The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Have there been any detailed, cross-cultural surveys into global moral priorities? by Amber Dawn
    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: Have there been any detailed, cross-cultural surveys into global moral priorities?, published by Amber Dawn on February 6, 2023 on The Effective Altruism Forum.
    Has there been any research done (either within or outside of EA) about most people’s moral priorities, and/or about the priorities of recipients of philanthropy? I’m thinking of things like, surveys of large groups of people across many cultures which asked them ‘is it more important to be healthy, or wealthy, or have more choices, or prevent risks that might hurt your grandchildren?’
    What motivates this question is something like: there’s been a lot of talk about democratizing EA. But even if more EA community members had input into funding decisions, that’s still not really democratic. I want to know: what does the average person worldwide think that wealthy philanthropists should do with their money?
    GiveWell commissioned some research into the preferences of people in some low income communities, similar to the beneficiaries of many of their top charities. However, they only asked about whether they valued saving the lives of younger vs older people, and how much they valued saving years of life vs increasing income. It would be interesting to read more holistic surveys that asked about other things that people might value, including things that charities might not straightforwardly be able to provide (like more political participation, or less oppression). (You could use as a basis, for example, the capability approach, or Spencer Greenberg’s work on intrinsic values.)
    This might be useful for longtermists as well as those who focus on global health and poverty in the nearer term. You could ask people how much they value risk mitigation vs increases in wellbeing, for example; or you could use people’s answers to try to shape a future that fits more people’s values.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Project Idea: Lots of Cause-area-specific Online Unconferences by Linda Linsefors
    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: Project Idea: Lots of Cause-area-specific Online Unconferences, published by Linda Linsefors on February 6, 2023 on The Effective Altruism Forum.
    Here's a project idea I have that's been laying around for some time. I think this would be high impact, but I need help to make it happen. It don't have to be this exact setup (which I describe below) as long as the output are more online unconferences for the EA community.
    I don't do grant applications anymore, due to past event that I'm still emotionally healing from. However if you want to write and submit an application on my behalf, that would be very welcome.
    Project description can also be found here: Lots of EA Online Unconferences - Google Docs
    Total budget: $40k - 85k.
    Summary: We need more online events - from my experience running TAISU and AISU, there are many smart people who can’t access existing events because of geographical constraints. I expect this to be true of other cause areas too. I want to hire one person for one year and teach them how to run an online unconference. They’ll run 3-4 events on different EA topics of their choice. If things go well, they secure more funding before the end of the year, and keep going.
    Motivation
    Online unconferences are low cost and relatively low effort to run, and are also really great. I think this type of event has the highest cost to value ratio out of any event type that I know of.
    Why more events?
    EA is growing faster than the infrastructure can keep up with. I’m mostly familiar with AI Safety but I expect the situation to be similar in other areas. Every AI Safety event I know about gets a massive number of applications. There is definitely room for more events.
    Why cause area specific?
    Generic EA events are great, but when you’ve picked what problem to work on, it’s much more valuable to connect with people who are interested in the same cause area as you. Those are the people you can have high context discussion with. Those are the people you can team up with to get work done.
    I count “EA Meta” (e.g. ops, events etc) as a cause area for the purpose of this project, because it seems valuable for people who work in this area to meet up and exchange ideas, and find common solutions to common problems.
    Why unconferences?
    Unconferences are great in many ways, but they are especially great in a rapidly growing movement, where the people at the top, can’t keep up with or even keep track of the need for infrastructure. Unconferences are self organised around the needs of interests of the participants. This makes them more adaptive but also lower effort to run relatively to other types of events.
    I’m not saying that unconferences are a one-size-fits-all type of event. There is definitely room for more specialised events too. But everything equal, I think EA would be better off with more unconferences.
    Why online?
    Online and offline have different pros and cons.
    One big advantage with online events is that they are much more accessible. This is important! Even if we have lots of in person events, it’s worth having online events too. Accessibility helps in two ways. Firstly there are people who can’t attend otherwise for various reasons. Just paying for their travel may not help since the barrier might be time or visa. Secondly, it lowers the cost (in time and money) to attend for everyone. People who would not be willing to pay the time cost for an entire in-person event often choose to join in for just a few sessions of an online event, benefiting both them and everyone else. Allowing people to pick and choose among sessions, rather than commit to a full event, opens up for more specialised sessions.
    The other main advantage of online events is that they are almost infinitely scalable, much less work to organise, and much lower organising costs. A single person can d...
