LessWrong 2.0 Reader

Launched a few days ago, the plan is:

• Kids tickets are $50
• There's daycare purchasable on-site from 10am to 7pm, for like $10/hour if you book ahead of time or $30/hour if you use it on-the-day
• If you want connection to a nanny for outside of those hours we have a service that can help with that at $45/hour.

Happy to get feedback on this, still figuring out what exactly helps parents and how to set it up right.

Sorry, what? I thought the fear was that we don't know how to make helpful AI at all. (And that people who think they're being helped by seductively helpful-sounding LLM assistants are being misled by surface appearances; the shoggoth underneath has its own desires that we won't like when it's powerful enough to persue them autonomously.) In contrast, this almost makes it sound like you think it is plausible to align AI to its user's intent, but that this would be bad if the users aren't one of "us"—you know, the good alignment researchers who want to use AI to take over the universe, totally unlike those evil capabilities researchers who want to use AI to produce economically valuable goods and services.

My steelman of this (though to be clear I think your comment makes good points): 

There is a large difference between a system being helpful and a system being aligned. Ultimately AI existential risk is a coordination problem where I expect catastrophic consequences because a bunch of people want to build AGI without making it safe. Therefore making technologies that in a naive and short-term sense just help AGI developers build whatever they want to build will have bad consequences. If I trusted everyone to use their intelligence just for good things, we wouldn't have anthropogenic existential risk on our hands.

Some of those technologies might end up useful for also getting the AI to be more properly aligned, or maybe to help with work that reduces the risk of AI catastrophe some other way, though my current sense is that kind of work is pretty different and doesn't benefit remotely as much from generically locally-helpful AI.

In-general I feel pretty sad about conflating "alignment" with "short-term intent alignment". I think the two problems are related but have really important crucial differences, I don't think the latter generalizes that well to the former (for all the usual sycophancy/treacherous-turn reasons), and indeed progress on the latter IMO mostly makes the world marginally worse because the thing it is most likely to be used for is developing existentially dangerous AI systems faster.

Edit: Another really important dimension to model here is also not just the effect of that kind of research on what individual researchers will do, but what effect this kind of research will have on what the market wants to invest in. My standard story of doom is centrally rooted in there being very strong short-term economic incentives to build more capable AGI for any individual, enabling people to make billions to trillions of dollars, while the downside risk is a distributed negative externality that is not at all priced into the costs of AI development. Developing applications of AI that make a lot of money without accounting for the negative extinction externalities therefore can be really quite bad for the world. 

Fixed. Thanks.

Oh, I wasn't saying that student debt is variable interest, just making a point about debt and inflation in general.

I'm very sympathetic to the idea of being very careful about publishing things that could spread capabilities ideas. However, I think there are several important things missing from your world model, which cause me to believe that following your advice would substantially hurt alignment progress.

(To be clear, none of this applies to alignment people working directly on capabilities, who should, like, not. Rather, this is about alignment researchers accidentally advancing capabilities by talking to capabilities people)

• It's genuinely hard to come up with ideas that help capabilities a lot. I think you are severely underestimating how hard it is, and how much insight is required. I think one issue here is that most papers on arxiv are garbage and don't actually make any progress, but those papers are not the ones that are pushing AGI forward anyways.
• Even if you try very hard to do so, it's still very hard to convince people that you're right if you don't have a ton of clout via a legible reputation of being right a lot. Everyone has an agenda they're convinced will solve AGI and is too busy trying to convince everyone else to work on their agenda.
• High level ideas are generally not that valuable in and of themselves. People generally learn to ignore ideas unless they have strong empirical evidence of correctness (or endorsement of highly respected researchers) because there are simply too many ideas. The valuable thing is not the idea itself, but the knowledge of which ideas are actually correct.
• I think deeply understanding top tier capabilities researchers' views on how to achieve AGI is actually extremely valuable for thinking about alignment. Even if you disagree on object level views, understanding how very smart people come to their conclusions is very valuable.
• I think alignment discourse is greatly harmed by people being too scared to say things. When it bleeds over to being too scared to think about capabilities related topics for fear of accidentally generating something dangerous, I think this is even more harmful.

Yeah, I am not super familiar with PCA, but my understanding is that while both PCA and referential containment can be used to extract lower-dimensional or more compact representations, they operate on different types of data structures (feature vectors vs. graphs/hypergraphs) and have different objectives (capturing maximum variance vs. identifying self-contained conceptual chunks). Referential containment is more focused on finding semantically meaningful and contextually relevant substructures within a causal or relational knowledge representation. It also tries to address the opposite direction, basically how to break existing concepts apart when zooming into the representations, and I am not sure if something like that is done with PCA.

I had Claude 3 read this post and compare the two. Here it is, if you are interested (keep in mind that Claude tends to be very friendly and encouraging, so it might be valuing referential containment too highly):
Similarities:

• Both aim to identify patterns and structure in high-dimensional data.
• Both can be used to reduce the dimensionality of data by finding lower-dimensional representations.
• Both rely on analyzing statistical dependencies or correlations between variables/features.

Differences:

1. PCA is a linear dimensionality reduction technique that projects data onto a lower-dimensional subspace spanned by the principal components (eigenvectors of the covariance matrix). Referential containment is not inherently a dimensionality reduction technique but rather a way to identify self-contained or tightly coupled substructures within a causal graph or hypergraph representation.
2. PCA operates on numerical feature vectors, while referential containment deals with graph/hypergraph structures representing conceptual relationships and causal dependencies.
3. PCA finds orthogonal linear combinations of features that capture maximum variance. Referential containment aims to identify subgraphs with strong internal connections and minimal external connections, which may not necessarily correspond to maximum variance directions.
4. PCA is an unsupervised technique that does not consider semantics or labeled data. Referential containment, as proposed in the context of the PSCA architecture, can potentially leverage semantic information and learned representations to guide the identification of meaningful conceptual chunks.
5. PCA produces a fixed set of principal components for a given dataset. Referential containment, as described, is more flexible and can adapt the level of granularity or resolution based on context, objectives, and resource constraints.

