The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - AI Neorealism: a threat model & success criterion for existential safety by davidad (David A. Dalrymple)
    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: AI Neorealism: a threat model & success criterion for existential safety, published by davidad (David A. Dalrymple) on December 15, 2022 on The AI Alignment Forum.
    Threat Model
    There are many ways for AI systems to cause a catastrophe from which Earth-originating life could never recover. All of the following seem plausible to me:
    Misuse: An AI system could help a human or group of humans to destroy or to permanently take over (and lock their values into) the world. The AI could be:
    An oracle AI (e.g. a question-answering LLM)
    An LLM simulating an intent-aligned agent and taking real-world actions via APIs
    An intent-aligned RL agent
    An interaction of multiple systems
    Power-Seeking: An AI system could destroy or permanently take over the world on its own account, by leveraging advanced instruments of force projection. The AI could be:
    An LLM simulating a misaligned agent
    "Specification gaming": An RL agent that is aligned to a formal objective and Goodharts to catastrophe
    "Goal misgeneralization": A surprise mesa-optimiser (most likely in model-free RL, but could conceivably arise through evolutionary processes in any iterative algorithm which has or learns sufficiently reality-like structure)
    An interaction of multiple systems, participating in coordination mechanisms that exclude humans
    Economic Squeeze: an AI system could acquire nearly all means of production through a gradual process of individually innocent economic transactions, thereby squeezing humanity out of resource allocation decisions and removing most human influence over the future.
    This would most likely be an "interaction of multiple systems".
    A single RL agent, or a unipolar tree of agents, might also do this, especially if they are successfully aligned to avoid use of force against humans.
    Superpersuasion: an AI system could generate stimuli which reliably cause humans to adopt its arbitrary goals. The AI could be:
    An LLM merely extrapolating from persuasive human text
    An RL agent trained on human approval
    A surprise mesa-optimiser
    Some mixture of the above
    Many AIs, collectively shaping a new human culture with an alien ideology
    Security Dilemma: If AI-enabled technological advancements turn out to be offence-dominant, and if partial alignment success leads AIs to be unable to make credible commitments to each other (e.g. due to corrigibility), the equilibrium strategy for AI-enabled militaries may involve high-risk preemptive strikes and increasingly escalated retaliation to a point of existential catastrophe.
    This would almost surely be a multipolar failure mode.
    But, instead of trying to enumerate all possible failure modes and then trying to shape incentives to make them less likely to come up, I typically use a quasi-worst-case assumption in which I assume that, perhaps as a matter of bad luck with random initialisation,
    Unlike a typical understanding of a "worst-case assumption," the last clause leaves open the possibility of hiding concrete facts about our world from an arbitrarily powerful model, and the framing in terms of functions highlights an ontology of AI that respects extensional equivalence, where imputations of "deceptive mesa-optimisers hiding inside" are discarded in favour of "capable but misaligned outputs on out-of-distribution inputs".
    On the other hand, unlike a typical "prosaic" threat model, in the neorealist threat model one does not rely on empirical facts about the inductive biases of the kind of network architectures that are practically successful. A realist justification for this is that there may be a phase transition as architectures scale up which drastically changes both their capabilities profile and this kind of inductive bias (vaguely analogous to the evolution of cultural knowledge-transfer within biological life).
    One can make progress with this assumption...
    6 min
  • EA - Saving Lives vs Creating Lives by Richard Y Chappell
    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: Saving Lives vs Creating Lives, published by Richard Y Chappell on December 15, 2022 on The Effective Altruism Forum.
    tl;dr: Total utilitarianism treats saving lives and creating new lives as equivalent (all else equal). This seems wrong: funding fertility is not an adequate substitute for bednets. We can avoid this result by giving separate weight to both person-directed and undirected (or "impersonal") reasons. We have weak impersonal reasons to bring an extra life into existence, while we have both impersonal and person-directed reasons to aid an existing individual. This commonsense alternative to totalism still entails longtermism, as zillions of weak impersonal reasons to bring new lives into existence can add up to overwhelmingly strong reasons to prevent human extinction.