    8 min
  • LW - Here's Why I'm Hesitant To Respond In More Depth by DirectedEvolution
    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: Here's Why I'm Hesitant To Respond In More Depth, published by DirectedEvolution on February 6, 2023 on LessWrong.
    Hello and welcome!
    Thank you in advance for taking the time to read this post.
    I mean that sincerely.
    I'm replying with this post because something about your comment seems to have triggered my Hesitation Reaction.
    Explaining what that means requires too much nuance to write it out every time. But I think I miss out on a lot of good conversations because of this Hesitation Reaction, which is essentially one of self-protection.
    So I created this post to explain my Hestitation Reaction in case you find it helpful in engaging with me differently, hopefully in ways we both prefer, in our original thread of conversation.
    My Hesitation Reaction, in brief
    To start, these are my own emotions, which are my responsibility, not yours. The point of articulating them isn't to blame you for them!!!
    The point is that your comment may have had the (non-blameworthy) side effect of bringing these emotions up in me, and that it will for practical purposes be difficult for me to continue having the interesting discussion we might both like to have as a result, without some sort of hopefully minor adjustment to get us back on track.
    My hope is that exposing these facts about my emotions in this level of depth will facilitate us mutually and collaboratively figuring out some productive next steps. That's not on you to sort out on your own. I'm up for my half of the bargain too.
    On an emotional level, my Hesitation Reaction is often a combination of:
    Kind of a chilled emotional feeling, a sort of pulling back. It's the reaction you have when somebody is noticeably less warm or interested or generous about something you're excited about than you had anticipated or hoped for.
    Anxiety that I'm being either trolled or drawn into a contest over status, expertise or authority.
    Frustration that my goals in posting feel like they're not being recognized.
    A feeling as though I'm being singled out as less worthy, implicitly excluded or marginalized somehow.
    Here is another picture of an elephant seal to lighten the mood:
    Making me Hesitate is not a crime
    Look, I really get it! It's tough to judge how much effort to put in, what needs to be said, who the heck the person you're talking to is and what their motivations are. Sometimes, we all leave comments that make the other person Hesitate, when we didn't mean to. And that's not necessarily a sign that we did anything wrong at all.
    In fact, the main reason I wrote this is because I suspect I'm leaving a lot of value on the table by having low-grade negative reactions to Hestitation-provoking comments that actually could have been the start of a really interesting conversation.
    So please don't take the fact that I linked you to this post as a criticism of you, or of your comment. This is the most important thing I want you to take away from this post. Instead, I linked you here because, through no fault of your own, your comment provoked my Hestitation Reaction. Assuming good faith, I want to highlight some reasons why that might have been, and invite you to take whatever steps you like to address that reaction, very much including nothing at all!
    I hope you and anybody else who sees the link will perceive this post as my best attempt to expose a part of my perhaps neurotic psychology for your direct inspection, not as a criticism or as a status move.
    Common drivers of my Hesitation Reaction.
    Note that these are all my subjective perceptions, which don't necessarily reflect your intentions, how other people would see the situation, or the facts of the matter. Almost certainly only a small subset of these apply in this specific case.
    Overemphasizing a genuine problem in my post that I don't consider to be among the highest-priority problem...
    6 min
  • EA - Moving community discussion to a separate tab (a test we might run) by Lizka
    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: Moving community discussion to a separate tab (a test we might run), published by Lizka on February 6, 2023 on The Effective Altruism Forum.
    TL;DR: we (the Forum team) might make a pretty significant change on Frontpage for a month as a test. We’re sharing the outline of the proposal to get feedback from the community. The change we’ll test is hiding “Community” posts from the Frontpage for everybody by default, and moving those posts to a prominent tab.
    Where we are right now:
    We’re probably going to run the test, although we might not depending on the feedback we get as a result of this post and we might significantly change our approach. After the test, it’s pretty unclear to us whether we’ll keep the change — and we might want to extend the test period — but we think that we will probably want to keep something like the change after the month goes by.
    We will review the feedback we get, and if we decide to go for it, we’ll likely start the test within a week or two.
    Summary of the change we’d test: moving community discussion to a separate tab
    For the test, we’d move all “Community” posts off the top section of the Frontpage for everyone by default (by setting the default topic filter for “Community” to “hidden,” meaning that people will be able to opt out of the change). We will probably also remove “Community” posts from Recent Discussion on the Frontpage (you will be able to change this back to the default in your user settings).
    Then we’ll probably either add a tab on the Frontpage or a section lower down on the Frontpage, which would feature a limited number of “Community” posts. You could then load more articles or visit the tab.
