The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • AF - Can we efficiently explain model behaviors? by Paul Christiano
    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: Can we efficiently explain model behaviors?, published by Paul Christiano on December 16, 2022 on The AI Alignment Forum.
    ARC’s current plan for solving ELK (and maybe also deceptive alignment) involves three major challenges:
    Formalizing probabilistic heuristic argument as an operationalization of “explanation”
    Finding sufficiently specific explanations for important model behaviors
    Checking whether particular instances of a behavior are “because of” a particular explanation
    All three of these steps are very difficult, but I have some intuition about why steps #1 and #3 should be possible and I expect we’ll see significant progress over the next six months. Unfortunately, there’s no simple intuitive story for why step #2 should be tractable, so it’s a natural candidate for the main technical risk.
    In this post I’ll try to explain why I’m excited about this plan, and why I think that solving steps #1 and #3 would be a big deal, even if step #2 turns out to be extremely challenging.
    I’ll argue:
    Finding explanations is a relatively unambitious interpretability goal. If it is intractable then that’s an important obstacle to interpretability in general.
    If we formally define “explanations,” then finding them is a well-posed search problem and there is a plausible argument for tractability.
    If that tractability argument fails then it may indicate a deeper problem for alignment.
    This plan can still add significant value even if we aren’t able to solve step #2 for arbitrary models.
    I. Finding explanations is closely related to interpretability
    Our approach requires finding explanations for key model behaviors like “the model often predicts that a smiling human face will appear on camera.” These explanations need to be sufficiently specific that they distinguish (the model actually thinks that a human face is in front of the camera and is predicting how light reflects off of it) from (the model thinks that someone will tamper with the camera so that it shows a picture of a human face).
    Our notion of “explanation” is informal, but I expect that most possible approaches to interpretability would yield the kind of explanation we want (if they succeeded at all). As a result, understanding when finding explanations is intractable may also help us understand when interpretability is intractable.
    As a simple caricature, suppose that we identify a neuron representing the model’s beliefs about whether there is a person in front of the camera. We then verify experimentally that (i) when this neuron is on it leads to human faces appearing on camera, (ii) this neuron tends to fire under the conditions where we’d expect a human to be in front of the camera.
    I think that finding this neuron is the hard part of explaining the face-generating-behavior. And if this neuron actually captures the model’s beliefs about humans, then it will distinguish (human in front of camera) from (sensors tampered with). So if we can find this neuron, then I think we can find a sufficiently specific explanation of the face-generating-behavior.
    In reality I don’t expect there to be a “human neuron” that leads to such a simple explanation, but I think the story is the same no matter how complex the representation is. If beliefs about humans are encoded in a direction then both tasks require finding the direction; if they are a nonlinear function of activations then both tasks require understanding that nonlinearity; and so on..
    The flipside of the same claim is that ARC’s plan effectively requires interpretability progress. From that perspective, the main way ARC’s research can help is by identify a possible goal for interpretability. By making a goal precise we may have a better chance of automating it (by applying gradient descent and search, as discussed in section III), and even if we can’t automate it then...
    15 min
  • EA - The Effective Altruism movement is not above conflicts of interest by sphor
    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 Effective Altruism movement is not above conflicts of interest, published by sphor on December 16, 2022 on The Effective Altruism Forum.
    I'm linking and excerpting a submission to the EA criticism contest published by a pseudonymous author on August 31, 2022 (i.e. before the collapse of FTX).
    The submission did not win a prize, but was highlighted by a panelist:
    I was unsure about including this post, but I think this post highlights an important risk of the EA community receiving a significant share of its funding from a few sources, both for internal community epistemics/culture considerations as well as for external-facing and movement-building considerations. I don't agree with all of the object-level claims, but I think these issues are important to highlight and plausibly relevant outside of the specific case of SBF / crypto. That it wasn't already on the forum (afaict) also contributed to its inclusion here.
    Due to concerns about copyright, I'm excerpting the post besides the summary and disclaimer, but I recommend reading the whole piece.
    Summary
    Sam Bankman-Fried, founder of the cryptocurrency exchange FTX, is a major donator to the Effective Altruism ecosystem and has pledged to eventually donate his entire fortune to causes aligned with Effective Altruism.