    Killing vs Failing to Create
    I think the strongest objection to total utilitarianism is that it risks collapsing the theoretical distinction between killing and failing to create. (Of course, there would still be good practical reasons to maintain such a distinction in practice; but I think there’s a principled distinction here that our theories ought to accommodate.) While I think it’s straightforwardly good to bring more awesome lives into existence, and so failing to create an awesome life constitutes a missed opportunity for doing good, premature death is not just a “missed opportunity” for a good future, it’s harmful in a way that should especially concern us.
    For example, we clearly have much stronger moral reasons to save the life of a young child (e.g. by funding anti-malarial bednets) than to simply cause an extra child to exist (e.g. by funding fertility treatments or incentivizing procreation). If totalism can’t accommodate this moral datum, that would seem a serious problem for the view.
    How can we best accommodate this datum? I think there may be two distinct intuitions in the vicinity that I’d want to accommodate:
    (1) Something about the intrinsic badness of (undesired) death.(2) Counting both person-directed and undirected (“impersonal”) moral reasons.
    The Intrinsic Harm of Death
    Most of the harm of death is comparative: not bad in itself, but worse than the alternative of living on. Importantly, we only have reason to avoid comparative harms in ways that secure the better alternative. To see this, suppose that if you save a child’s life, they’ll live two more decades and then die from an illness that robs them of five decades more life. That latter death is then really bad for them. Does it follow that you shouldn’t save the child’s life after all (since it exposes them to a more harmful death later)? Of course not. The later death is worse compared to living the five decades extra, but letting them die now would do them even less good, no matter that the early death — in depriving them of just two decades of life — is not “as bad” (comparatively speaking) as the later death would be (in a different context with a different point of comparison).
    So we should not aim to minimize comparative harms of this sort: that would lead us badly astray. But it’s a tricky question whether the harm of death is purely comparative. In ‘Value Receptacles’ (2015, p. 323), I argued that it plausibly is not:
    Besides preventing the creation of future goods, death is also positively disvaluable insofar as it involves the interruption and thwarting of important life plans, projects, and goals. If such thwarting has sufficient disvalue, it could well outweigh the slight increase in hedonic value obtained in the replacement scenario [where one person is “struck down in the prime of life and replaced with a marginally happier substitute”].
    Thwarted goals and projects may make death positively bad to some extent. But the extent must be limited. However tragic it is to die in one’s teens (say), I don’t ...
    9 min
  • EA - I went to the Progress Summit. Here’s What I Learned. by Nick Corvino
    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: I went to the Progress Summit. Here’s What I Learned., published by Nick Corvino on December 14, 2022 on The Effective Altruism Forum.
    I attended the Progress Summit in Hollywood yesterday, hosted by The Atlantic. Progress studies and EA have overlap, so I thought it would be useful to give my thoughts on the event. In general, the main difference I perceived was people attending not because they wanted to maximize their positive impact but rather because they were intellectually interested in socially responsible progress. And some other stuff.
    Right after I left, I told my friend it felt like a “grown up version of an EA conference.” I’m 22, and was probably the youngest person there. Everything felt more professional (cocktails, food, outfits, etc.) and the operations seemed smoother than any EA event I’ve been to.
    The facilitators were all Atlantic writers, such as Ross Anderson and Derek Thompson, and they were significantly more eloquent and better at holding people’s attention than any EA event I’ve been to. I could definitely tell the difference in their training. That being said, the talks felt fluffy and often skirted around intellectual issues for the sake of a smooth conversation.
    Networking was less direct. More small talk, less intensity.
    The event was catered towards investors/venture capitalists. Speakers were trying to make their product sound appealing so investors will fund them, which I thought was slightly bad for epistemics, often ignoring the risks of their products (e.g. AI slaughter bots).
