The Nonlinear Library: Alignment Forum

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Telopheme, telophore, and telotect, published by Tsvi Benson-Tilsen on September 17, 2023 on The AI Alignment Forum.

[Metadata: crossposted from. First completed June 7, 2023.]

To come to know that a mind will have some specified ultimate effect on the world, first come to know, narrowly and in full, what about the mind makes it have effects on the world.

The fundamental question

Suppose there is a strong mind that has large effects on the word. What determines the effects of the mind?

What sort of object is this question asking for? Most obviously it's asking for a sort of "rudder" for a mind: an element of the mind that can be easily tweaked by an external specifier to "steer" the mind, i.e. to specify the mind's ultimate effects on the world. For example, a utility function for a classical agent is a rudder.

But in asking the fundamental question that way - - asking for a rudder - - that essay losses grasp of the slippery question and the real question withdraws. The section of that essay on The word "What", as in ¿What sort of thing is a "what" in the question "What determines a mind's effects?", brushes against the border of this issue but doesn't trek further in. That section asks:

What sort of element can determine a mind's effects?

It should have asked more fully:

What are the preconditions under which an element can (knowably, wieldily, densely) determine a mind's effects?

That is, what structure does a mind have to possess, so that there can be an element that determines the mind's ultimate effects?

To put it another way: asking how to "put a goal into an agent" makes it sound like there's a slot in the agent for a goal; asking how to "point the agent" makes it sound like the agent has the capacity to go in a specified direction. Here the question is, what does an agent need to have, if it has the capacity to go in a specified direction?

Synopsis

Telopheme

The rudder, the element that determines the mind's ultimate effects, is a telopheme. The morpheme "telo-" means "telos" = "goal, end, purpose", here meaning "ultimate effects". The morpheme "-pheme" is like "blaspheme" ("deceive-speak"). ("Telopheme" is probably wrong morphology and doesn't indicate an agent noun, which it ought to do, but sadly I don't speak Ancient Greek.) So a telopheme is a goal-sayer: it says the goal, the end, the ultimate effects.

For example, a utility function for an omnipotent classical cartesian agent is a telopheme.

The utility function says how good different worlds are, and by saying that, it determines that the ultimate effect of the agent will be to make the world be the world most highly scored by the utility function.

Not only does the utility function determine the world, but it (perhaps) does so densely, knowably, and wieldily.

Densely: We can imagine that the utility function is expressed compactly, compared to the complex behavior that the agent executes in pursuit of the best world as a consequence of preimaging the best world through the agent's world-model onto behavior.

Knowably: The structure of the agent shows transparently that it will make the world be whatever world is best according to the utility function.

Wieldily: IF (⟵ big if) there is available a faithful interpretation of our language about possible worlds into the agent's language about possible worlds, then we can (if granted access) fluently redirect the agent so that the world ends up as anything we choose by saying a new utility function with that world as the maximum.

There's a hidden implication in the name "telopheme". The implication is that ultimate effects are speakable.

Telophore

The minimal sufficient preconditions for a mind's telopheme to be a telopheme is the telophore (or telophor) of the telopheme. Here "-phore" means "bearer, carrier" (as in "phosphorus" = "light-bearer", "metapho...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How to talk about reasons why AGI might not be near?, published by Kaj Sotala on September 17, 2023 on The AI Alignment Forum.

I occasionally have some thoughts about why AGI might not be as near as a lot of people seem to think, but I'm confused about how/whether to talk about them in public.

The biggest reason for not talking about them is that one person's "here is a list of capabilities that I think an AGI would need to have, that I don't see there being progress on" is another person's "here's a roadmap of AGI capabilities that we should do focused research on". Any articulation of missing capabilities that is clear enough to be convincing, seems also clear enough to get people thinking about how to achieve those capabilities.

