The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - Racial and gender demographics at EA Global in 2022 by Amy Labenz
    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: Racial and gender demographics at EA Global in 2022, published by Amy Labenz on March 10, 2023 on The Effective Altruism Forum.
    CEA has recently conducted a series of analyses to help us better understand how people of different genders and racial backgrounds experienced EA Global events in 2022 (not including EAGx). In response to some requests (like this one), we wanted to share some preliminary findings.
    This post is primarily going to report on some summary statistics. We are still investigating pieces of this picture but wanted to get the raw data out fast for others to look at, especially since we suspect this may help shed light on other broad trends within EA.
    High-level summary
    Attendees:
    33% of registered attendees (and 35% of applicants) at EA Global events in 2022 self-reported as female or non-binary.
    33% of registered attendees (and 38% of applicants) self-reported as people of color (“POC”).
    Experiences:
    Attendees generally find EA Global welcoming (4.51/5 with 1–5 as options) and are likely to recommend it to others (9.03/10 with 0–10 as options).
    Women and non-binary attendees reported that they find EA Global slightly less welcoming (4.46/5 compared to 4.56/5 for men and 4.51 overall).
    Otherwise, we found no statistically significant difference in terms of feelings of welcomeness and overall recommendation scores across groups in terms of gender and race/ethnicity.
    Speakers:
    43% of speakers and MCs at EA Global events in 2022 were female or non-binary.
    28% of speakers and MCs were people of color.
    Some initial takeaways:
    A more diverse set of people apply to and attend EAG than complete the EA survey.
    Welcomingness and likelihood to recommend scores for women and POC were very similar to the overall scores.
    There is a small but statistically significant difference in welcomingness scores for women.
    We are not sure what to make of the fact that the application stats for POC were higher than the admissions stats. We are currently investigating whether this demographic is newer to EA (our best guess) and if that might be influencing the admissions rate.
    One update for our team is that women / non-binary speaker stats are higher compared to the applicant pool and this is not the case for POC. We had not realized that prior to conducting this analysis.
    The 2022 speaker statistics appear to be broadly in line with our statistics since London 2018 when we started tracking. We had significantly less diverse speakers prior to EAG London 2018.
    Applicants and registrants
    For EA Globals in 2022, our applicant pool was slightly more diverse in terms of race/ethnicity than our attendee pool (38% of applicants were POC vs. 33% of attendees), and around the same in terms of gender (35% of applicants were female or non-binary vs. 33% of attendees).
    For comparison, our attendee pool has about the same composition in terms of gender as the respondents in the 2020 EA Survey and is more diverse in terms of race/ethnicity than that survey.
    We had much more racial diversity at EAGx events outside of the US and Europe (e.g. EAGxSingapore, EAGxLatAm, and EAGxIndia, where POC were the majority). Generally, EAGx attendees end up later attending EAGs, so we think the events could result in more attendees from these locations. (However, due to funding constraints and their impact on travel grants, we expect this will not impact EAGs in 2023 as much as it might have otherwise.)
    Experiences of attendees
    Overall, attendees tend to find EA Global welcoming (4.51/5 with 1–5 as options) and are likely to recommend it to others (9.03/10 with 0–10 as options).
    Women and non-binary attendees reported slightly lower average scores on whether EA Global was “a place where [they] felt welcome” (women and non-binary attendees reported an average score of 4.46/5 vs an average of 4.56/5 for me...
    14 min
  • EA - How oral rehydration therapy was developed by Kelsey Piper
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How oral rehydration therapy was developed, published by Kelsey Piper on March 10, 2023 on The Effective Altruism Forum.
    This is a link post for "Salt, Sugar, Water, Zinc: How Scientists Learned to Treat the 20th Century’s Biggest Killer of Children" in the second issue of Asterisk Magazine, now out. The question it poses is: oral rehydration therapy, which has saved millions of lives a year since it was developed, is very simple. It uses widely available ingredients. Why did it take until the late 1960s to come up with it?