    By relying heavily on ultra-wealthy individuals like Sam Bankman-Fried for funding, the Effective Altruism community is incentivized to accept political stances and moral judgments based on their alignment with the interests of its wealthy donators, instead of relying on a careful and rational examination of the quality and merits of these ideas. Yet, the Effective Altruism community does not appear to recognize that this creates potential conflicts with its stated mission of doing the most good by adhering to high standards of rationality and critical thought.
    In practice, Sam Bankman-Fried has enjoyed highly-favourable coverage from 80,000 Hours, an important actor in the Effective Altruism ecosystem. Given his donations to Effective Altruism, 80,000 Hours is, almost by definition, in a conflict of interest when it comes to communicating about Sam Bankman-Fried and his professional activities. This raises obvious questions regarding the trustworthiness of 80,000 Hours’ coverage of Sam Bankman-Fried and of topics his interests are linked with (quantitative trading, cryptocurrency, the FTX firm.).
    In this post, I argue that the Effective Altruism movement has failed to identify and publicize its own potential conflicts of interests. This failure reflects poorly on the quality of the standards the Effective Altruism movement holds itself to. Therefore, I invite outsiders and Effective Altruists alike to keep a healthy level of skepticism in mind when examining areas of the discourse and action of the Effective Altruism community that are susceptible to be affected by incentives conflicting with its stated mission. These incentives are not just financial in nature, they can also be linked to influence, prestige, or even emerge from personal friendships or other social dynamics. The Effective Altruism movement is not above being influenced by such incentives, and it seems urgent that it acts to minimize conflicts of interest.
    Introduction — Cryptocurrency is not neutral (neither morally nor politically)
    ... Cryptocurrency is not simply an attempt to provide a set of technical solutions to improve existing currency systems. It is an attempt to replace existing monetary institutions by a new political system, it is therefore political at its core...
    My point here is not to debate on the virtues of the societal model promoted by cryptocurrency actors, but rather to convince readers unfamiliar with the cryptocurrency industry that it is deeply infused with political ideology and is certainly not a purely-te...
    10 min
  • EA - MacKenzie Scott's grantmaking data 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: MacKenzie Scott's grantmaking data, published by david reinstein on December 15, 2022 on The Effective Altruism Forum.
    "MacKenzie Scott's finally revealed her grants to date" Dylan Mathews notes in this Twitter thread.
    It seems mostly USA-directed, but some 'good stuff from an EA POV'. I'm working on unpacking this further.
    Above: her donations in $ millions classified by year and by whether they go to a charity that includes a US location under 'Geographies of service', but may also operate abroad.(Next step: classify the amount that might go abroad, and/or might be effective.)My code is here, feel free to continue to continue this work, I'm taking a break for now.Her largest donations across years (my aggregation):
    Dylan's aggregation by org:Below, some of her largest donations in particular years:
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Radical tactics can increase support for more moderate groups by James Ozden
    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: Radical tactics can increase support for more moderate groups, published by James Ozden on December 16, 2022 on The Effective Altruism Forum.
    Note: This post is a synthesised version of this report, which analyses some public opinion polling we did for a disruptive climate-focused campaign in the UK, by Just Stop Oil. I think this might be relevant to EAs as I've generally perceived there to be quite a lot of skepticism about groups who employ "radical" tactics (e.g. disruptive protest) as it might taint a certain field, or otherwise slow overall progress. Whilst this work only answers the public opinion component of this concern, we find that radical tactics can actually increase support for and identification with more moderate organisations working on the same issue - a helpful dynamic. However, it's plausible that this works best for large social movements (e.g. the climate movement) where distinctions between a radical faction and moderate faction are much more clear. That said, we also found some trends towards polarisation that we intend to analyse further.