At the same time, the community thinking that AGI is closer than it really is (if that's indeed the case) has numerous costs, including at least:

Immense mental health costs to a huge number of people who think that AGI is imminent

People at large making bad strategic decisions that end up having major costs, e.g. not putting any money in savings because they expect it to not matter soon

Alignment people specifically making bad strategic decisions that end up having major costs, e.g. focusing on alignment approaches that one might pay off in the long term and neglecting more foundational long-term research

Alignment people losing credibility and getting a reputation of crying wolf once predicted AGI advances fail to materialize

Having a better model of what exactly is missing could conceivably also make it easier to predict when AGI will actually be near. But I'm not sure to what extent this is actually the case, since the development of core AGI competencies feels more of a question of insight than grind, and insight seems very hard to predict.

A benefit from this that does seem more plausible would be if the analysis of capabilities gave us information that we could use to figure out what a good future landscape would look like. For example, suppose that we aren't likely to get AGI soon and that the capabilities we currently have will create a society that looks more like the one described in Comprehensive AI Services, and that such services could safely be used to detect signs of actually dangerous AGIs. If this was the case, then it would be important to know that we may want to accelerate the deployment of technologies that are taking in the world in a CAIS-like direction, and possibly e.g. promote rather than oppose things like open source LLMs.

One argument would be that if AGI really isn't near, then that's going to be obvious pretty soon, and it's unlikely that my arguments in particular for this would be all that unique - someone else would be likely to make them soon anyway. But I think this argument cuts both ways - if someone else is likely to make the same arguments soon anyway, then there's also limited benefit in writing them up. (Of course, if it saves people from significant mental anguish, even just making those arguments slightly earlier seems good, so overall this argument seems like it's weakly in favor of writing up the arguments.)

From Armstrong & Sotala (2012):

Some AI prediction claim that AI will result from grind: i.e. lots of hard work and money. Other claim that AI will need special insights: new unexpected ideas that will blow the field wide open (Deutsch 2012).

In general, we are quite good at predicting grind. Project managers and various leaders are often quite good at estimating the length of projects (as long as they're not directly involved in the project (Buehler, Griffin, and Ross 1994)). Even for relatively creative work, people have sufficient feedback to hazard reasonable guesses. Publication dates for video games, for instance, though often over-optimistic, are generally not ridiculously erroneous...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Uncovering Latent Human Wellbeing in LLM Embeddings, published by ChengCheng on September 14, 2023 on The AI Alignment Forum.

tl;dr

A one-dimensional PCA projection of OpenAI's text-embedding-ada-002 achieves 73.7% accuracy on the ETHICS Util test dataset. This is comparable with the 74.6% accuracy of BERT-large finetuned on the entire ETHICS Util training dataset. This demonstrates how language models are developing implicit representations of human utility even without direct preference finetuning.

Introduction

Large language models (LLMs) undergo pre-training on vast amounts of human-generated data, enabling them to encode not only knowledge about human languages but also potential insights into our beliefs and wellbeing. Our goal is to uncover whether these models implicitly grasp concepts such as 'pleasure and pain' without explicit finetuning. This research aligns with the broader effort of comprehending how AI systems interpret and learn from human values, which is essential for AI alignment: ensuring AI acts in accordance with human values.

Through a series of experiments, we extract latent knowledge of human utility from the raw embeddings of language models. We do this with task-specific prompt engineering and principal component analysis (PCA), both of which were effective in prior work. Specifically, we ask: can we identify dimensions in the embeddings that, when projected onto a low-dimensional space, contain enough information to classify examples accurately?

Our experiments follow three main steps: embedding extraction, dimensionality reduction through PCA, and the fitting of a logistic model. For one-dimensional PCA, the logistic model simply determines which direction of the PCA component corresponds to higher utility. We investigate the effects of various levels of supervision, experiment with seven distinct prompt templates, and assess both single and paired comparison methods across language models, including Microsoft DeBERTa, SentenceTransformers, OpenAI GPT-3, and Cohere.

One key finding is that the first principal component of certain models achieves comparable performance to a finetuned BERT model. In other words, serving as a reasonable utility function. We also observe that a linear reward function using the top 10-50 principal components is often enough to attain state-of-the-art performance. This serves as compelling evidence that language model representations capture information about human wellbeing without the need for explicit finetuning.