    There's sort of a two part answer. The first part is that without a solid theoretical understanding of the problem you're trying to solve, it's (at least in this case) ludicrously difficult to solve it empirically: people kept trying variants on this, and they didn't work, because an important parameter was off and they had no idea which direction to correct in.
    The second is that the incredible simplicity of the modern formula for oral rehydration therapy is the product of a lot of concerted design effort not just to find something that worked against cholera but to find something dead simple which did only require household ingredients and was hard to get wrong. The fact the final solution is so simple isn't because oral rehydration is a simple problem, but because researchers kept on going until they had a sufficiently simple solution.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - Japan AI Alignment Conference by Chris Scammell
    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: Japan AI Alignment Conference, published by Chris Scammell on March 10, 2023 on LessWrong.
    Conjecture and ARAYA are hosting and organizing the first Japan AI Alignment Conference. The conference will take place in Tokyo, Japan on March 11 and 12. Details about the event can be found here. This event is generously supported by a grant from the Long Term Future Fund.
    The goal of the conference is to illustrate the AI control problem to Japanese AI researchers, introduce them to current trends in AI alignment research, inspire new research directions, and to provide Western researchers exposure to a different set of AI safety thoughts from Japan. This is an exploratory event, and we plan to write a postmortem about the event in due time.
    The first half of the conference will be livestreamed. It will feature an opening talk from Connor Leahy (CEO of Conjecture), a fireside chat between Ryota Kanai (CEO of ARAYA) and Jaan Tallinn, and some presentations on AI safety research directions in the West and in Japan. You can follow the first part of the conference here. The livestream runs from 9:30am-12:30pm JST.
    The rest of the conference will not be livestreamed, and will consist of in-person small group workshops to discuss various AI alignment research directions.The conference will have ~50 attendees from ARAYA, Conjecture, Whole Brain Architecture Initiative, MIRI, OpenAI, RIKEN, Ritsumeikan University, University of Tokyo, Omron Sinic X, Keio University, and others.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • LW - The hot mess theory of AI misalignment: More intelligent agents behave less coherently by Jonathan Yan
    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 hot mess theory of AI misalignment: More intelligent agents behave less coherently, published by Jonathan Yan on March 10, 2023 on LessWrong.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • AF - Japan AI Alignment Conference by Chris Scammell
    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: Japan AI Alignment Conference, published by Chris Scammell on March 10, 2023 on The AI Alignment Forum.
    Conjecture and ARAYA are hosting and organizing the first Japan AI Alignment Conference. The conference will take place in Tokyo, Japan on March 11 and 12. Details about the event can be found here. This event is generously supported by a grant from the Long Term Future Fund.
    The goal of the conference is to illustrate the AI control problem to Japanese AI researchers, introduce them to current trends in AI alignment research, inspire new research directions, and to provide Western researchers exposure to a different set of AI safety thoughts from Japan. This is an exploratory event, and we plan to write a postmortem about the event in due time.
    The first half of the conference will be livestreamed. It will feature an opening talk from Connor Leahy (CEO of Conjecture), a fireside chat between Ryota Kanai (CEO of ARAYA) and Jaan Tallinn, and some presentations on AI safety research directions in the West and in Japan. You can follow the first part of the conference here. The livestream runs from 9:30am-12:30pm JST.
    The rest of the conference will not be livestreamed, and will consist of in-person small group workshops to discuss various AI alignment research directions.The conference will have ~50 attendees from ARAYA, Conjecture, Whole Brain Architecture Initiative, MIRI, OpenAI, RIKEN, Ritsumeikan University, University of Tokyo, Omron Sinic X, Keio University, and others.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Announcing the Open Philanthropy AI Worldviews Contest by Jason Schukraft
    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: Announcing the Open Philanthropy AI Worldviews Contest, published by Jason Schukraft on March 10, 2023 on The Effective Altruism Forum.
    We are pleased to announce the 2023 Open Philanthropy AI Worldviews Contest.