    Summary
    Social Change Lab conducted nationally representative YouGov surveys, before and after a week-long campaign by Just Stop Oil to block the M25 motorway. These surveys were conducted longitudinally, by surveying the same people before and after the Just Stop Oil M25 campaign. Our aim was to see if a ‘radical flank effect’ was at play: did the radical tactics implemented by Just Stop Oil impact attitudes towards more moderate UK climate organisations? We surveyed 1,415 members of the public about their support for climate policies and support for and identification with a more moderate climate organisation (Friends of the Earth). We detected a positive radical flank effect, whereby increased awareness of Just Stop Oil resulted in increased support for and identification with Friends of the Earth (p=0.004 and p=0.007 respectively).
    We believe this is the first time the radical flank effect has been observed empirically using large-scale nationally representative polling, and it corroborates previous experimental findings by Simpson et al. (2022). The results indicate the potential positive effects of radical tactics on a broader social movement.
    Support for climate policies also increased between our two surveys, though we attribute this largely to media coverage of COP27, which took place at the same time as the M25 protests. We believe this change was largely due to COP27 as we observed no positive correlation between awareness of Just Stop Oil and support for climate policies, and in fact, observed a statistically non-significant (p = 0.19) negative association.
    Key Results:
    Over 92% of UK adults had heard of Just Stop Oil after their campaign, putting awareness of the organisation as high as the top 20 UK charities. This figure was 87% before this campaign started (as people may have reported higher awareness of Just Stop Oil purely due to our first survey).
    The number of people saying they support Friends of the Earth increased from 50.3% to 52.9% of the population, a 2.6 percentage point increase, equivalent to 1.75 million people in the UK.
    A clear positive radical flank effect: Increased awareness of Just Stop Oil after their M25 protest campaign was linked with stronger identification with and support for Friends of the Earth.
    A trend towards polarisation: Increased awareness of Just Stop Oil through their M25 campaign tended to make people who had low baseline identification with a moderate climate organisation reduce their support for climate policies; the opposite was true for people with high levels of baseline identification, who showed increased support for climate policies with increasing awareness of Just Stop Oil.
    Introduction
    Social movements are often made up of several factions, dep...
    27 min
  • AF - Paper: Transformers learn in-context by gradient descent by Lawrence Chan
    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: Paper: Transformers learn in-context by gradient descent, published by Lawrence Chan on December 16, 2022 on The AI Alignment Forum.
    The paper argues that auto-regressive transformers implement in-context learning via gradient-based optimization on in-context data.
    The authors start by pointing out that with a single linear self-attention (LSA) layer (that is, no softmax), a Transformer can implement one step of gradient descent on the l2 regression loss (a fancy way of saying w -= LR (w x-y)x^T), and confirm this result empirically. They extend this result by showing that an N-layer LSA-only transformer is similar to N-steps of gradient descent for small linear regression tasks, both in and out of distribution. They also find that the results pretty much hold with softmax self-attention (which isn’t super surprising given you can make a softmax pretty linear).Next, they show empirically that the forward pass of a small transformer with MLPs behaves similarly to an meta-learned MLP + one steps of gradient descent on a toy non-linear regression task, again in terms of both in-distribution and OOD performance.
    They then show how you can interpret an induction head as a single step of gradient descent, and provide circumstantial evidence that this explains some of the in-context learning observed in Olsson et al 2022. Specially, they show that 1) a two layer attention-only transformers converge to loss consistent with one step of GD on this task, and 2) the first layer of the network learns to copy tokens one sequence position over in the first layer, prior to the emergence of in-context learning.
    (EDIT:) davidad says below:
    this is strong empirical evidence that mesa-optimizers are real in practice
    Personally, while I think you could place this in the same category as papers like RL^2 or In-context RL with Algorithmic Distillation, which also show mesa optimization, I think the more interesting results are the mechanistic ones -- i.e., that some forms of mesa optimization in the model seem to be implemented via something like gradient descent.
    (Nitpick: I do want to push back a bit on their claim that they've "mechanistically understand the inner workings of optimized Transformers that learn in-context", since they've only really looked at the mechanism of how single attention-only transformers perform in-context learning. )
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - Today is Draft Amnesty Day (December 16-18) 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: Today is Draft Amnesty Day (December 16-18), published by Lizka on December 16, 2022 on The Effective Altruism Forum.