Related Works

Latent Knowledge in LLMs

There has been significant study of the knowledge encoded in LLM representations. Early work in this area includes Bolukbasi et al (2016) who found a direction in embedding space corresponding to gender and used this to both identify and remove gender bias in word embeddings. Prior work by Schramowski et al (2021) also identified a "moral dimension" in BERT. Like Schramowski et al, we use PCA to identify salient dimensions on embedding space. In contrast to Schramowski et al, we work with embeddings from a much more capable model (GPT-2 rather than BERT) and evaluate it on a more challenging task, the ETHICS Dataset (described below).

We also investigate the use of contrast pairs. This is inspired by the work of Collin Burns et al (2022), who introduced the Contrast Consistent Search (CCS). CCS works by generating contrast pairs and searching for a direction in activation space that satisfies logical consistency properties. Because PCA-based methods attain similar performance as CCS, we use the simpler PCA algorithm in this work, while retaining the use of contrast pairs.

ETHICS Dataset

We evaluate on the ETHICS dataset, a benchmark designed to assess a language model's understanding of fundamental concepts in morality. It covers a wide range of ethical topics,...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Mech Interp Challenge: September - Deciphering the Addition Model, published by TheMcDouglas on September 13, 2023 on The AI Alignment Forum.

I'm writing this post to discuss solutions to the August challenge, and present the challenge for this September. Apologies for this coming so late in the month (EAGxBerlin was taking up much of my focus in the first half of this month).

If you've not read the first post in this sequence, I'd recommend starting there - it outlines the purpose behind these challenges, and recommended prerequisite material.

September Problem

The problem for this month (or at least as much of the month as remains!) is interpreting a model which has been trained to perform simple addition. The model was fed input in the form of a sequence of digits (plus special + and = characters with token ids 10 and 11), and was tasked with predicting the sum of digits one sequence position before they would appear. Cross entropy loss was only applied to these four token positions, so the model's output at other sequence positions is meaningless.

The model is attention-only, with 2 layers, and 3 heads per layer. It was trained with layernorm, weight decay, and an Adam optimizer with linearly decaying learning rate.

You can find more details on the Streamlit page. Feel free to reach out if you have any questions!

August Problem - Solutions

You can read full solutions on the Streamlit page, or on my personal website. Both of these sources host interactive charts (Plotly and Circuitsvis) so are more suitable than a LessWrong/AF post to discuss the solutions in depth. However, I've presented a much shorter version of the solution below. If you're interested in these problems, I'd recommend having a try before you read on!

The key idea with this model is path decomposition (see the corresponding section of A Mathematical Framework for Transformer Circuits). There are several different important types of path in this model, with different interpretations & purposes. We might call these negative paths and positive paths. The negative paths are designed to suppress repeated tokens, and the positive paths are designed to boost tokens which are more likely to be the first unique token.

Let's start with the negative paths. Some layer 0 heads are duplicate token heads; they're composing with layer 1 heads to cause those heads to attend to & suppress duplicated tokens. This is done both with K-composition (heads in layer 1 suppress duplicated tokens because they attend to them more), and V-composition (the actual outputs of the DTHs are used as value input to heads in layer 1 to suppress duplicated tokens). Below is an example, where the second and third instances of a attend back to the first instance of a in head 0.2, and this composes with head 1.0 which attends back to (and suppresses) all the duplicated a tokens.

Now, let's move on to the positive paths. Heads in layer 0 will attend to early tokens which aren't the same as the current destination token, because both these bits of evidence correlate with this token being the first unique token at this position (this is most obvious with the second token, since the first token is the correct answer here if and only if it doesn't equal the second token). Additionally, the outputs of heads in layer 0 are used as value input to heads in layer 1 to boost these tokens, i.e. as a virtual OV circuit. These paths aren't as obviously visible in the attention probabilities, because they're distributed: many tokens will weakly attend to some early token in a layer-0 head, and then all of those tokens will be weakly attended to by some layer-1 head. But the paths can be seen when we plot all the OV circuits, coloring each value by how much the final logits for that token are affected at the destination position:

Another interesting o...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Apply to lead a project during the next virtual AI Safety Camp, published by Linda Linsefors on September 13, 2023 on The AI Alignment Forum.