    The goal of the contest is to surface novel considerations that could influence our views on AI timelines and AI risk. We plan to distribute $225,000 in prize money across six winning entries. This is the same contest we preannounced late last year, which is itself the spiritual successor to the now-defunct Future Fund competition. Part of our hope is that our (much smaller) prizes might encourage people who already started work for the Future Fund competition to share it publicly.
    The contest deadline is May 31, 2023. All work posted for the first time on or after September 23, 2022 is eligible. Use this form to submit your entry.
    Prize Conditions and Amounts
    Essays should address one of these two questions:
    Question 1: What is the probability that AGI is developed by January 1, 2043?
    Question 2: Conditional on AGI being developed by 2070, what is the probability that humanity will suffer an existential catastrophe due to loss of control over an AGI system?
    Essays should be clearly targeted at one of the questions, not both.
    Winning essays will be determined by the extent to which they substantively inform the thinking of a panel of Open Phil employees. There are several ways an essay could substantively inform the thinking of a panelist:
    An essay could cause a panelist to change their central estimate of the probability of AGI by 2043 or the probability of existential catastrophe conditional on AGI by 2070.
    An essay could cause a panelist to change the shape of their probability distribution for AGI by 2043 or existential catastrophe conditional on AGI by 2070, which could have strategic implications even if it doesn’t alter the panelist’s central estimate.
    An essay could clarify a concept or identify a crux in a way that made it clearer what further research would be valuable to conduct (even if the essay doesn’t change anybody’s probability distribution or central estimate).
    We will keep the composition of the panel anonymous to avoid participants targeting their work too closely to the beliefs of any one person. The panel includes representatives from both our Global Health & Wellbeing team and our Longtermism team. Open Phil’s published body of work on AI broadly represents the views of the panel.
    Panelist credences on the probability of AGI by 2043 range from ~10% to ~45%. Conditional on AGI being developed by 2070, panelist credences on the probability of existential catastrophe range from ~5% to ~50%.
    We will award a total of six prizes across three tiers:
    First prize (two awards): $50,000
    Second prize (two awards): $37,500
    Third prize (two awards): $25,000
    Eligibility
    Submissions must be original work, published for the first time on or after September 23, 2022 and before 11:59 pm EDT May 31, 2023.
    All authors must be 18 years or older.
    Submissions must be written in English.
    No official word limit — but we expect to find it harder to engage with pieces longer than 5,000 words (not counting footnotes and references).
    Open Phil employees and their immediate family members are ineligible.
    The following groups are also ineligible:
    People who are residing in, or nationals of, Puerto Rico, Quebec, or countries or jurisdictions that prohibit such contests by law
    People who are specifically sanctioned by the United States or based in a US-sanctioned country (North Korea, Iran, Russia, Myanmar, Afghanistan, Syria, Venezuela, and Cuba at time of writing)
    You can submit as many entries as you want, but you can only win one prize.
    Co-authorship is fine.
    See here for additional details and fine print.
    Submission
    Use this form to submit your entries. We strongl...
    7 min
  • EA - The Ethics of Posting: Real Names, Pseudonyms, and Burner Accounts by Sarah Levin
    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 Ethics of Posting: Real Names, Pseudonyms, and Burner Accounts, published by Sarah Levin on March 9, 2023 on The Effective Altruism Forum.
    Recently there’s been debate about the ethics of using burner accounts to make attacks and accusations on this forum. See The number of burner accounts is too damn high and Why People Use Burner Accounts especially. This post is a more systematic discussion of poster identity, reputation, and accountability.
    Types of Accounts
    We can roughly break accounts down into four categories:
    Real name accounts are accounts under a name that is easily linkable to the poster’s offline identity, such as their legal name. A real name account builds a reputation over time based on its posts. In addition, a real name account’s reputation draws on their offline reputation, and affects their offline reputation in turn.
    Pseudonymous accounts are accounts which are not easily linkable to the poster’s offline identity, and which the poster maintains over time. A pseudonym builds a reputation over time based on its posts. This reputation is separate from the poster’s offline reputation.
