Ideas for improving communications

About

What: I want our communications to be more successful in being better at convincing more people of the risk, convincing them that supporting a pause is good for their concerns too, and in reaching wider audiences.
Why: Convincing people of the existence of serious risks is hard. Making it easier gets more support for AI risk mitigation. Convincing people more concerned with different risks to join forces is a force multiplier and can benefit everyone, and more concerned people helps in being taken seriously.

1. This is not a project proposal.
2. This is a collection of ideas on how to improve our communications.
3. For some of these ideas, it would be a good idea to create a dedicated project to implement that idea.
4. For some of these ideas, it might be better to take them into account in other projects, existing or new.
5. Some of these ideas might be bad in practice, even if they might sound good at first.
6. I can't commit to any specific involvement in this. But optimizing communications is important, so I hope others can start working on at least parts of it.
7. Ideas not guaranteed to be original or even mine.

How you can help

* Evaluate these ideas
* Level of involvement: Tiny to big
* Skills: none / think / write / research / marketing / lobbying / outreach
* Please consider and discuss which of these ideas are worthwhile, and whether to incorporate them in existing projects or spin up new ones. Remember that sometimes "there are people who want to work on this" trumps "this is more effective in theory". Scope might range from writing a single comment here, to doing serious research on the efficacy of various approaches.
* Coordinate efforts at improving communications
* Level of involvement: Big
* Skills: Think / write / project management
* Make sure good ideas from here are used. Try to balance effectiveness and feasibility. See evaluation task description.

primary-violetOP•6/4/24, 12:34 PM

I wanted to get these out quickly and not get bogged down thinking about which might already have been presented or discussed or what would be the ideal presentation format, and I ended up highjacking a specific outreach project with this

. I'm copypasting them here now, which may explain odd references to what other people said there.

primary-violetOP•6/4/24, 12:36 PM

This is relevant, but also to that other project, so link instead of removing original: Recruiter team

primary-violetOP•6/4/24, 12:36 PM

So, ideas related to optimizing communications is what initially lead me to start taking action. I have limited bandwidth, I'm not a great communicator myself, and there's probably loads of people with better communication skills and experience or familiarity with at least some areas of communication optimization. And I would expect such people to have better ideas or even experience in how to organize such efforts, which is why I think it's better for someone else to write up project proposals for these things. But just in case I thought of something others haven't already, I'll post my thoughts on the subject here.

ETA: These are complex topics, and I'm kind of used to thinking about some of them in terms of concepts, which might not be commonly used outside of some circles, such as https://www.lesswrong.com/tag/inferential-distance , which is one important factor in difficulty of successful communication about complex ideas. So, that's a useful concept to keep in mind. Ask me to clarify if I use other weird terms and fail to explain them

Inferential Distance - LessWrong

Inferential Distance between two people with respect to an item of knowledge is the amount of steps or concepts a person needs to share before they can successfully communicate the object level point. This can be thought of as the missing foundation or building block concepts needed to think clearly about a specific thing.

In Expecting Short In...

primary-violetOP•6/4/24, 12:36 PM

1. Firstly, poor framing of the problem is kind of dangerous, because it carries the risk of you, and the whole concept as being written off as overblown panicking over something trivial or mostly irrelevant, or as an apocalypse cult at worst, and that can be hard to change later.
2. So, these are complex topics, and yet, there is an incredibly wide range of levels of knowledge at which people think that they can successfully reason about AI risks, and there is a huge variety of ways in which people form mental models of AI systems, which result in all kinds of combinations of ideas about various AI risk related topics, but mostly they lead people to dismiss many significant classes of safety concerns. So, if you want to be successful at convincing people that certain risks need to be taken seriously, you need to be able to model how they think about these risks in order to successfully explain some aspects of AIs, such as why they are extremely alien to humans despite all appearances, to identify key mistaken assumptions that lead to false conclusions about safety and how to best poke holes in those assumptions. Failure in this makes the results poor choice of framing even worse.
3. This makes talking about the risk of unaligned superintelligence especially hard, as there are often very long inferential distances to be crossed in order to even being able to start talking about the same thing, and I think fears related to these kinds of failure modes have seriously limited people's willingness to even try explaining existential risks more widely.
4. Also, a related important concept is the Overton window, which basically means the range of ideas which it is widely acceptable to discuss and also be taken seriously. Global catastrophic risk from AI in general has only recently entered the Overton window, and unaligned superintelligence did it even later. But the Overton window is not the same everywhere. This is an important consideration when doing wider outreach..

