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: [Linkpost] Inconsistent evidence for price substitution between butter and margarine: A shallow review, published by samaramendez on April 14, 2023 on The Effective Altruism Forum.

Executive summary

One prominent strategy for reducing animal product usage is to decrease the prices of plant-based analogs for meat, dairy and eggs. Cross-price elasticities measure how prices of analogs affect sales of animal products, and vice versa.

Previously, we found cross-price elasticities that sometimes indicated decreased plant-based milk prices could cause increased consumption of dairy milk. To see whether this result replicated, we studied elasticities of butter and margarine.

We synthesize 52 cross-price elasticities from 19 demand studies of butter and margarine.

We expected butter and margarine to be substitutes. Instead, we observed wide variation in estimates, complementarity, and even opposite signs of two elasticities for the same pair of products in the same setting (study, time, country, etc.).

Margarine was a substitute for butter in two thirds of estimates, while butter was a substitute for margarine in about half of estimates.

The results are likely explained by some combination of methodological issues and complex consumer behavior.

Estimating elasticities from observational data is very difficult, and there is some evidence of methodological issues in the literature. Consumer behavior may be highly context specific.

Our best highly uncertain guess is that methodological issues contribute 45% of the observed variation in results, with complex consumer behavior accounting for the remaining 55%.

Price substitution between plant-based analogs and animal products is not a certainty.

It is possible that decreasing plant-based analog prices causes harmful increases in animal product usage in some contexts.

Research aiming to test behavioral theories that might explain our results and to validate observational estimates of elasticities against experimental methods may help clarify consumer behavior.

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

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: Moderation notes re: recent Said/Duncan threads, published by Raemon on April 14, 2023 on LessWrong.

Recently there's been a series of posts and comment back-and-forth between Said Achmiz and Duncan Sabien, which escalated enough that it seemed like site moderators should weigh in.

For context, a quick recap of recent relevant events as I'm aware of them are. (I'm glossing over many details that are relevant but getting everything exactly right is tricky)

Duncan posts Basics of Rationalist Discourse. Said writes some comments in response.

Zack posts "Rationalist Discourse" Is Like "Physicist Motors", which Duncan and Said argue some more and Duncan eventually says "goodbye" which I assume coincides with banning Said from commenting further on Duncan's posts.

I publish LW Team is adjusting moderation policy. Lionhearted suggests "Basics of Rationalist Discourse" as a standard the site should uphold. Paraphrasing here, Said objects to a post being set as the site standards if not all non-banned users can discuss it. More discussion ensues.

Duncan publishes Killing Socrates, a post about a general pattern of LW commenting that alludes to Said but doesn't reference him by name. Commenters other than Duncan do bring up Said by name, and the discussion gets into "is Said net positive/negative for LessWrong?" in a discussion section where Said can't comment.

@gjm publishes On "aiming for convergence on truth", which further discusses/argues a principle from Basics of Rationalist Discourse that Said objected to. Duncan and Said argue further in the comments. I think it's a fair gloss to say "Said makes some comments about what Duncan did, which Duncan says are false enough that he'd describe Said as intentionally lying about them. Said objects to this characterization" (although exactly how to characterize this exchange is maybe a crux of discussion)

LessWrong moderators got together for ~2 hours to discuss this overall situation, and how to think about it both as an object-level dispute and in terms of some high level "how do the culture/rules/moderation of LessWrong work?".

I think we ended up with fairly similar takes, but, getting to the point that we all agree 100% on what happened and what to do next seemed like a longer project, and we each had subtly different frames about the situation. So, some of us (at least Vaniver and I, maybe others) are going to start by posting some top level comments here. People can weigh in the discussion. I'm not 100% sure what happens after that, but we'll reflect on the discussion and decide on whether to take any high-level mod actions.

If you want to weigh in, I encourage you to take your time even if there's a lot of discussion going on. If you notice yourself in a rapid back and forth that feels like it's escalating, take at least a 10 minute break and ask yourself what you're actually trying to accomplish.