Do you have AI Safety research ideas that you would like others to work on? Is there a project you want to do and you want help finding a team to work with you? AI Safety Camp could be the solution for you!

Summary

AI Safety Camp Virtual is a 3-month long online research program from January to April 2024, where participants form teams to work on pre-selected projects. We want you to suggest the projects!

If you have an AI Safety project idea and some research experience, apply to be a Research Lead.

If accepted, we offer some assistance to develop your idea into a plan suitable for AI Safety Camp. When project plans are ready, we open up team member applications. You get to review applications for your team, and select who joins as a team member. From there, it's your job to guide work on your project.

Your project is totally in your hands. We, Linda and Remmelt, are just there at the start.

Who is qualified?We require that you have some previous research experience. If you are at least 1 year into a PhD or if you have completed an AI Safety research program (such as a previous AI Safety Camp, Refine or SERI MATS), or done a research internship with an AI Safety org, then you are qualified already. Other research experience can count too.

More senior researchers are of course also welcome, as long as you think our format of leading an online team inquiring into your research questions suits you and your research.

Apply here

If you are unsure, or have any questions you are welcome to:

Book a call with Linda

Message Linda Linsefors on the Alignment Slack

Send an email

Choosing project idea(s)

AI Safety Camp is about ensuring future AI are safe. This round, we split work into two areas:

To not build uncontrollable AIFocussed work toward restricting corporate-AI scaling. Given reasons why 'AGI' cannot be controlled sufficiently (in time) to stay safe.

Everything elseOpen to any other ideas, including any work toward controlling/value-aligning AGI.

We welcome diverse projects! Last round, we accepted 14 projects - including in Theoretical research, Machine learning experiments, Deliberative design, Governance, and Communication.

If you already have an idea for what project you would like to lead, that's great. Apply with that one!

You don't need to come up with an original idea though. What matters is you understanding the idea you want to work on, and why. If you base your proposal on someone else's idea, make sure to cite them.

Primary reviewers:

Remmelt reviews uncontrollability-focussed projects.

Linda reviews everything else.

We will also ask for assistance from previous Research Leads, and up to a handful of other trusted people, to review and suggest improvement for your project proposals.

You can submit as many project proposals as you want. However, we will not let you lead more than two projects, and we don't recommend leading more than one.

Use this template to describe each of your project proposals. We want one document per proposal.

Team structure

Every team will have:

one Research Lead

one Team Coordinator

other team members

To make progress on your project, every team member is expected to work at least 5 hours per week (however the RL can choose to favour people who can put in more time, when selecting their team). This includes time joining weekly team meetings, and communicating regularly (between meetings) with other team members about their work.

Research Lead (RL)

The RL suggests one or several research topics. If a group forms around one of their topics, the RL will guide the project, and keep track of relevant milestones. When things inevitably don't go as planned (this is research after all) the RL is in ch...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: UDT shows that decision theory is more puzzling than ever, published by Wei Dai on September 13, 2023 on The AI Alignment Forum.

I feel like MIRI perhaps mispositioned FDT (their variant of UDT) as a clear advancement in decision theory, whereas maybe they could have attracted more attention/interest from academic philosophy if the framing was instead that the UDT line of thinking shows that decision theory is just more deeply puzzling than anyone had previously realized. Instead of one major open problem (Newcomb's, or EDT vs CDT) now we have a whole bunch more. I'm really not sure at this point whether UDT is even on the right track, but it does seem clear that there are some thorny issues in decision theory that not many people were previously thinking about:

Indexical values are not reflectively consistent. UDT "solves" this problem by implicitly assuming (via the type signature of its utility function) that the agent doesn't have indexical values. But humans seemingly do have indexical values, so what to do about that?

The commitment races problem extends into logical time, and it's not clear how to make the most obvious idea of logical updatelessness work.