    In general, the majority of the attendees seemed bullish on “the progress of technology,” and didn’t touch much on the potential risks of things like AGI or biorisk. If they did address the risks, it was invariably in relation to (1) the economy, (2) climate change, or (3) war. Of the people I spoke with, <20% had heard of misalignment or existential risk. I didn’t get the impression that anyone at the event didn’t take existential risk seriously. Rather, it felt like they had not heard about it in the progress studies ecosystem.
    Overall, I think the Progress studies community seems decently aligned with what EAs care about, and could become more-so in the coming years. The event had decent epistemics and was less intimidating than an EA conference. I think many people who feel that EA is too intense, cares too much about longtermism, or uses too much jargon could find progress studies as a suitable alternative. If the movement known as EA dissolved (God forbid) I think progress studies could absorb many of the folks.
    Notable events:
    “How mRNA Technology Can Save the World”
    Most of the people I talked to here had never considered the risks posed by biotechnology (beyond class inequality stuff)
    “Drones and AI: The Future of Military Technology”
    The concern most people had was causing a war (which seems good), but because much of what Brian Schimpf talked about was technology used for deterrence, most people then seemed bullish on the positive impacts of this technology (I was not).
    “How Artificial Intelligence Can Revolutionize Creativity”
    I didn’t go to this one, but I heard from someone who did that they talked about GPT-3 and Dall-e positively, and didn't mention the potential risks posed by capabilities advancements.
    “The Long View”
    Didn’t go, unfortunately.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min
  • LW - Discovering Latent Knowledge in Language Models Without Supervision by Xodarap
    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: Discovering Latent Knowledge in Language Models Without Supervision, published by Xodarap on December 14, 2022 on LessWrong.
    Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4% on average.
    We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.
    Some discussion on Twitter including Eliezer calling it "Very dignified work!"
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - Predicting GPU performance by Marius Hobbhahn
    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: Predicting GPU performance, published by Marius Hobbhahn on December 14, 2022 on LessWrong.
    We develop a simple model that predicts progress in the performance of field-effect transistor-based GPUs under the assumption that transistors can no longer miniaturize after scaling down to roughly the size of a single silicon atom. We construct a composite model from a performance model (a model of how GPU performance relates to the features of that GPU), and a feature model (a model of how GPU features change over time given the constraints imposed by the physical limits of miniaturization), each of which are fit on a dataset of 1948 GPUs released between 2006 and 2021. We find that almost all progress can be explained by two variables: transistor size and the number of cores. Using estimates of the physical limits informed by the relevant literature, our model predicts that GPU progress will stop roughly between 2027 and 2035, due to decreases in transistor size.
    In the limit, we can expect that current field-effect transistor-based GPUs, without any paradigm-altering technological advances, will be able to achieve a peak theoretical performance of 1e14 and 1e15 FLOP/s in single-precision performance.
    While there are many other drivers of GPU performance and efficiency improvements (such as memory optimization, improved utilization, and so on), decreasing the size of transistors has historically been a great, and arguably the dominant, driver of GPU performance improvements. Our work therefore strongly suggests that it will become significantly harder to achieve GPU performance improvements around the mid-2030s within the longstanding field-effect transistor-based paradigm.
    This post is just the executive summary, you can find the full report on the Epoch website.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - The US Attorney for the SDNY asked that FTX money be returned to victims. What are the moral and legal consequences to EA? by Fermi–Dirac Distribution
    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 US Attorney for the SDNY asked that FTX money be returned to victims. What are the moral and legal consequences to EA?, published by Fermi–Dirac Distribution on December 14, 2022 on The Effective Altruism Forum.
    In a December 13, 2022 press conference announcing the indictment of Sam Bankman-Fried, Damian Williams, the United States Attorney for the Southern District of New York, said the following:
    To any person, entity or political campaign that has received stolen customer money, we ask that you work with us to return that money to the innocent victims.