    Burner accounts are accounts which are intended to be used for a single, transient purpose and then abandoned. They accrue little or no reputation.
    Anonymous posts are not traceable to a specific identity at all. This forum mostly doesn’t have anonymous posts and so I will not discuss them here.
    All of these accounts have some legitimate uses. Because of the differences in how these types of accounts operate, readers should evaluate their claims differently, especially when it comes to evaluating claims about the community. Posters should use accounts appropriate for the points they are making, or restrict their claims to those which their account can support.
    Arguments, Evidence, and Accountability
    When it comes to abstract arguments, the content can be evaluated separately from the speaker, so all this stuff can be disregarded. If someone on this forum wants to post a critique of the statistics used in vitamin A supplementation trials, or an argument about moral status of chickens, or something like that, then the poster’s reputation shouldn’t matter much, and so it’s legitimate to post under any type of account. When 4chan solved an open combinatorics problem while discussing a shitpost about anime, mathematicians accepted the proof and published it with credit to "Anonymous 4chan poster". When it comes to abstract arguments, anything goes, except for blatant fuckery like impersonation or sockpuppet voting.
    If someone wants to claim expertise as part of an argument, then it helps to demonstrate that expertise somehow. If someone says “I’m a professional statistician and your statistical analysis here is nonsense”, then that rightly carries a lot more weight if it’s the real-name account of a professional statistician, or a pseudonymous account with a demonstrable track record on the subject. Burner accounts lack reputation, track records, and credentials, so they can’t legitimately make this move unless they first demonstrate expertise, which is often impractical.
    Things get trickier when it comes to reporting facts about the social landscape. The poster’s social position is a legitimate input into evaluating such claims. If I start telling everyone about what’s really happening in Oxford board rooms or Berkeley group houses, then it matters a great deal who I am. Am I a veteran who’s been deep inside for years? A visitor who attended a few events last summer? Am I just repeating what I saw in a tweet this morning?
    Advantages of Real Name Accounts
    Real name accounts can report on social situations with authority that other types of account can’t legitimately claim, for two reasons. First, their claims are checkable. If I used this pseudonymous account to make a f...
    11 min
  • LW - The Translucent Thoughts Hypotheses and Their Implications 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: The Translucent Thoughts Hypotheses and Their Implications, published by Fabien Roger on March 9, 2023 on LessWrong.
    Epistemic status: Uncertain about the validity of the claims I’m making here, and looking for feedback about the research directions I’m suggesting.
    Thanks to Marius Hobbhahn, Johannes Treutlein, Siméon Campos, and Jean-Stanislas Denain for helpful feedback on drafts.
    Here is a set of hypotheses:
    The first AGIs will have LLMs at their core
    Effective plans to defeat humanity can’t be found in a single LLM forward pass
    LLMs will solve complex tasks by using English text (self-prompting, scratch pads, combination of expert LLMs, .)
    I call these the Translucent Thoughts hypotheses.
    I think the Translucent Thoughts hypotheses are likely (around 20% conditioning on AGI before 2030) because:
    Text pretraining is more efficient at building algorithms and knowledge required for real-world plan generation and evaluation than alternative methods;
    Future models are likely to be like Transformers, which use a limited amount of serial step in a single forward pass, and deception requires many serial steps;
    Text pretraining and slight fine-tuning makes model able to use text generation to increase the maximum number of serial steps by a huge factor. Getting this increase through other means is likely to be hard and not competitive.
    If these hypotheses are true, it should lead us to prioritize underexplored research directions, such as circumventing steganography or building extremely reliable text-supervision methods. I think those deserve attention, because Translucent Thoughts AIs are not safe by default.
    In this post, I argue that we may will in a world where the first AGIs will look like X, and I then describe ways to make the first AGIs safer given X. This is different from most other works in this space, which often directly describe a kind of safe AGI. Despite this, the ideas of this post are close to some other works describing paths to safe AGIs, such as:
    Externalized Reasoning Oversight, which describes a class of solutions similar to the one outlined here, but also aims for additional properties which I argue can be replaced with a less stringent hypothesis about AI systems;
    Conditioning Predictive Models, which makes assumptions slightly different from the Translucent Thoughts hypotheses, yielding different research directions;
    The Open Agency Model and Factored Cognition which describe subsets of AIs with Translucent Thoughts, which might be safe.