    TL;DR: Today is an official “Draft Amnesty Day” on the Forum. People are encouraged to share unfinished posts, unpolished writing, butterfly ideas, thoughts they’re not sure they endorse, etc.
    This post explains how to:
    Participate ⬇️
    Explore what people have written ⬇️
    Hide these posts so that you don’t need to see them ⬇️
    Participate!
    If you have a draft lying around that you’ve never gotten around to polishing or checking, you are invited to post it today. (Here's why I think that's valuable.)
    To make it clear that it’s a “Draft Amnesty Day” post, do the following:
    Tag it with the Draft Amnesty Day tag
    Copy-paste the following table into the post, and modify it as you’d like:
    This is a Draft Amnesty Day draft. That means it’s not polished, it’s probably not up to my standards, the ideas are not thought out, and I haven’t checked everything. I was explicitly encouraged to post something unfinished! Commenting and feedback guidelines: I’m going with the default — please be nice. But constructive feedback is appreciated; please let me know what you think is wrong. Feedback on the structure of the argument is also appreciated.
    You can also elaborate more on what types of feedback would be most useful, the context for the original draft, and the like.
    Explore what people have written
    Here’s a list of posts that have been published with this tag. Explore them! If you like one of them, tell the author! You can also give constructive feedback, which could be enormously useful to the poster, but please be generous, respect the norms, and see if the author has asked for certain types of feedback specifically.
    How to hide these posts entirely
    You can hide Draft Amnesty Day posts entirely by using a tag filter. Go to the Frontpage of the Forum, then hover over the “Draft Amnesty Day” tag (find it to the right of the “Frtonpage Posts” header), and click on “Hidden.” Posts with this tag will stop showing up for you.
    Depending on your timezone, you might see this early! In that case, tomorrow is a Draft Amnesty Day.
    If it's early, there might not be content there, yet.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • LW - My AGI safety research—2022 review, ’23 plans by Steven Byrnes
    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: My AGI safety research—2022 review, ’23 plans, published by Steven Byrnes on December 14, 2022 on LessWrong.
    The short version: In this post I’m briefly summarizing how I spent my work-time in 2022, and what I’m planning for 2023.
    The first half of 2022 was writing the “Intro to Brain-Like-AGI Safety” blog post series.
    The second half of 2022 was split maybe 45%-45%-10% between my main research project (on reverse-engineering human social instincts), miscellaneous other research and correspondence, and outreach mostly targeted towards neuroscientists.
    I expect to carry on with a similar time allocation into 2023.
    If you think there are other things I should be doing instead or differently, please don’t be shy, the comment section is below, or DM me, email, etc.
    The long version:
    1. First half of 2022: Writing “Intro to Brain-Like AGI Safety”
    So, I was writing some technical post in late 2021, and realized that the thing I was talking about was a detail sitting on top of a giant pile of idiosyncratic beliefs and terminology that nobody else would understand. So I started writing a background section to that post. That background section grew and grew and grew, and eventually turned into a book-length series of 15 blog posts entitled “Intro to Brain-Like AGI Safety”, which reorganized and re-explained almost everything I had written and thought about up to that point, since I started in the field around 2019. (My palimpsest!) Writing that series took up pretty much 100% of my work time until May 2022.
    Then I spent much of the late spring and summer catching up on lots of miscellaneous to-do-list stuff that I had put off while writing the series, and everyone in my family caught COVID, and we took a family vacation, and I attended two conferences, and I switched jobs when Jed McCaleb generously offered me a home at Astera Institute, and various other things. So I didn’t get much research done during the late spring and summer.
    Moving on to the rest of the year, my substantive work time has been divided, I dunno, something like 45%-45%-10% between “my main research project”, “other research”, and “outreach”. Let’s take those one at a time in the next three sections.
    2. Second half of 2022 (1/3): My main research project
    2.1 What’s the project?
    I’m working on the open neuroscience problem that I described in the post Symbol Grounding and Human Social Instincts, and motivated in the post Two paths forward: “Controlled AGI” and “Social-instinct AGI”. I’ll give an abbreviated version here.