    In the same press conference, he alleges that FTX has been stealing customer money since 2019:
    First, we charge that from 2019 until earlier this year, Bankman-Fried and his co-conspirators stole billions of dollars from FTX customers.
    He also alleged the following about SBF's political contributions:
    Those contributions were disguised to look like they were coming from wealthy co-conspirators, when in fact, the contributions were funded by Alameda Research with stolen customer money.
    This might be relevant to EA orgs, since a lot of SBF's political contributions seem to have been made around the same time as his EA donations (i.e. this year), and so likely came from the same source.
    I made a comment pointing out some of this information yesterday, but so far that this is getting less attention than I expected given the seriousness of the matter, hence this top-level post.
    What are the moral and legal consequences of the information in the indictment, and of Williams's request, to EA orgs and individuals that have received money from FTX?
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - It Takes Two Paracetamol? by Eli
    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: It Takes Two Paracetamol?, published by Eli on December 13, 2022 on LessWrong.
    Note: Depending on where you live paracetamol is also known as acetaminophen, APAP and Tylenol.
    I have had a version of the following conversation a surprising number of times:
    "I have a headache."
    "Aah, that sucks. Did you take some paracetamol?"
    "Yes, but it still hurts."
    "How many did you take?"
    The fact that I even ask this question is because I have been noticing a trend.
    "One."
    "You should take two, it makes a big difference."
    "Nah, I'm fine with one."
    "But you still have a headache."
    "Yeah, but it doesn't make much of a difference. Paracetamol just doesn't work that well for me."
    Everyone is free to choose how many paracetamol they take, but I've always had this hunch that there is a big difference between taking one or two paracetamol. Where taking one has almost no effect on my headaches, but taking two makes a big difference. Effectively I'm making two claims:
    1000mg paracetamol works better than 500mg paracetamol.
    500mg paracetamol doesn't work much better than placebo.
    Today I chose to finally resolve this mystery and laugh in the face of my friends! Or I will admit in shame that I was wrong...
    Claim 2: 500mg vs placebo
    There is a Cochrane Review from 2016 that tries to answer exactly my question. The review looks at pain relief with paracetamol for tension-type headaches, the most common form of headaches. They reference two studies that have compared 500mg/650mg with placebo that also fit their inclusion criteria.
    Steiner 2003, which comes to the following conclusion:
    They didn't find a significant result when comparing 500mg paracetamol and placebo! But they did find a significant result when comparing 1000mg paracetamol and placebo.
    Dahlöf 1996 is less conclusive. This study only contains 29 participants, where only 14 of them took paracetamol at a dose of 500mg.
    In this study both 500mg and 1000mg performed equally well, but with the low participant numbers it is hard to draw any conclusions.
    The Cochraine Review itself comes to the conclusion that they couldn't find a significant difference between 500mg paracetamol and placebo, but that the quality of evidence is low.
    Claim 1: 1000mg vs 500mg
    The same Cochrane Review concludes that a 1000mg paracetamol outperforms placebo. There is high quality evidence for this being the case when looking at
    Pain-free at two hours: the proportion of attacks/participants who were pain-free at two hours with paracetamol.
    Pain-free or mild pain at two hours: the proportion of attacks/participants who were pain-free or had only mild pain at two hours with paracetamol.
    There is moderate quality evidence for this being the case when looking at
    Pain-free at one hour: The proportion of attacks/participants who were pain-free or had only mild pain at one hour with paracetamol.
    Pain-free at four hours: The proportion of attacks/participants who were pain-free at four hours with paracetamol.
    Although it might seem that I'm moving towards a clear win for my case, I'm afraid I have to highlight one interesting point which is exemplified by the work of Steiner 2003.
    Their data leads to three seemingly incompatible statements:
    They find a significant difference between 1000mg paracetamol and placebo.
    They don't find a significant difference between 500mg paracetamol and placebo.