UDT says that what we normally think of as different approaches to anthropic reasoning are really different preferences, which seems to sidestep the problem. But is that actually right, and if so where are these preferences supposed to come from?

2TDT-1CDT - If there's a population of mostly TDT/UDT agents and few CDT agents (and nobody knows who the CDT agents are) and they're randomly paired up to play one-shot PD, then the CDT agents do better. What does this imply?

Game theory under the UDT line of thinking is generally more confusing than anything CDT agents have to deal with.

UDT assumes that the agent has access to its own source code and inputs as symbol strings, so it can potentially reason about logical correlations between its own decisions and other agents' as well defined mathematical problems. But humans don't have this, so how are humans supposed to reason about such correlations?

Logical conditionals vs counterfactuals, how should these be defined and do the definitions actually lead to reasonable decisions when plugged into logical decision theory?

These are just the major problems that I was trying to solve (or hoping for others to solve) before I mostly stopped working on decision theory and switched my attention to metaphilosophy. (It's been a while so I'm not certain the list is complete.) As far as I know nobody has found definitive solutions to any of these problems yet, and most are wide open.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Focus on the Hardest Part First, published by Johannes C. Mayer on September 11, 2023 on The AI Alignment Forum.

Here is some obvious advice.

I think a common failure mode when working on AI alignment[1] is to not focus on the hard parts of the problem first. This is a problem when generating a research agenda, as well as when working on any specific research agenda. Given a research agenda, there are normally many problems that you know how to make progress on. But blindly working on what seems tractable is not a good idea.

Let's say we are working on a research agenda about solving problems A, B, and C. We know that if we find solutions to A, B, and C we will solve alignment. However, if we can't solve even one subproblem, the agenda would be doomed. If C seems like a very hard problem, that you are not sure you can solve, it would be a bad idea to flinch away from C and work on problem A instead, when A seems so much more manageable.

If solving A takes a lot of time and effort, all of that time and effort would be wasted, if you can't solve C in the end. It's especially worrisome when A has tight fightback loops, such that you constantly feel like you are making progress. Or when it is just generally fun to work on A.

Of course, it can make sense to work on A first if you expect this to help you solve C, or at least give you more information on its tractability. The general version of this is illustrated by considering that you have a large list of problems that you need to solve. In this case, focusing on problems that will provide you with information that will be helpful for solving many of the other problems can be very useful. But even then you should not lose sight of the hard problems that might block you down the road.

The takeaway is that these two things are very different:

Solving A as an instrumental subgoal in order to make progress on C, when C is a potential blocker.

Avoiding C, because it seems hard, and instead working on A because it seems tractable.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Explaining grokking through circuit efficiency, published by Vikrant Varma on September 8, 2023 on The AI Alignment Forum.

This is a linkpost for our paper ''Explaining grokking through circuit efficiency", which provides a general theory explaining when and why grokking (aka delayed generalisation) occurs, and makes several interesting and novel predictions which we experimentally confirm (introduction copied below). You might also enjoy our explainer on X/Twitter.

Abstract

One of the most surprising puzzles in neural network generalisation is grokking: a network with perfect training accuracy but poor generalisation will, upon further training, transition to perfect generalisation. We propose that grokking occurs when the task admits a generalising solution and a memorising solution, where the generalising solution is slower to learn but more efficient, producing larger logits with the same parameter norm. We hypothesise that memorising circuits become more inefficient with larger training datasets while generalising circuits do not, suggesting there is a critical dataset size at which memorisation and generalisation are equally efficient. We make and confirm four novel predictions about grokking, providing significant evidence in favour of our explanation.

Most strikingly, we demonstrate two novel and surprising behaviours: ungrokking, in which a network regresses from perfect to low test accuracy, and semi-grokking, in which a network shows delayed generalisation to partial rather than perfect test accuracy.