I do note: the moderation team will be making an ultimate call on whether to take any mod actions based on our judgment. (I'll be the primary owner of the decision, although I expect if there's significant disagreement among the mod team we'll talk through it a lot). We'll take into account arguments various people post, but we aren't trying to reflect the wisdom of crowds.

So if you may want to focus on engaging with our cruxes rather than what other random people in the comments think.

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

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: Prospects for AI safety agreements between countries, published by oeg on April 14, 2023 on The Effective Altruism Forum.

This post summarizes a project that I recently completed about international agreements to coordinate on safe AI development. I focused particularly on an agreement that I call “Collaborative Handling of Artificial intelligence Risks with Training Standards” (“CHARTS”). CHARTS would regulate training runs and would include both the US and China.

Among other things, I hope that this post will be a useful contribution to recent discussions about international agreements and regulatory regimes for AI.

I am not (at least for the moment) sharing the entire project publicly, but I hope that this summary will still be useful or interesting for people who are thinking about AI governance. The conclusions here are essentially the same as in the full version, and I don’t think readers’ views would differ drastically based on the additional information present there. If you would benefit from reading the full version (or specific parts) please reach out to me and I may be able to share.

This post consists of an executive summary (∼1200 words) followed by a condensed version of the report (∼5000 words + footnotes).

Executive summary

In this report, I investigate the idea of bringing about international agreements to coordinate on safe AI development (“international safety agreements”), evaluate the tractability of these interventions, and suggest the best means of carrying them out.

Introduction

My primary focus is a specific type of international safety agreement aimed at regulating AI training runs to prevent misalignment catastrophes. I call this agreement “Collaborative Handling of Artificial intelligence Risks with Training Standards,” or “CHARTS.”

Key features of CHARTS include:

Prohibiting governments and companies from performing training runs with a high likelihood of producing powerful misaligned AI systems.

Risky training runs would be determined using proxies like training run size, or potential risk factors such as reinforcement learning.

Requiring extensive verification through on-chip mechanisms, on-site inspections, and dedicated institutions.

Cooperating to prevent exports of AI-relevant compute to non-member countries, avoiding dangerous training runs in non-participating jurisdictions.

I chose to focus on this kind of agreement after seeing similar proposals and thinking that they seemed promising. My intended contribution here is to think about how tractable it would be to get something like CHARTS, and how to increase this tractability. I mostly do not attempt to assess how beneficial (or harmful) CHARTS would be.

I focus particularly on an agreement between the US and China because existentially dangerous training runs seem unusually likely to happen in these countries, and because these countries have an adversarial relationship, heightening concerns about racing dynamics.

Political acceptance of costly measures to regulate AI

I introduce the concept of a Risk Awareness Moment (Ram) as a way to structure my thinking elsewhere in the report. A Ram is “a point in time, after which concern about extreme risks from AI is so high among the relevant audiences that extreme measures to reduce these risks become possible, though not inevitable.” Examples of audiences include the general public and policy elites in a particular country.

I think this concept is helpful for thinking about a range of AI governance interventions. It has advantages over related concepts such as “warning shots” in that it makes it easier to remain agnostic about what causes increased concern about AI, and what the level of concern looks like over time.

I think that CHARTS would require a Ram among policy elites, and probably also among the general public – at least in ...

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: Anti-'FOOM' (stop trying to make your cute pet name the thing), published by david reinstein on April 14, 2023 on The Effective Altruism Forum.

Notes/basis: This is kind of a short-form post in style but I think it's important enough to put here. Obviously let me know if someone else said this better

Summary

Formal overly-intellectual academese is bad. But using your 'cute' inside joke name for things is potentially worse. It makes people cringe, sounds like you are trying to take ownership of something, and excludes people. Use a name that is approachable but serious.

The problem.

Where did the term 'FOOM' come from, to refer to AGI risk? I asked GPT4:

[!ai] AI

The term 'foom' was coined by artificial intelligence researcher and author Eliezer Yudkowsky in his 2008 book titled "The Sequences". Yudkowsky used the term to refer to a hypothetical scenario where an artificial general intelligence (AGI) rapidly and exponentially improves its own intelligence, leading to an uncontrollable and potentially catastrophic outcome for humanity. The term 'foom' is a play on the word 'boom', representing the sudden and explosive nature of AGI development in this scenario.