    They don't find a significant difference between 500mg and 1000mg paracetamol.
    It is funny that these three statements can all true at the same time. At this point in time we don't have the data to show that a 1000mg outperforms 500mg.
    Conclusion
    To be honest, I didn't expect the results I found. In hindsight it is obvious that these statements can all be true at the same time, but when starting out I hadn't considered it an option.
    I think I can safely say the following going...
    6 min
  • EA - EA's Achievements in 2022 by ElliotJDavies
    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: EA's Achievements in 2022, published by ElliotJDavies on December 14, 2022 on The Effective Altruism Forum.
    As the year is nearing its end, it would be good to crowd source achievements made by Effective Altruists this year, and collect them into a single thread.
    What got done in 2022, to to make the world safer or better, for animals, people or future beings:
    How many mosquito nets were distributed? What organisations got founded? What papers got published? What incubation grants where made? How have organisations scaled up? What laws got passed?
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • AF - Extracting and Evaluating Causal Direction in LLMs' Activations by Fabien Roger
    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: Extracting and Evaluating Causal Direction in LLMs' Activations, published by Fabien Roger on December 14, 2022 on The AI Alignment Forum.
    This post was written by Fabien at SaferAI. Simeon has prompted Fabien in relevant directions and has provided valuable feedback.
    Thanks to Jean-Stanislas Denain, Alexandre Variengien, Charbel-Raphael Segerie, and Nicole Nohemi for providing helpful feedback on early experiments and drafts of this post.
    In this post
    I describe a method to evaluate that a specific direction encodes information relative to a specific concept,
    I use it to evaluate how good directions found through probing techniques are,
    I present a way of using causal intervention to find directions that maximally encode information relative to a specific concept, which I call Causal Direction Extraction (CDE),
    I present some preliminary results about the directions found through Causal Direction Extraction
    One of the main goals of this post is to get suggestions on further experiments to run. I think the current data I gathered is not strong enough to prove or disprove claims like “CDE finds a direction, which mostly captures the concept of gender in GPT-J”, but further experiments could.
    The code for all experiments is available here.
    How to Evaluate How Much a Direction Encodes a Concept
    What It Means for a Direction to Encode a Concept
    Let’s say you have a neural network N. Let N<=L be the network up to a given layer L, and N>L be the network after that, such that for all input x, N>L(N<=L(x))=N(x).
    The claim that a (normalized) direction d in the space of activations A encodes a target concept between layers L and L+1 can be understood as follows:
    Activations after layer L can be cleanly cut into two pieces:
    c(x), the projection of N<=L(x) along d, which only encodes information relative to the target concept
    rest(x), the projection of N<=L(x) orthogonal to d, which only encodes all information not related to the target concept
    Another way to put it is that there exists a function c:XRd which only depends on properties of x related to the concept and a function rest:X(Rd)⊥ which depends on every other characteristic of the input, such that for all input x, N(x)=N>L(c(x)+rest(x)).
    Using Activation Patching to Evaluate How Much a Direction Encodes a Concept
    With this definition of what it means for a direction to encode a concept, you can use activation patching (introduced by Redwood’s Indirect Object Identification paper, Wang 2022) to quantify how well this hypothesis works. The setup here is simpler than in the original paper, since we patch the activations directly in the residual stream (between two layers instead of at the output of one attention head).
    Let’s take the concept of gender, and let’s run the network on the two following sentences: xA = “but her favorite color is”, xB = “but his favorite color is”, which only differ by the concept of gender (we’ll discuss this claim later). If the hypothesis is correct, then rest(xA)=rest(xB), which means we can patch activations after layer L:
    N(xA)=N>L(c(xA)+rest(xB))
    Let’s define Nd(xA,xB)=N>L(c(xA)+rest(xB)). The first argument is the one we should determine the output of the function if the hypothesis is correct, the second argument is a “distraction”.