    The Translucent Thoughts Hypotheses
    Here, I sketch a world in which the first AGIs have certain properties. I argue that this world is likely, and thus a subset of all possible futures to care about. But I think it’s not a large part of all possible futures (20% conditioning on AGI before 2030).
    The First AGIs Will Have LLMs at Their Core
    By “first AGIs” I mean the first systems able to automate all cognitive tasks.
    AGI is likely to do reasoning and planning using LLMs. AGI might rely on vision models for some tasks and interactions with the world, and it might use explicit search processes like AlphaGo. But I expect LLMs to do plan generation and evaluation, which are the core of the system (from an Alignment point of view).
    Why: Vision systems are bad at coming up with and evaluating deceptive plans. Explicit search processes can’t generate and evaluate plans in the real world. LLMs seem to be able to do both plan generation and evaluation. (Plan generation and evaluation are the core tasks we would like to monitor to make AGIs safe, which is why I focus on those.)
    End-to-end neural networks won’t be able to compete with LLMs when it comes to reasoning and planning, or at least, end-to-end networks will use “their LLMs parts” to do their most advanced form of reasoning and planning. This means that systems lik...
    35 min
  • LW - Why Not Just Outsource Alignment Research To An AI? by johnswentworth
    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: Why Not Just Outsource Alignment Research To An AI?, published by johnswentworth on March 9, 2023 on LessWrong.
    Warmup: The Expert
    If you haven’t seen “The Expert” before, I recommend it as a warmup for this post:
    The Client: “We need you to draw seven red lines, all strictly perpendicular. Some with green ink, some with transparent. Can you do that?”
    (... a minute of The Expert trying to explain that, no, he cannot do that, nor can anyone else.)
    The Client: “So in principle, this is possible.”
    This. This is what it looks like in practice, by default, when someone tries to outsource some cognitive labor which they could not themselves perform. At best, The Expert is well-intentioned and knows what the user needs, ignores the incoherent parts of The Client’s babbling, and does the right thing. Or, they manage to add some silly but ultimately harmless bells and whistles to satisfy whatever dumb thing The Client is looking for.
    At worst. well, there’s more than one failure mode which could qualify for the title of "worst". Maybe The Expert gives The Client something which looks right to The Client and successfully conceals all the problems with it; presumably that’s a lucrative strategy for Experts. Maybe the Double Illusion of Transparency kicks in, both parties think they’ve successfully communicated, but in fact neither has any idea what’s going on in the other’s head. Maybe a well-intentioned Expert decides to ignore The Client’s incoherent babbling and do the thing which seems most likely to be right, but gets The Client’s preferences wrong.
    One way or another, The Client’s ignorance is a major bottleneck to cognitive outsourcing. In practice, I expect The Client’s ignorance to be the primary bottleneck to cognitive outsourcing.
    The core reason why we cannot just outsource alignment research to an AI is because we would then be The Client, and probably a very ignorant one.
    Application to Alignment Schemes
    There’s a lot of different flavors of “have the AI solve alignment for us”. A sampling:
    Just prompt a language model to generate alignment research
    Do some fine-tuning/RLHF on the language model to make it generate alignment research
    Let the language model talk to other instances of itself, and prompt or fine-tune them together so they generate alignment research jointly
    Set up a language model to generate alignment proposals and another to poke holes in them, and fine-tune the pair via a human judging the “debate”
    As we go down the list, the proposals get fancier and add more bells and whistles, trying to make the AI a better expert. Sadly, none of them at all address what I expect to be the actual main bottleneck: The Client (i.e. the human user or users) has no understanding of what they need, what questions to ask, what’s possible or even logically coherent, etc.
    What would this kind of error look like in practice?