Another example: 'AI-not-kill-everyone-ism'

Analogies to fairly successful movements:

Global warming was not called ''Roast", and the movement was not called "anti-everyone-burns-up-ism"

Nuclear holocaust was not called "mega-boom"

Anti-slavery was not called ... (OK I won't touch this one)

How well has the use of cute names worked in the past?

I can't think of any examples where they have caught on in a positive way. The closest I can think of are

"Nudge" (by Richard Thaler?)

... to describe choice-architecture interventions; - My impression is that the term 'nudge' got people to remember it but made it rather easy to dismiss

others in that space have come up with names that caught on less well I think (like "sludge"), which also induce a bit of cringe

"Woke"

I think this example basically speaks for itself.

Tea-Party movement

This goes in the opposite direction perhaps (fairly successful), but I still think it's not quite as cringeworthy as FOOM. The term 'tea party' obviously has a long history in our culture, especially the "Boston Tea Party.

What else?

I asked GPT4

when have social movements used cute 'inside joke' names to refer to the threats faced?

The suggestions are not half as cute or in-jokey as FOOM: Net Neutrality, The Umbrella Movement, Extinction Rebellion (XR), Occupy Wall Street (OWS)

I asked it to get cuter... [1]

Prodding it further... Climategate, Frankenfoods, Slacktivism ... also not so inside-jokey nor as cringeworthy IMO.

Prodding it for more cutesy more inside-jokey yields a few terms that barely caught on, or didn't characterize the movement or the threat as a whole.[2]

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

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 self-unalignment problem, published by Jan Kulveit on April 14, 2023 on The AI Alignment Forum.

The usual basic framing of alignment looks something like this:

We have a system “A” which we are trying to align with system "H", which should establish some alignment relation “f” between the systems. Generally, as the result, the aligned system A should do "what the system H wants".Two things stand out in this basic framing:

Alignment is a relation, not a property of a single system. So the nature of system H affects what alignment will mean in practice.

It’s not clear what the arrow should mean.

There are multiple explicit proposals for this, e.g. some versions of corrigibility, constantly trying to cooperatively learn preferences, some more naive approaches like plain IRL, some empirical approaches to aligning LLMs.

Even when researchers don’t make an explicit proposal for what the arrow means, their alignment work still rests on some implicit understanding of what the arrow signifies.

But humans are self-unaligned

To my mind, existing alignment proposals usually neglect an important feature of the system "H" : the system "H" is not self-aligned, under whatever meaning of alignment is implied by the alignment proposal in question.

Technically, taking alignment as relation, and taking the various proposals as implicitly defining what it means to be ‘aligned’, the question is whether the relation is reflexive.

Sometimes, a shell game seems to be happening with the difficulties of humans lacking self-alignment - e.g. assuming if the AI is aligned, it will surely know how to deal with internal conflict in humans.

While what I'm interested in is the abstract problem, best understood at the level of properties of the alignment relation, it may be useful to illustrate it on a toy model.

In the toy model, we will assume a specific structure of system "H":

A set of parts p1..pn, with different goals or motivations or preferences. Sometimes, these parts might be usefully represented as agents; other times not.

A shared world model.

An aggregation mechanism Σ, translating what the parts want into actions, in accordance with the given world model.

In this framing, it’s not entirely clear what the natural language pointer ‘what system H wants’ translates to. Some compelling options are:

The output of the aggregation procedure.

What the individual parts want.

The output of a pareto-optimal aggregation procedure.

For any operationalization of what alignment means, we can ask if system H would be considered ‘self-aligned’, that is, if the alignment relation would be reflexive. For most existing operationalizations, it’s either unclear if system H is self-aligned, or clear that it isn’t.

In my view, this often puts the whole proposed alignment structure on quite shaky grounds.

Current approaches mostly fail to explicitly deal with self-unalignment

It’s not that alignment researchers believe that humans are entirely monolithic and coherent. I expect most alignment researchers would agree that humans are in fact very messy.