    There’s a chance that we’ll tweak our classification of what we mean by “Community” posts, but currently, it’s something like what’s described in this footnote.
    Here are two mockups of the layouts we’re considering testing (if we go for one of these, we’ll probably continue testing and improving it):
    A “Community” tab
    A section for “Community” posts on the Frontpage
    Why consider doing this at all?
    We’re pretty worried about the Forum voting structure driving more engagement to some topics systematically (without the community endorsing this prioritization), and related issues.
    Some topics and discussions tend to get a lot more Forum karma and attention than others, as Ben and Lizka outlined in this post. These tend to be:
    About the community (because it interests almost everyone at least a little bit)
    Accessible to everyone, or on topics where everyone has an opinion
    This phenomenon means that posts that interest a smaller fraction of the community — even if they interest their target readers a lot more and are more useful to their audience (and even if you consider that some other posts’ audiences are larger), will get a lot less attention than might be helpful.
    We want to do something to address this imbalance, and we think that separating out “Community” posts might get us a lot of the benefit while still letting important conversations happen. (We checked a lot of posts we thought were getting extra engagement via this effect; almost all were “Community” posts.)
    Note that this mechanism “overvalues” “Community” posts whether or not you think that “Community” posts are currently over-valued by the average Forum user (the Forum team has mixed opinions on this); given whatever the community actually cares about, things that are easier to discuss or think about will get more engagement. (See more in this comment.)
    We also want the Frontpage to feature different kinds of content that’s particularly interesting to different groups, instead of being full of posts that interest the majority of Forum users to at least some extent (in particular, posts about the EA community). This is how we get cross-pollination between different cause ...
    12 min
  • EA - Unjournal's 1st eval is up: Resilient foods paper (Denkenberger et al) & AMA ~48 hours by david reinstein
    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: Unjournal's 1st eval is up: Resilient foods paper (Denkenberger et al) & AMA ~48 hours, published by david reinstein on February 6, 2023 on The Effective Altruism Forum.
    The Unjournal: reporting some progress
    (Link: our main page, see also our post Unjournal: Call for participants and research.)
    Our group (curating articles and evaluations) is now live on Sciety HERE.
    The first evaluated research project (paper) has now been posted HERE .
    First evaluation: Denkenberger et al
    Our first evaluation is for "Long Term Cost-Effectiveness of Resilient Foods for Global Catastrophes Compared to Artificial General Intelligence Safety", by David Denkenberger, Anders Sandberg, Ross Tieman, and Joshua M. Pearce, published in the International Journal of Disaster Risk Reduction.These three reports and ratings (see a sample below) come from three experts with (what we believe to be) complementary backgrounds (note, these evaluators agreed to be identified rather than remain anonymous):
    Alex Bates: An award-winning cost-effectiveness analyst with some background in considering long-term and existential risks
    Scott Janzwood: A political scientist and Research Director at the Cascade Institute
    Anca Hanea: A senior researcher and applied probabilist based at the Centre of Excellence for Biosecurity Risk Analysis (CEBRA) at the University of Melbourne. She has done prominent research into eliciting and aggregating (expert) judgments, working with the RepliCATS project.
    These evaluations were, overall, fairly involved. They engaged with specific details of the paper as well as overall themes, directions, and implications. While they were largely positive about the paper, they did not seem to pull punches. Some examples of their feedback and evaluation below (direct quotes).
    Extract of evaluation content
    Bates:
    I’d be surprised if I ever again read a paper with such potential importance to global priorities.
    My view is that it would be premature to reallocate funding from AI Risk reduction to resilient food on the basis of this paper alone. I think the paper would have benefitted from more attention being paid to the underlying theory of cost-effectiveness motivating the investigation. Decisions made in places seem to have multiplied uncertainty which could have been resolved with a more consistent approach to analysis.
    The most serious conceptual issue which I think needs to be resolved before this can happen is to demonstrate that ‘do nothing’ would be less cost-effective than investing $86m in resilient foods, given that the ‘do nothing’ approach would potentially include strong market dynamics leaning towards resilient foods.".
    Janzwood
    the authors’ cost-effectiveness model, which attempts to decrease uncertainty about the potential uncertainty-reducing and harm/likelihood-reducing 'power' of resilient food R&D and compare it to R&D on AGI safety, is an important contribution"
    It would have been useful to see a brief discussion of some of these acknowledged epistemic uncertainties (e.g., the impact of resilient foods on public health, immunology, and disease resistance) to emphasize that some epistemic uncertainty could be reduced by exactly the kind of resilient food R&D they are advocating for.