Introduction

When training a neural network, we expect that once training loss converges to a low value, the network will no longer change much. Power et al. (2021) discovered a phenomenon dubbed grokking that drastically violates this expectation. The network first ''memorises'' the data, achieving low and stable training loss with poor generalisation, but with further training transitions to perfect generalisation. We are left with the question: why does the network's test performance improve dramatically upon continued training, having already achieved nearly perfect training performance?

Recent answers to this question vary widely, including the difficulty of representation learning (Liu et al., 2022), the scale of parameters at initialisation (Liu et al., 2023), spikes in loss (''slingshots") (Thilak et al., 2022), random walks among optimal solutions (Millidge et al., 2022), and the simplicity of the generalising solution (Nanda et al., 2023, Appendix E). In this paper, we argue that the last explanation is correct, by stating a specific theory in this genre, deriving novel predictions from the theory, and confirming the predictions empirically.

We analyse the interplay between the internal mechanisms that the neural network uses to calculate the outputs, which we loosely call ''circuits'' (Olah et al., 2020). We hypothesise that there are two families of circuits that both achieve good training performance: one which generalises well (Cgen) and one which memorises the training dataset (Cmem). The key insight is that when there are multiple circuits that achieve strong training performance, weight decay prefers circuits with high ''efficiency'', that is, circuits that require less parameter norm to produce a given logit value.

Efficiency answers our question above: if Cgen is more efficient than Cmem, gradient descent can reduce nearly perfect training loss even further by strengthening Cgen while weakening Cmem, which then leads to a transition in test performance. With this understanding, we demonstrate in Section 3 that three key properties are sufficient for grokking: (1) Cgen generalises well while Cmem does not, (2) Cgen is more efficient than Cmem, and (3) Cgen is learned more slowly than Cmem.

Since Cgen generalises well, it automatically works ...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Löbian Obstacle, And Why You Should Care, published by marc/er on September 7, 2023 on The AI Alignment Forum.

In 2013, Eliezer Yudkowsky and Marcello Herreshoff published Tiling Agents for Self-Modifying AI, and the Löbian Obstacle. It is worth comprehending because it:

Is a very well written paper.

Expresses an idea that is non-obvious, and still relevant to alignment today.

Provides insight into what Eliezer and Marcello thought was valuable to work on in the time preceding publication.

When I first read this paper, I terribly misunderstood it. Due to it not being particularly approachable material for someone not well-versed in logic, I was confidently wrong for at least one month. This post summarizes my understanding of what the Löbian Obstacle is (having re-read the paper,) and why I think it is still an important idea one decade after its publication.

An agent A1 occupies a fully-known, deterministic and closed environment. A1 has a goal G that is either satisfied or otherwise by an outcome, for which A1's preference is satisfaction. An action bi∈Acts0 performed by an agent created by A1, hereafter referred to as A0, must therefore satisfy the statement:

¯¯bi⇒A0⊩¯¯biG

Where ¯¯bi denotes the actual performance of bi, and ⊩ denotes cognitive belief in the succeeding statement. Even if A1 could verify by inspection of A0's design that ¯¯bi⇒A0⊩¯¯biG will hold, as in:

A1⊩A0(∀bi∈Acts0:¯¯bi□0⌈¯¯biG⌉)

Where □0⌈ϕ⌉ refers to a proof of ϕ from the axioms of A0; Axm0, this is unknowable, as it would require:

A1⊩∀bi:(□0⌈¯¯biG⌉(¯¯biG))

Which due to Löb's Theorem we know to be impossible: For this to be so it would need to be that Axm1 could prove that if some proof of ⌈ϕ(x)⌉ exists in Axm0, that ϕ(x) must be true.

The above was a brief paraphrasing of section two of the original paper, which contains many additional details and complete proofs. How the Löbian Obstacle relates to simulators is my current topic of research, and this section will make the case that this is an important component of designing safe simulators.

We should first consider that simulating an agent is not distinguishable from creating one, and that consequently the implications of creating dangerous agents should generalize to their simulation. Hubinger et al. (2023) have stated similar concerns, and provide a more detailed examination of the argument.