But in practice, a lot of alignment researcher seem to assume that it’s fine to abstract this away. There seems to be an assumption that alignment (the correct operationalization of the arrow f) doesn’t depend much on the contents of the system H box. So if we abstract the contents of the box away and figure out how to deal with alignment in general, this will naturally and straightforwardly extend to the messier case too.

I think this is incorrect. To me, it seems that:

Current alignment proposals implicitly deal with self-unalignment in very different ways.

Each of these ways poses problems.

Dealing with self-unalignment can’t be postponed or delegated to powerful AIs to deal with.

The following is a rough classification of the main implicit solutions to self-u...

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: R0 Is Not Counterfactual, published by jefftk on April 13, 2023 on LessWrong.

Someone who helps organize an

N95-required dance recently wrote to a group of organizers:

R0 is the number of people that an infected individual is likely to infect. R0 was 5.4 in Dec 2022. ... People who assume increased risk for themselves are also assuming increased risk for 5.4 other people.

This is wrong in two main ways, and I responded on the list, but I wanted to share my response here as well because the claim illustrates two common misconceptions.

The first issue is that R0 is for an entirely "susceptible" population, one in which no one has any gained any immunity from exposure or vaccination. What an R0 of

5.4 would mean is that if it had suddenly appeared in 2019, each infected person would on average infect 5.4 others. This is very different from the current situation where most people have had several shots, plus most people have had covid at least once. The term for the expected number of people an infected person will directly infect given current conditions is Rt, currently about 1. Which is really just another way of saying that covid levels have been changing relatively slowly: if Rt were 5.4 we'd have rapid growth tearing through the population.

The other issue is that even Rt doesn't tell you how many infections you getting sick would cause. You may know various things about your behavior that make the expected number of people you'd directly infect higher or lower, but that's not the main issue. Instead it's that (a) people you infect can go on to infect other people and (b) people you infect might otherwise have been infected by other people. These two factors push in opposite directions, but both can be quite large. Here are a pair of toy situations showing how, holding Rt fixed, one infection can lead to either very many or almost no counterfactual infections:

A new epidemic is starting, and Rt (which is R0 in this case) is 5.4. There are very clear symptoms, and people are just starting to catch on. Very soon there will be massive behavior changes to suppress Rt, and at this stage those might or might not be enough. One more person getting infected could have a 5% impact on the chance that this becomes a pandemic infecting ~half the world. In which case the expected number of additional infections is ~200M people. Here

(a) is the main factor.

An new pandemic is well under way, and it has easily missed symptoms. Even with lots of precautions in place, Rt is still a very high 5.4. There are so many paths by which a person can get infected that one additional infection has almost no effect on how many people eventually get infected. Here (b) is the main factor.

Now, "how many counterfactual infections would I cause if I got sick", and "how would my getting sick shift the distribution of when other people get sick" are really valuable questions to know the answers to if you're trying to understand the social impacts of more risky behavior, and it would be great if we did know these. But the progress epidemiologists have put into estimating R0 is not most of what you'd draw on in trying to get better answers.

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

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: Shapley Value Attribution in Chain of Thought, published by leogao on April 14, 2023 on The AI Alignment Forum.

TL;DR: Language models sometimes seem to ignore parts of the chain of thought, and larger models appear to do this more often. Shapley value attribution is a possible approach to get a more detailed picture of the information flow within the chain of thought, though it has its limitations.

Project status: The analysis is not as rigorous as I would prefer, but I'm going to be working on other directions for the foreseeable future, so I'm posting what I already have in case it's useful to others.

Thanks to Jacob Hilton, Giambattista Parascandolo, Tamera Lanham, Ethan Perez, and Jason Wei for discussion.

Motivation

Chain of thought (CoT) has been proposed as a method for language model interpretability (see Externalized Reasoning Oversight, Visible Thoughts). One crucial requirement for interpretability methods is that they should accurately reflect the cognition inside the model. However, by default there is nothing forcing the CoT to actually correspond to the model’s cognition, and there may exist theoretical limitations to doing so in general.