    Hanea:
    The structure of the models is not discussed. How did [they] decide that this is a robust structure (no sensitivity to structure performed as far as I understood)"
    It is unclear if the compiled data sets are compatible. I think the quantification of the model should be documented better or in a more compact way."
    The authors also responded in detail. Some excerpts:
    The evaluations provided well thought out and constructively critical analysis of the work, pointing out several assumptions which could impact findings of the paper while also recognizing the value of the work in spite of s...
    6 min
  • AF - Conditioning Predictive Models: The case for competitiveness by Evan Hubinger
    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: Conditioning Predictive Models: The case for competitiveness, published by Evan Hubinger on February 6, 2023 on The AI Alignment Forum.
    This is the third of seven posts in the Conditioning Predictive Models Sequence based on the forthcoming paper “Conditioning Predictive Models: Risks and Strategies” by Evan Hubinger, Adam Jermyn, Johannes Treutlein, Rubi Hudson, and Kate Woolverton. Each post in the sequence corresponds to a different section of the paper. We will be releasing posts gradually over the course of the next week or so to give people time to read and digest them as they come out.
    3. The case for competitiveness
    In addition to ensuring that we can condition predictive models safely, for such an approach to work as a way to actually reduce AI existential risk, we also need it to be the case that it is competitive—that is, that it doesn’t impose too much of an alignment tax. Following “How do we become confident in the safety of a machine learning system?” we’ll distinguish between two different aspects of competitiveness here that we’ll need to address:
    Training rationale competitiveness [(Implementation competitiveness)]: how hard the training rationale [(getting the model we want)] is to execute. That is, a proposal should fail on training rationale competitiveness if its training rationale is significantly more difficult to implement—e.g. because of compute or data requirements—than competing alternatives.
    Training goal competitiveness [(Performance competitiveness)]: whether, if successfully achieved, the training goal [(the model we want)] would be powerful enough to compete with other AI systems. That is, a proposal should fail on training goal competitiveness if it would be easily outcompeted by other AI systems that might exist in the world.
    To make these concepts easier to keep track of absent the full training stories ontology, we’ll call training rationale competitiveness implementation competitiveness, since it describes the difficulty of implementing the proposal, and training goal competitiveness performance competitiveness, since it describes the achievable performance for the resulting model.
    Implementation competitiveness
    The most generally capable models today, large language models, seem to be well-described as predictive models. That may change, but we think it is also at least quite plausible that the first human-level AGI will be some sort of predictive model, likely similar in structure to current LLMs.
    Furthermore, LLM pre-training in particular seems to be where most of the capabilities of the most advanced current models come from: the vast majority of compute spent training large language models is spent in pre-training, not fine-tuning. Additionally, our guess is that the fine-tuning that is done is best modeled as targeting existing capabilities rather than introducing entirely new capabilities.
    Assuming that, after pre-training, LLMs are well-understood as predictive models, that suggests two possibilities for how to think about different fine-tuning regimes:
    The fine-tuning resulted in a particular conditional of the original pre-trained predictive model.
    The fine-tuning targeted the capabilities by turning the predictive model into one that is no longer well-understood as predictive.
    In the first case, the conditioning predictive models approach would simply be a variation on the exact techniques currently used at the forefront of capabilities, making it hopefully implementation competitive by default.[1] The main way we think such an implementation competitiveness argument could fail is if the fine-tuning necessary to get the sort of conditionals we describe here is substantially harder than alternative fine-tuning paradigms.
    In particular, we think it is likely the case that our proposed solutions will add some amount of o...
    18 min
  • LW - Childhoods of exceptional people by Henrik Karlsson
    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: Childhoods of exceptional people, published by Henrik Karlsson on February 6, 2023 on LessWrong.
    Let’s start with one of those insights that are as obvious as they are easy to forget: if you want to master something, you should study the highest achievements of your field. If you want to learn writing, read great writers, etc.
    But this is not what parents usually do when they think about how to educate their kids. The default for a parent is rather to imitate their peers and outsource the big decisions to bureaucracies. But what would we learn if we studied the highest achievements?
    Thinking about this question, I wrote down a list of twenty names—von Neumann, Tolstoy, Curie, Pascal, etc—selected on the highly scientific criteria “a random Swedish person can recall their name and think, Sounds like a genius to me”. That list is to me a good first approximation of what an exceptional result in the field of child-rearing looks like. I ordered a few piles of biographies, read, and took notes. Trying to be a little less biased in my sample, I asked myself if I could recall anyone exceptional that did not fit the patterns I saw in the biographies, which I could, and so I ordered a few more biographies.