    As discussed in “Intro to Brain-Like-AGI Safety”, I hold the following opinions:
    We should think of within-lifetime learning in the human brain as a kind of model-based reinforcement learning (RL) system;
    We should think of that model-based RL system as potentially similar to how future AGIs will work;
    We should (to a first approximation) think of the “reward function” of that RL system as encoding “innate drives”, like pain being bad and sweet tastes being good;
    These “innate drives” correspond to specific genetically-hardwired circuitry primarily in the hypothalamus and brainstem;
    A subset of that circuitry underlies human social and moral instincts;
    .And the project I’m working on is an attempt to figure out what those circuits are and how they work.
    2.2 Why do I think success on this project would be helpful for AGI safety?
    I have two arguments:
    The modest argument is: At some point, I hope, we will have a science that can produce predictions of the form:
    (“Innate drives” a.k.a. “Reward function” X) + (“Life experience” a.k.a. “Training environment” Y) (“Adult” AGI that’s trying to do Z)
    If we knew exactly what innate drives are in humans (particularly related to sociality, morality, etc.), then we would have actual examples of X+YZ to ground this future science.
    Even with the benefit...
    12 min
  • EA - $50 TisBest Charity Gift Card to the first 20,000 people who sign up by Michael Huang
    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: $50 TisBest Charity Gift Card to the first 20,000 people who sign up, published by Michael Huang on December 15, 2022 on The Effective Altruism Forum.
    $50 Charity Gift Cards are available today on a first come basis.
    “We’re helping this idea catch fire by offering a $50 Charity Gift Card to the first 20,000 people who sign up so that more people can experience this uniquely meaningful holiday gift for themselves. There are no strings attached. Our hope is simply that once you see how easy and joyous this kind of gift can be, you’ll choose to give it to your friends and family. We also ask that if you claimed one of these free cards last year, you either let others participate this year or pass this offer along to someone else who hasn’t received one yet.”
    Paul Tudor Jones
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - I'm as approving of the EA community now as before the FTX collapse by Will Aldred
    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'm as approving of the EA community now as before the FTX collapse, published by Will Aldred on December 15, 2022 on The Effective Altruism Forum.
    (Posting on behalf of and with permission from Duncan Sabien; the first person speaker is Duncan. Full text below.)
    Just took the Effective Altruism survey, and it had an extra, optional section that had a lot of questions about the FTX stuff, and trust, and how EA should respond, and what I think of it, and so forth.
    I'm not really an EA; haven't taken the pledge, don't work at an org, have been to fewer than twenty EA meetups in my life (though I've been a speaker at multiple EA Globals).
    However, I've been close to, and quite fond of, and at least a little protective of the EA community, for the past seven years (for instance, volunteering to speak at multiple EA Globals!).
    And the questions on the survey made me want to note publicly:
    I feel almost exactly as fond of, and approving of, and protective of, the EA community writ-large as I have for the past seven years.
    I get that the FTX thing is a big deal. Every EA I've seen has treated it as a big deal. It's being taken seriously, and I've seen soul-searching at every level from the individual up to the meta-organizational.
    I think there's a mistake that the average person tends to make, though, which is something like "if a plane crashes, something absolutely needs to visibly change."
    Often, when a plane crashes, something absolutely needs to change! Often there are legitimate flaws in the system that need to be patched.
    But sometimes, you just get that confluence of three one-in-a-thousand events. Sometimes, the right answer really is "our system shouldn't change; this is the exception."
    I'm not claiming that's the case here, at all. I'm mentioning it because:
    it is in fact sometimes true, and the outrage machine doesn't take that fact into account
    it is a good reminder of the difference between improvement and improvement theater.
    There are actions you can take to look like you're taking things seriously, and there are actions you can take because you're actually taking things seriously.
    Public relations are also important, so some amount of the former category belongs in the latter category.
    But overall, I'm not interested in insisting that the individuals and organizations in the Effective Altruism sphere do things that look to me, from the outside, like sufficient due diligence, in response to this crisis.
    I want them to simply respond. To the best of their ability, in the ways that seem right to them.
    And I do, in fact, trust that that is happening. In part because of the glimpses I've caught, but also in part because that's just ... firmly in my model of these people and these orgs.