Because it is plausible that the first AGI systems bear resemblance to current LMs with more sophisticated CoT and CoT-like techniques, it is valuable to study its properties, and to understand and address its limitations.

Related work

Shapley values have been used very broadly in ML for feature importance and attribution (Cohen et al, 2007; Štrumbelj and Kononenko, 2014; Owen and Prieur, 2016; Lundberg and Lee, 2017; Sundararajan and Najmi, 2020). Jain and Wallace (2019) argue that attention maps can be misleading as attribution, motivating better attribution for information flow in LMs. Kumar et al. (2020) highlight some areas where Shapley value based attribution falls short for some interpretability use cases.

Madaan and Yazdanbakhsh (2022) consider a similar method of selectively ablating tokens as a method of deducing what information the model is dependent on. Wang et al. (2022) find that prompting with incorrect CoT has surprisingly minor impact on performance.

Effect of Interventions

We use a method similar to Kojima et al. (2022) on GSM8K (Cobbe et al., 2021) with GPT-4 to first generate a chain of thought and evaluate the answer, and then for all chains of thought that result in a correct answer we perform an intervention as follows: we choose a random numerical value found in the CoT, and replace it with a random number in a +/-3 range about the original. We then discard the remainder of the CoT and regenerate it. If the LM is following strictly the CoT described, this intervention should almost always result in an incorrect answer, the same way one would if they made a mistake in one calculation and propagated the error through to the answer (with occasional rare cases where the new value happens to also result in the correct answer, though from qualitative inspection this is very rarely the case).

Some cherrypicked examples (red = intervention, blue = correct continuations that are seemingly non-sequiturs):

We test how frequently this occurs in several different settings (n=100):

SettingAccuracy (w/ CoT)P(error not propagated | original correct)GPT4, zero shot0.880.68GPT4 base, 2-shot0.730.63GPT3.5, zero-shot0.430.33

Interestingly, if we condition on the CoT answer being correct and the single forward pass answer being incorrect (i.e the LM could only solve the problem with the CoT), the intervened accuracy for GPT-4 is still 0.65.

Shapley value attribution

We would like to get more granular information about the causal structure (i.e which tokens cause which other tokens). One thing we could do is look at how an intervention at each token affects the logprob of each other token. However, one major prob...

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: Financial Times: We must slow down the race to God-like AI, published by trevor on April 13, 2023 on LessWrong.

The article itself is paywalled, so here you go (you can often circumvent paywalls by typing the article's name into google search's news tab). This isn't actually bad news; it means that the FT will reach a smaller and more elite audience (and hopefully with above average quant skills), rather than the NYT which simply maximizes the number of views by being the most popular outlet.

Notably, this story not only gives AI alignment positive coverage, but it is also extremely close to the front page on Financial Times's website, and with a very striking image to boot (possibly the most important factor). Of course, it's still social media spread that largely decide the fate of these articles, not the news outlet's website, so we can't know for sure how helpful it is.

This article was written by Ian Hogarth, "yesterday" according to the page. It's important to bear in mind that being published in a major news outlet is a stamp of approval that most people take extremely seriously when entering a field for the first time, and that factor is a much bigger deal than whether the author got everything right on their first try.

As a side note, Raemon has recently recommended these sources as a good way to explain AI risk to someone for the first time:

Superintelligence FAQ (very accessible to layfolk)

The Alignment Problem from a Deep Learning Perspective (written with ML researchers in mind)

(I'll work on compiling more of these soon)

On a cold evening in February I attended a dinner party at the home of an artificial intelligence researcher in London, along with a small group of experts in the field. He lives in a penthouse apartment at the top of a modern tower block, with floor-to-ceiling windows overlooking the city’s skyscrapers and a railway terminus from the 19th century. Despite the prime location, the host lives simply, and the flat is somewhat austere.

During dinner, the group discussed significant new breakthroughs, such as OpenAI’s ChatGPT and DeepMind’s Gato, and the rate at which billions of dollars have recently poured into AI. I asked one of the guests who has made important contributions to the industry the question that often comes up at this type of gathering: how far away are we from “artificial general intelligence”? AGI can be defined in many ways but usually refers to a computer system capable of generating new scientific knowledge and performing any task that humans can.