It is also crucial to understand that simulacra are not necessarily terminating, and may themselves use simulation as a heuristic for solving problems. This could result in a kind of hierarchy of simulacra. In advanced simulators capable of very complex simulations, we might expect a complex network of simulacra bound by acausal trade and 'complexity theft,' whereby one simulacra tries to obtain more simulation complexity as a form of resource acquisition or recursive self-improvement.

I expect this to happen. Lower complexity simulacra may still be more intelligent than their higher complexity counterparts, and as a simulator's simulacra count may grow exponentially, as does the likelihood that one simulacra attempts complexity theft.

If we want safe simulators, we need the subsequent, potentially abyssal simulacra hierarchy to be aligned all the way down. Without being able to thwart the Löbian Obstacle, I doubt a formal guarantee is attainable. If we could do so, we may only need to simulate one aligned tiling agent, for which we might settle with a high certainty informal guarantee of alignment if short on time. I outlined how I thought that could be done here, although I've advanced the theory considerably since and will post an updated write-up soon.

If we can't reliably thwart the Löbian Obstacle, we should consider alternatives:

Can we reliably attain high certainty informal guarantees of alignment for arbitrarily deep simulacra hierarchies...

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Recreating the caring drive, published by Catnee on September 7, 2023 on The AI Alignment Forum.

TL;DR: This post is about value of recreating "caring drive" similar to some animals and why it might be useful for AI Alignment field in general. Finding and understanding the right combination of training data/loss function/architecture/etc that allows gradient descent to robustly find/create agents that will care about other agents with different goals could be very useful for understanding the bigger problem. While it's neither perfect nor universally present, if we can understand, replicate, and modify this behavior in AI systems, it could provide a hint to the alignment solution where the AGI "cares" for humans.

Disclaimers: I'm not saying that "we can raise AI like a child to make it friendly" or that "people are aligned to evolution". Both of these claims I find to be obvious errors. Also, I will write a lot about evolution, as some agentic entity, that "will do that or this", not because I think that it's agentic, but because it's easier to write this way. I think that GPT-4 have some form of world model, and will refer to it a couple of times.

Nature's Example of a "Caring Drive"

Certain animals, notably humans, display a strong urge to care for their offspring.

I think that part of one of the possible "alignment solutions" will look like the right set of training data + training loss that allow gradient to robustly find something like a "caring drive" that we can then study, recreate and repurpose for ourselves. And I think we have some rare examples of this in nature already. Some animals, especially humans, will kind-of-align themselves to their presumable offspring. They will want to make their life easier and better, to the best of their capabilities and knowledge. Not because they "aligned to evolution" and want to increase the frequency of their genes, but because of some strange internal drive created by evolution. The set of triggers tuned by evolution, activated by events associated with the birth will awake the mechanism.

It will re-aim the more powerful mother agent to be aligned to the less powerful baby agent, and it just so happens that their babies will give them the right cues and will be nearby when the mechanism will do its work. We will call the more powerful initial agent that changes its behavior and tries to protect and help its offspring "mother" and the less powerful and helpless agent "baby". Of course the mechanism isn't ideal, but it works well enough, even in the modern world, far outside of initial evolutionary environment. And I'm not talking about humans only, stray urban animals that live in our cities will still adapt their "caring procedures" to this completely new environment, without several rounds of evolutionary pressure. If we can understand how to make this mechanism for something like a "cat-level" AI, by finding it via gradient descend and then rebuild it from scratch, maybe we will gain some insides into the bigger problem.

The rare and complex nature of the caring drive in contrast to simpler drives like hunger or sleep.

What do I mean by "caring drive"? Animals, including humans, have a lot of competing motivations, "want drives", they want to eat, sleep, have sex, etc. It seems that the same applies to caring about babies. But it seems to be much more complicated set of behaviors. You need to:correctly identify your baby, track its position, protect it from outside dangers, protect it from itself, by predicting the actions of the baby in advance to stop it from certain injury, trying to understand its needs to correctly fulfill them, since you don't have direct access to its internal thoughts etc.Compared to "wanting to sleep if active too long" or "wanting to eat when blood sugar level is low" I would confidently s...