    This kept going for an unhealthy amount of time.
    I sampled writers (Virginia Woolf, Lev Tolstoy), mathematicians (John von Neumann, Blaise Pascal, Alan Turing), philosophers (Bertrand Russell, René Descartes), and composers (Mozart, Bach), trying to get a diverse sample.
    In this essay, I am going to detail a few of the patterns that have struck me after having skimmed 42 biographies. I will sort the claims so that I start with more universal patterns and end with patterns that are less common.
    Exceptional people grow up in exceptional milieus
    This seems to be true for >95 percent of the people I looked at.
    These naked apes, the humans, are intensely social animals. They obsessively internalize values, ideas, skills, and desires from the people who surround them. It is therefore not surprising that those who grow up to be exceptional tend to have spent their formative years surrounded by adults who were exceptional.
    Virginia Woolf never attended school. Her father, Leslie Stephen, who, along with their tutors, educated Virginia and her sister, was an editor, critic, and biographer “complicatedly hated” by his daughter and of such standing that he could invite Henry James, Thomas Hardy, and Alfred Lord Tennyson to dine and converse with his children. Leslie Stephen described his circle, in which Virginia grew up, as “most of the literary people of mark . . . clever young writers and barristers, chiefly of the radical persuasion . . . we used to meet on Wednesday and Sunday evenings, to smoke and drink and discuss the universe and the reform movement.” When they went to the Hebrides in the summers, Leslie brought along painters and philosophers, who would hang out and work in their summer house while the children played.
    This parental obsession with curating a rich intellectual milieu comes through in nearly all of the biographies. As I wrote in First we shape our social graph; then it shapes us:
    Michel Montaigne’s father employed only servants who were fluent in Latin, curating a classical culture, so Montaigne would learn Latin as his mother tongue. J.S. Mill spent his childhood at his father’s desk, helping his father write a treatise on economics, running over to Jeremy Bentham’s house to borrow books and discuss ideas.
    Blaise Pascal, too, was homeschooled by his father. His father chose not to teach him math. (The father, Etienne, had a passion for mathematics that he felt was slightly unhealthy. He feared mathematics would distract Pascal from less intrinsically rewarding pursuits, such as literature, much like modern parents fear TikTok.) Pascal had to teach himself. Wh...
    25 min
  • EA - Podcast with Oli Habryka on LessWrong / Lightcone Infrastructure by DanielFilan
    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: Podcast with Oli Habryka on LessWrong / Lightcone Infrastructure, published by DanielFilan on February 6, 2023 on The Effective Altruism Forum.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • EA - Shallow investigation: Loneliness by Em
    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: Shallow investigation: Loneliness, published by Em on February 6, 2023 on The Effective Altruism Forum.
    This is a shallow report examining loneliness.
    In a nutshell
    Importance: 4-4.5/5 Unfortunately, it is very common to experience loneliness and particularly so later in life, so the importance of loneliness as a cause area seems high. There is evidence that loneliness impacts a lot of health and economic domains. More worryingly, there is evidence of a ‘vicious cycle’ between health and loneliness such that lonely people become sicker and sicker people tend to have fewer opportunities to find social contact. Some evidence suggests more people may become lonelier over time.
    Tractability: 2/5 Current loneliness interventions seem very costly. There is also mixed evidence about the effectiveness of interventions, with some being effective and some not alleviating loneliness. Further challenges relate to the lack of essential data about the prevalence of loneliness, methodological, sampling, targeting and mechanistic problems in interventions. Gaps in knowledge are more pronounced in LMICs.
    Neglectedness: 3-3.5/5 Funding, relevant NGOs and charities, as well as public awareness campaigns are present and even seem to be increasing in high-income settings. At the same time, there is more pronounced neglectedness in LMICs.[1] Further, there is still a stigma about discussing loneliness and mental health issues.
    Key uncertainties
    Key uncertainty 1: How effective and sustainable are anti-loneliness interventions in the long run?
    The current interventional evidence base is predominantly cross-sectional or short in duration, marked by relatively small and not globally representative samples, so it is hard to make strong causal claims as well as claims with high certainty for generalizability. There are few longitudinal interventions, so little is known about the long-term impacts and potential sustainability of such projects.
    Key uncertainty 2: Can anti-loneliness interventions be scaled?
