really showing HMMs? I thought each state in an HMM had a set of transition probabilities and another set of emission probabilities, which at each step are sampled independently. In these diagrams, the two processes are coupled. If "Even Ys" were a conventional HMM, $S_E$ would sometimes emit X and transition to $S_O$, which would result in some even and some odd runs of Y. Are these a special variant of HMM, or some other type of state machine? And would these results apply to conventional HMMs with separate transition and emission probabilities?

Yes, I train them one at a time, constraining each new vector to be orthogonal to the older ones (this was not clear in the post, so thanks for asking!). 

I haven't experimented with this, but you could also imagine using only "soft" orthogonality constraints (e.g., penalizing pairwise cosine similarities between vectors). 

Thanks for your comment! Yes, I’d say that roughly sums things up.

As I alluded to in the introduction, I think the advantage of having a latent perturbation (as opposed to just eliciting an interesting sequence of tokens) is due to i) generalization and ii) applications to mechanistic anomaly detection. Generalization is important in situations where the backdoored behavior is subtle. In this case, even if you had a non-latent way of eliciting the behavior (e.g. through adversarial inputs) you don't know whether you've simply elicited a weird sequence of tokens or if there’s something systematic about it. But if you have a latent perturbation then you can run the perturbed model over many inputs to confirm that this is a generalizable behavior (e.g., systematic but subtle coding errors meant to sabotage you) rather than a weird sequence of tokens.

Moreover, this might chain well with mechanistic anomaly detection. Quoting the post [AF(p) · GW(p)]: “before deployment, train a large set of unsupervised model perturbations, then during deployment, if the model's activations look similar to one of the previously learned perturbations, auditors can go back and evaluate the behaviors encoded by that perturbation more thoroughly to see if there is anything problematic with these behaviors.”

I agree that it would be helpful to spell out/test the advantages of latent perturbations more concretely in followup work (for example, trying to show that mechanistic anomaly detection actually works using the learned steering vectors).

Just got to this today. I've come up with a candidate solution just to try to survive, but haven't had a chance yet to check & confirm that it'll work, or to try to get clever and reduce the number of soldiers I'm using.

I will probably work on this some more tomorrow.

corp, org

I maybe want to clarify: there will still be presentations at LessOnline, we're just trying to design the event such that they're clearly more of a secondary thing.

If we know a meteor is about to hit earth, having only D days to prepare what is the rational thing to person P to do?

There are values of P and D, such that throwing a big party party before the earth explodes might be the best choice

Virtual watercoolers

As I mentioned in some recent Shortform posts, I recently listened to the Bayesian Conspiracy podcast's episode on the LessOnline festival and it got me thinking.

One thing I think is cool is that Ben Pace was saying how the valuable thing about these festivals isn't the presentations, it's the time spent mingling in between the presentations, and so they decided with LessOnline to just ditch the presentations and make it all about mingling. Which got me thinking about mingling.

It seems plausible to me that such mingling can and should happen more online. And I wonder whether an important thing about mingling in the physical world is that, how do I say this, you're just in the same physical space, next to each other, with nothing else you're supposed to be doing, and in fact what you're supposed to be doing is talking to one another.

Well, I guess you're not supposed to be talking to one another. It's also cool if you just want to hang out and sip on a drink or something. It's similar to the office water cooler: it's cool if you're just hanging out drinking some water, but it's also normal to chit chat with your coworkers.

I wonder whether it'd be good to design a virtual watercooler. A digital place that mimicks aspects of the situations I've been describing (festivals, office watercoolers).

1. By being available in the virtual watercooler it's implied that you're pretty available to chit chat with, but it's also cool if you're just hanging out doing something low key like sipping a drink.
2. You shouldn't be doing something more substantial though.
3. The virtual watercooler should be organized around a certain theme. It should attract a certain group of people and filter out people who don't fit in. Just like festivals and office water coolers.

In particular, this feels to me like something that might be worth exploring for LessWrong.

Note: I know that there are various Slack and Discord groups but they don't meet conditions (1) or (2).

I'd take an agnostic view on whether LLMs are doing search internally. Crucially, though, I think the relevant output to be searching over is distributions of tokens, rather than the actual token that gets chosen. Search is not required to generate a single distribution over next tokens. 

I agree that external search via scaffolding can also be done, and would be much easier to identify, but without understanding the internals it's hard to know how powerful the search process will be.

More dakka with festivals

In the rationality community people are currently excited about the LessOnline festival. Furthermore, my impression is that similar festivals are generally quite successful: people enjoy them, have stimulating discussions, form new relationships, are exposed to new and interesting ideas, express that they got a lot out of it, etc.

So then, this feels to me like a situation where More Dakka [? · GW] applies. Organize more festivals!

How? Who? I dunno, but these seem like questions worth discussing.

Some initial thoughts:

1. Assurance contracts [? · GW] seem like quite the promising tool.
2. You probably don't need a hero license [LW · GW] to go out and organize a festival.
3. Trying to organize a festival probably isn't risky. It doesn't seem like it'd involve too much time or money.