    How much this holds can be quantified in different ways:
    Measure the KL-Divergence between the relevant output N(xA) and the mixed output Nd(xA,xB)
    Measure the probability of a specific token in the mixed output Nd(xA,xB) (“pink” for example) and see how much closer it is to the relevant output N(xA) than the irrelevant N(xB). More precisely, we can define the success rate as Nd(xA,xB)pink−N(xB)pinkN(xA)pink−N(xB)pink. A success rate of 1 means that the probability of “ pink” in the mixed output is the probability given to pink by the model on the relevant i...
    19 min
  • EA - What Rethink Priorities General Longtermism Team Did in 2022, and Updates in Light of the Current Situation by Linch
    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: What Rethink Priorities General Longtermism Team Did in 2022, and Updates in Light of the Current Situation, published by Linch on December 14, 2022 on The Effective Altruism Forum.
    Summary
    Rethink Priorities’ General Longtermism team, led by me, has existed for just under a year. In this post, I summarize our work so far.
    Our initial theory of change centered around:
    Primarily, facilitating the creation of faster and better longtermist “megaprojects” (though in practice, we focused more on somewhat scalable longtermist projects).
    Secondarily, improving strategy clarity about which “intermediate goals” longtermists should pursue (though in practice, we focused more opportunistically on miscellaneous high-impact research questions). (more)
    We had ~5 Full-time Equivalents (FTE) on average (a total of 54.5 FTE months of research work up until the end of November) and spent ~$721,000. (more)
    (Shareable) Outputs and outcomes of our work this year include:
    I encouraged the creation of and supported the Rethink Priorities’ Special Projects team, which provides fiscal sponsorship to external entrepreneurial projects.
    Marie and Renan (based on prior work by Renan, Max, and me) made a simple model for prioritizing between longtermist projects and identifying ones that seemed especially promising to research further. (more)
    Our fellows and research assistant (Emma, Jam, Joe, Max, Marie) completed 13 research “speedruns” (~10h shallow dives) into specific longtermist projects. (more)
    Our fellows and research assistant (Emma, Jam, Joe, Max, Marie) completed further research on several longtermist projects, including air sterilization techniques, whistleblowing, the AI safety recruiting pipeline, and infrastructure to support independent researchers. (more)
    Renan cofounded and ran Condor Camp, a project that aims to find and engage world-class talent in Brazil for longtermist causes while also field-building longtermism and supporting EA community building in the country. (more)
    Ben cofounded and ran Pathfinder, a project to help mid-career professionals to find the highest impact work (more)
    I initiated a founder search for multiple promising projects, including:
    Shelters and other civilizational resilience work (resulting in recommending grants to Tereza Fildrova, who helped organize the SHELTER weekend, and Ulrik Horn, who is exploiting his fit for work in this area). (more)
    An early warning forecasting center (resulting in working with Alex D to explore founding a project in this area). (more)
    Ben researched nanotechnology strategy and made a database of resources relevant to this area. (more)
    Separately from my work at Rethink Priorities, I was a guest fund manager for EA Funds’ Long-Term Future Fund, and a regranter for the Future Fund. (more)
    The recent changes to the EA funding situation have significantly affected our team’s strategy, in that megaprojects now seem less relevant, and in that new research questions might have become especially important in light of the FTX crash. (more) However, I still think it’s very plausible that we continue to focus on entrepreneurial longtermist projects as our main research area. We’re currently in the process of reorienting and setting our strategy for 2023. (more)
    You can help our team by contributing ideas for highly impactful research projects, funding us, expressing interest in working with us, and giving feedback on the work and plans outlined in this post. (more)
    Preamble
    I’ve led the new General Longtermism team at Rethink Priorities for slightly under a year. Recent changes in the EA funding landscape have had a large impact on our work. I’ve decided that now is a good time to write a detailed summary of our work so far, as well as how recent events have affected our work.
    The primary purpose of this article is to be informa...
    34 min