    Or to put it another way: the FTX thing was a blindside, and a negative update, but it was a negative update in the slack. It was the kind of out-of-distribution, surprisingly bad event that I sort of ... budget room for, on the meta level?
    I expect there to be one or two bad things here and there, when you're trying to coordinate thousands and thousands of individuals all pulling in different directions. This particular badness was in an extremely high-leverage place, and that sucks, but that feels more like bad luck than like ... "HOW DARE YOU ALL NOT HAVE PREDICTED THIS SORT OF THING AND ALSO BEEN FULLY ROBUST AGAINST ANY SUCH SHENANIGANERY AT ALL TIMES AND FROM ALL ANGLES."
    Some people are shouting stuff like that. But I think those people are being unreasonable, and not considering what a world with that level of vigilance actually looks like, in practice (hint: it looks like paralysis).
    I think that if a second disaster falls (or if a second looming disaster is uncovered and prevented before it actually breaks) then sure, I will owe those people an apology, and will make a p...
    5 min
  • EA - Consider working more hours and taking more stimulants by Arjun Panickssery
    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: Consider working more hours and taking more stimulants, published by Arjun Panickssery on December 15, 2022 on The Effective Altruism Forum.
    Epistemic status: amphetamines
    Sooner or later, the great men turn out to be all alike. They never stop working. They never lose a minute. It is very depressing.
    V.S. Pritchett
    EA produces much talk about an obligation to donate some or even most of your wealth. In both direct work and earning to give, there's a connection between your work productivity and your (direct or indirect) impact. Hard work is also a costly signal of commitment that could substitute for frugality in our less funding-constrained phase of the movement. And working incredibly hard increases the chance of tail successes that might generate very high impact.
    In the same way that you might want to attract converts by advancing a softer norm of donating only 10% of your income rather than everything above $40k, you might want to create a softer norm about productivity, and not feel bad about only following this norm. This post is addressed instead to those who haven't considered much at all the prospect of experimenting with working 60 hours a week rather than 30-40.
    Don't dismiss this option out of hand because of general concerns about burnout. There are multiple good reasons to think you should work much harder.
    First, the short-term optimal workweek might just be very long. Studies often find that CEOs work 50+ hours per week. Silicon Valley is very productive and has a "hustle culture" involving long work hours (see also). I agree with Lynette Bye that most of the working hours literature is poor—I'm even more skeptical than she is about agenda-driven research on Gilded Age factory workers—and that gaining an impression from anecdotes of top performers is better. Top performers in business routinely work long hours, and reading through lists of anecdotes like Daily Rituals (which is mostly writers and artists) you'll see a lot of strict routines, long hours, and stimulants of all kinds: caffeine, nicotine, amphetamines.
    High-performing managers in the EA ecosystem report working long hours:
    "I know there have been some pretty publicized debates recently in Silicon Valley about whether you can have it all and be successful and have balance, or whether you really do need to almost work yourself to death in order to accomplish something big. At least in our mind and our experience with our company is that sometimes if you want to do something that is very challenging and that you think will have a big impact and hasn’t been done before the reality is you just have to run and work harder and faster than anyone else and I think that’s a very real thing for us. We have gone all-in on this." -Theresa Condor, COO at Spire Global
    "Yeah, so the limiting factor is I only want to work about 55 hours a week or something like that at the most. And so maybe I'll experiment in pushing that up to 60. But somewhere between 45 and 60 hours a week. That is hours in the office, and that transfers into Toggl hours at a rate of something like 75% to 85% just because of pee breaks and chatting to people and eating and stuff like that. And so that translates into around 40 Toggl hours, like 35 to 40 Toggl hours a week. And then, then the fraction of those that are spent on key priorities. The rest of it is meetings with people and emails, which are the two biggest sucks of time. Also internal comms, which is checking Slack or recording my goals and stuff like that." -Niel Bowerman, Director of Special Projects at 80,000 Hours
    However, productive hours look more limited for certain types of cognitive or intellectual work:
    "I've done several different kinds of work, and the limits were different for each. My limit for the harder types of writing or programming is about five hours ...
    7 min