Hmm, I think the first bullet point is pretty precisely what I am talking about (though to be clear, I haven't read the paper in detail). 

I was specifically saying that trying to somehow get feedback from future tokens into the next token objective would probably do some interesting things and enable a bunch of cross-token optimization that currently isn't happening, which would improve performance on some tasks. This seems like what's going on here.

Agree that another major component of the paper is accelerating inference, which I wasn't talking about. I would have to read the paper in more detail to get a sense of how much it's just doing that, in which case I wouldn't think it's a good example.

I think these judgements would benefit from more concreteness: that rather than proposing a dichotomy of "capabilities research" (them, Bad) and "alignment research" (us, Good), you could be more specific about what kinds of work you want to see more and less of.

I agree that (say) Carmack and Sutton are doing a bad thing by declaring a goal to "build AGI" while dismissing the reasons that this is incredibly dangerous. But the thing that makes infohazard concerns so fraught is that there's a lot of work that potentially affects our civilization's trajectory into the machine intelligence transition in complicated ways, which makes it hard to draw a boundary around "trusted alignment researchers" in a principled and not self-serving way that doesn't collapse into "science and technology is bad".

We can agree that OpenAI as originally conceived was a bad idea. What about the people working on music generation? That's unambiguously "capabilities", but it's also not particularly optimized at ending the world that way "AGI for AGI's sake" projects are. If that's still bad even though music generation isn't going to end the world (because it's still directing attention and money into AI, increasing the incentive to build GPUs, &c.), where do you draw the line? Some of the researchers I cited in my most recent post are working on "build[ing] better models of primate visual cognition". Is that wrong? Should Judea Pearl not have published? Turing? Charles Babbage?

In asking these obnoxious questions, I'm not trying to make a reductio ad absurdum of caring about risk, or proposing an infinitely slippery slope where our only choices are between max accelerationism and a destroy-all-computers Butlerian Jihad. I just think it's important to notice that "Stop thinking about AI" kind of does amount to a Butlerian Jihad (and that publishing and thinking are not unrelated)?

I would bet that sunscreen dwarfs all other effects in the long run 

I think that "epistemic rationality" matches very well with what I am thinking of as level 3, which is my notion of intelligence. It is indeed applicable to non-agentic systems.
I am still thinking about whether to include meta-learning (referring to updating level 3 algorithms based on experience) and meta-processes above that in my concept of intelligence.

Would this layer of meta-learning be part of epistemic rationality, do you think? It becomes particularly relevant if the system is resource constrained and has to prioritize what to learn about, and/or cares about efficiency of learning. These constraints feel a bit less natural to introduce for a non-agentic system, other than if said system is set up by an agentic system for some purpose. 

In any case, instrumental rationality does not seem to cover all that I mean with Competence, perhaps something narrower, like "cognitive competence". I find it a bit difficult to systematically distinguish between cognitive competence and noncognitive competence, because the cognitive part of the system is also implemented by its embodiment - and there are various correlations between events in "morphospace" and "cognitive space". 
One way of resolving that might be to distinguish between on-surface properties of a markov blanket (corresponding to the system interfacing with its environment) and within-surface properties of that markov blanket (corresponding to integrated regulation systems and the cognition). There will still be feedback loops between those properties, so our mileage on obtaining clean distinctions into different competencies may vary. 
If you are interested in some more thoughts on that, you can check out my post on extended embodiment.

An independent idea is that it is apparently possible to divide "competence" or "instrumental rationality" into two independent axes: Generality, and intelligence proper (or perhaps: competence proper).

I have already commented a bit on meta-learning above, per default my level 3 would refer just to online learning, but I am thinking of including different levels of meta-learning because of the algorithmic similarities. 
Perhaps interestingly for you, I consider one of the primary purposes of meta-learning to refine a generally intelligent system into a more narrowly intelligent system, by improving its learning capabilities for a particular set of domains, in some sense biasing the cognition towards the kind of environment it seems to operate within (i.e. in terms of hypothesis generation, or which kinds of functions to use when approximating the behavior of an observed sequence).

Of course, unless the system loses its meta-learning capability, it will be able to respond to changes in its environment by re-aiming/updating its learning tendencies over time - so it is technically general if you give it some time, but ends up converging towards beneficial specialisation.

I think of generality of intelligence as relatively conceptually trivial. At the end of the day, a system is given a sequence of data via observation, and is now tasked with finding a function or set of functions that both corresponds to plausible transition rules of the given sequence, and has a reasonably high chance of correctly predicting the next element of the sequence (which is easy to train for by hiding later elements of the sequence from the modeling process and sequentially introducing them to test and potentially update the fit of the model). 
Computationally speaking, the set of total atomic functions that you would have to consider in order to be able to compositionally construct arbitrary transition rules for sequences of discrete data packages, is very small. The only mathematical requirement is Turing universality - basically the entire difficulty arises due to resource constraints.

This seems to match with your thoughts about the appearance of greater generality simply due to more processing power. A cognitive system that is provided with more processing power could use that either to search more deeply those regions of causal models that it naturally considers, or it could branch out to consider new regions within model-space. Many brains in the animal kingdom seem to implement a sort of limited generative simulation of their environment, so that could be considered as a fairly general problem domain.

I could try to write more on this, but I am curious what you think about this so far and if I come across as reasonably clear.