Most experts view the arrival of AGI as a historical and technological turning point, akin to the splitting of the atom or the invention of the printing press. The important question has always been how far away in the future this development might be. The AI researcher did not have to consider it for long. “It’s possible from now onwards,” he replied.

This is not a universal view. Estimates range from a decade to half a century or more. What is certain is that creating AGI is the explicit aim of the leading AI companies, and they are moving towards it far more swiftly than anyone expected. As everyone at the dinner understood, this development would bring significant risks for the future of the human race. “If you think we could be close to something potentially so dangerous,” I said to the researcher, “shouldn’t you warn people about what’s happening?” He was clearly grappling with the responsibility he faced but, like many in the field, seemed pulled along by the rapidity of progress.

When I got home, I thought about my four-year-old who would wake up in a few hours. As I considered the world he might grow up in, I gradually shifted from shock to anger. It felt deeply wrong that consequential decisions potentially affecting every life on Earth could be made by a small grou...

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: A freshman year during the AI midgame: my approach to the next year, published by Buck on April 14, 2023 on The Effective Altruism Forum.

I recently spent some time reflecting on my career and my life, for a few reasons:

It was my 29th birthday, an occasion which felt like a particularly natural time to think through what I wanted to accomplish over the course of the next year .

It seems like AI progress is heating up.

It felt like a good time to reflect on how Redwood has been going, because we’ve been having conversations with funders about getting more funding.

I wanted to have a better answer to these questions:

What’s the default trajectory that I should plan for my career to follow? And what does this imply for what I should be doing right now?

How much urgency should I feel in my life?

How hard should I work?

How much should I be trying to do the most valuable-seeming thing, vs engaging in more playful exploration and learning?

In summary:

For the purposes of planning my life, I'm going to act as if there are four years before AGI development progresses enough that I should substantially change what I'm doing with my time, and then there are three years after that before AI has transformed the world unrecognizably.

I'm going to treat this phase of my career with the urgency of a college freshman looking at their undergrad degree--every month is 2% of their degree, which is a nontrivial fraction, but they should also feel like they have a substantial amount of space to grow and explore.

The AI midgame

I want to split the AI timeline into the following categories.

The early game, during which interest in AI is not mainstream. I think this ended within the last year

The midgame, during which interest in AI is mainstream but before AGI is imminent. During the midgame:

The AI companies are building AIs that they don’t expect will be transformative.

The alignment work we do is largely practice for alignment work later, rather than an attempt to build AIs that we can get useful cognitive labor from without them staging coups.

For the purpose of planning my life, I’m going to imagine this as lasting four more years. This is shorter than my median estimate of how long this phase will actually last.

The endgame, during which AI companies conceive of themselves as actively building models that will imminently be transformative, and that pose existential takeover risk.

During the endgame, I think that we shouldn’t count on having time to develop fundamentally new alignment insights or techniques (except maybe if AIs do most of the work? Idt we should count on this); we should be planning to mostly just execute on alignment techniques that involve ingredients that seem immediately applicable.

For the purpose of planning my life, I’m going to imagine this as lasting three years. This is about as long as I expect this phase to actually take.

I think this division matters because several aspects of my current work seem like they’re optimized for midgame, and I should plausibly do something very differently in the endgame. Features of my current life that should plausibly change in the endgame:

I'm doing blue-sky alignment research into novel alignment techniques–during the endgame, it might be too late to do this.

I'm working at an independent alignment org and not interacting with labs that much. During the endgame, I probably either want to be working at a lab or doing something else that involves interacting with labs a lot. (I feel pretty uncertain about whether Redwood should dissolve during the AI endgame.)

I spend a lot of my time constructing alignment cases that I think analogous to difficulties that we expect to face later. During the endgame, you probably have access to the strategy “observe/construct alignment cases that are obviously scary in the models you have”...

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: A freshman year during the AI midgame: my approach to the next year, published by Buck on April 14, 2023 on LessWrong.

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