    Here’s one concrete example of the kind of failures I’d expect when a would-be outsourcer’s understanding falls short (from here):
    Somebody literally types “If we take the action you just proposed, will we be happy with the outcomes?” into a GPT prompt. Obviously that does not result in the AI giving its actual best-guess answers to the questions, but in this case it doesn't result in the AI thinking about how to deceive humans either. It just thinks about what text would follow that question if it appeared on the internet somewhere. And then I imagine someone with a bunch of interpretability tools saying "yup, it's just thinking about what text typically follows this question", and then that person's boss is like "great, it's not trying to deceive us, guess we can trust the answer", and they both just haven't really thought of the fact that the AI's response-text does not have anything in particular to do with whether the AI is aligned or whether t...
    15 min
  • AF - Why Not Just Outsource Alignment Research To An AI? by johnswentworth
    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: Why Not Just Outsource Alignment Research To An AI?, published by johnswentworth on March 9, 2023 on The AI Alignment Forum.
    Warmup: The Expert
    If you haven’t seen “The Expert” before, I recommend it as a warmup for this post:
    The Client: “We need you to draw seven red lines, all strictly perpendicular. Some with green ink, some with transparent. Can you do that?”
    (... a minute of The Expert trying to explain that, no, he cannot do that, nor can anyone else.)
    The Client: “So in principle, this is possible.”
    This. This is what it looks like in practice, by default, when someone tries to outsource some cognitive labor which they could not themselves perform. At best, The Expert is well-intentioned and knows what the user needs, ignores the incoherent parts of The Client’s babbling, and does the right thing. Or, they manage to add some silly but ultimately harmless bells and whistles to satisfy whatever dumb thing The Client is looking for.
    At worst. well, there’s more than one failure mode which could qualify for the title of "worst". Maybe The Expert gives The Client something which looks right to The Client and successfully conceals all the problems with it; presumably that’s a lucrative strategy for Experts. Maybe the Double Illusion of Transparency kicks in, both parties think they’ve successfully communicated, but in fact neither has any idea what’s going on in the other’s head. Maybe a well-intentioned Expert decides to ignore The Client’s incoherent babbling and do the thing which seems most likely to be right, but gets The Client’s preferences wrong.
    One way or another, The Client’s ignorance is a major bottleneck to cognitive outsourcing. In practice, I expect The Client’s ignorance to be the primary bottleneck to cognitive outsourcing.
    The core reason why we cannot just outsource alignment research to an AI is because we would then be The Client, and probably a very ignorant one.
    Application to Alignment Schemes
    There’s a lot of different flavors of “have the AI solve alignment for us”. A sampling:
    Just prompt a language model to generate alignment research
    Do some fine-tuning/RLHF on the language model to make it generate alignment research
    Let the language model talk to other instances of itself, and prompt or fine-tune them together so they generate alignment research jointly
    Set up a language model to generate alignment proposals and another to poke holes in them, and fine-tune the pair via a human judging the “debate”
    As we go down the list, the proposals get fancier and add more bells and whistles, trying to make the AI a better expert. Sadly, none of them at all address what I expect to be the actual main bottleneck: The Client (i.e. the human user or users) has no understanding of what they need, what questions to ask, what’s possible or even logically coherent, etc.
    What would this kind of error look like in practice?
    Here’s one concrete example of the kind of failures I’d expect when a would-be outsourcer’s understanding falls short (from here):
    Somebody literally types “If we take the action you just proposed, will we be happy with the outcomes?” into a GPT prompt. Obviously that does not result in the AI giving its actual best-guess answers to the questions, but in this case it doesn't result in the AI thinking about how to deceive humans either. It just thinks about what text would follow that question if it appeared on the internet somewhere. And then I imagine someone with a bunch of interpretability tools saying "yup, it's just thinking about what text typically follows this question", and then that person's boss is like "great, it's not trying to deceive us, guess we can trust the answer", and they both just haven't really thought of the fact that the AI's response-text does not have anything in particular to do with whether the AI is aligned...
    15 min