    Some of the promising interventions to tackle loneliness require significant involvement in terms of time and service provision, and so may be difficult to scale. Compounding this difficulty, many of the most socially isolated people live rurally or in poorly connected areas and so reaching them may be challenging.
    Key uncertainty 3: How much can the cost-effectiveness of loneliness interventions (practically) be improved in the short term?
    A surprising key target many current anti-loneliness interventions miss is actually recruiting participants who are lonely. I anticipate this is an essential part to understanding the real cost-effectiveness of available interventions. Beyond this, provision of a relevant intervention type for the relevant group as well as provision for a sufficient length of time are also warranted.
    Importance
    Problem Overview
    Question
    Answer
    Definition of the problem
    Loneliness is commonly defined as a discrepancy between a person’s desired and actual social relationships or opportunities for social contact. This can be a subjective (perceived) or objective discrepancy (social isolation). Loneliness is a risk factor for a variety of general and mental health conditions, mortality, loss of productivity, and poor self-esteem.
    Burden of the problem
    Health burden:
    £170 - £780 per lonely person per year in the UK or about £340 million - £1.56 billion for the lonely adult population in the UK annually.
    In the US, lonely adults pay an additional $1643 in annual healthcare coverage, which means the lonely adult population in the country may be paying an excess of $18.4 billion annually.
    In Australia, loneliness is estimated to cost $2.7 billion a year in excess costs linked to additional GP visits, more physical inactivity and smoking, equivalent to $1,565 per person.
    Productivi...
    57 min
  • LW - Research Direction: Be the AGI you want to see in the world by scottviteri
    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: Research Direction: Be the AGI you want to see in the world, published by scottviteri on February 5, 2023 on LessWrong.
    Be the AGI you want to see in the world.
    Epistemic status: highly speculative, authors are not neuroscientists.
    Summary
    It may be possible to enhance human intelligence via a brain-computer interface (BCI). We could put electrodes into a human brain, connect those electrodes to an artificial neural network, and train that network to predict and write neural activations.
    This may present a technique for gradual uploading of human minds to computers that doesn’t require technology sufficient to create AGI in the first place.
    The purpose of this article is to elicit feedback on this idea and, if it seems promising, encourage more work in this area.
    Introduction: Goal and Idea
    Goal
    We suspect that it may be possible to develop an enhancement system for human brains which satisfies the following desiderata:
    Competitiveness: Our enhancement system must be powerful enough to help humans escape the acute AGI risk period. This desideratum may be satisfied if enhanced humans are able to find a solution to the prosaic AI alignment problem.
    Timeliness: Our enhancement system must be fully developed before the advent of AGI.
    Value preservation: Our enhancement system must not severely distort the core values of the enhanced human.
    Idea
    Human brains are sharply constrained in size. Indeed, human birth is difficult because our brains (and thus skulls) are unusually large compared to other primates.
    If evolution progressed such that human brains evolved until it hit a size constraint, then it seems likely that further increasing human brain size would yield more than modest gains to human intelligence.
    One possible approach to increasing human brain size is by predicting real-time activations in the neocortex. The guiding intuition is that this is already how the brain extends itself, like how the cerebellum currently predicts activations in the telencephalon. Namely, Neuralink hardware currently supports three thousand EEG probes which can read from the brain, and Neuralink has the intention to add write access as well. Write access in this context means that the Neuralink probes could pipe brain activation data into a neural network which is trained to predict future brain activations, and the write probes could send those predictions as signals back into the brain. Scott believes that other hypotheses such as spike-timing dependent plasticity and predictive coding further support the hypothesis that the brain will learn to harness predictive signals from the EEG probes.
    Human brains are optimized for low energy expenditure rather than speed, and it can take tens of milliseconds for a single serial communication between neurons. Artificial neural networks could run faster. For instance, if a language model the size of Chinchilla could predict neural activations, then current neural network hardware could speed serial computations in the brain by two orders of magnitude (0.7ms to 70ms, of course depending on the distance between the neurons).
    However, 3e3 probes is essentially nothing in the context of a brain with 8e10 neurons. To keep the orders of magnitude consistent, this is like expecting to read and predict 10 or 100 activations out of a neural network with 100M. Our core hypothesis is that this is not the correct neural network analogy: since both the neural prosthesis and your brain are simultaneously learning, it is more like jointly training the two networks. In this setting, the loss could flow into the helper network through the EEG probes, and the main network could learn to supply inputs to the helper network such that it is especially helpful. For instance, if the Neuralink network has access to a compass, the internet, or 4D-shape rotation softwar...
    13 min