primary-violetOP•6/4/24, 12:37 PM

... Especially outside of anglosphere.
5. Considering these things together, I think it might be, as a general rule, better to lead wider outreach efforts with focus on other, more easily understood risks, of which there are many.
6. Another important strategic consideration in communication is that even if you think that all other risks are trivial considering the risk of unaligned superintelligence, making more people aware of, and willing to act on other risks will help in establishing regulations and other measures of risk mitigation, some of which can also help mitigate risk of unaligned superintelligence. And if outreach focusing on these other risks is more likely to succeed in getting people and governments to act, that could well make a bigger impact on x risk than outreach focusing on just that. Also, it makes more people learn more about AI risks more broadly and broadens the Overton window, thereby paving the way for introducing the idea of unaligned superintelligence and other x risks.

primary-violetOP•6/4/24, 12:37 PM

7. So, a framing and messaging guide, or rather a set of them for various target audiences should be created, and also variations focusing on different risks.
8. Because of (2), I think that something that probably goes far beyond typical framing and messaging guides might be warranted; a comprehensive guide to various failure modes in thinking about AI risks, and best practices in addressing each of them.
9. In addition to such guides, some kind of an outreach coaching program might be a good idea.
10. Additionally, high profile safety advocates would benefit from professional presentation coaching.

primary-violetOP•6/4/24, 12:38 PM

11. Guides for how to address those who think the risk is real, but for various reasons don't want to get behind a pause(divided efforts are weak, pause has a lot of momentum, has widespread support among experts, there are actual international talks about it between countries to at least some extent, even though a six month pause is insufficient, it would set a precedent for at least temporary bans on some research, it would create oversight and enforcement mechanisms which should help with faster action in the future if a new pause is needed, and maybe there is even a tiny chance that it could be extended before it is over, it would send an unambiguous signal that this is serious business, and just give more time to consider things, probably drive up funding for safety research etc. So many benefits, even if it does not directly impact your pet risk, or is not the best theoretically possible mitigation for it.)
12. Guides for how to address opposition to various other risk mitigation strategies: These can quickly get more complex or speculative, as the exact details of some risk mitigation measures might be speculative, which makes it even more speculative to say how the downsides of those mitigation measures might be addressed: But having ideas for possible solutions might go a long way towards making the idea palatable, and that could also be presented as an opportunity to influence the solution. Notable examples of possibly important mitigation measures that might get a lot of opposition from both open source and entrepreneurial people, about which I have some ideas are: (1) cryptographic controls on AI accelerators: see https://discord.com/channels/1100491867675709580/1245353843655708753/1245396193823948951 , (2) liability: see https://discord.com/channels/1100491867675709580/1119539399034683462/1246816780375162992 , (3) strict oversight requirements in general.

primary-violetOP•6/4/24, 12:39 PM

It's getting late, so I'll try to do this bit quickly. Ask for clarifications later if I condense it too much to make sense. So, a bunch of reasons for "why you should support a pause even if you don't believe in superintelligence", related especially to 6, 7 and 11 above:

13. Why you should take superintelligence seriously even if you don't believe in it. This is a difficult one, because people fail at understanding what probability even means. Also, because people are used to seeing hype bubbles. And because AI, like fusion is always 20 years away. But, generally, outside view and averaging experts are better predictors than individual experts. However, wrt. recent progress, experts failed by miserably underestimating how smart AI systems we could have at this hardware level and how quickly that could change. That should cause you to think that there is a much higher chance of superintelligence being possible in general, and that it could also be real a lot sooner than you thought. Since probability is hard, an oversimplified analogy might be helpful, such as Russian roulette. Average expert predictions are roughly in that ballpark for superintelligence / global catastrophic risk in a few years, I think? Would you play Russian roulette for getting a million a year? Would you be comfortable with Sam Altman or Mark Zuckerberg making that choice for you? For every single human on the planet?
14. The right to self determination: If the world collectively makes an informed decision to accept the risk, that might be OK, but it's definitely not OK for a minority to make this decision for everyone. The whole world should be in on it. Also, even if you discount the possibility of superintelligence, there might be pretty significant social upheaval and other risks. Again, mitigating these should be a global effort, involving representation from all affected. For some audiences, emphasizing this view might motivate them to take action, e.g. in helping global outreach.

primary-violetOP•6/4/24, 12:39 PM

15. Almost all of humanity is at risk of economic marginalization, even if we pause. But pause would slow that change, giving society more time to prepare. This is probably hard for most to believe based on how even the top models fail on some pretty simple tasks. That makes them forget that they still far outperform most humans in a lot of areas. And most people don't realize that such possibilities as fine tuning models, augmenting them with all kinds of plug-ins, including other, more specialized neural networks, rags, combining them in multi agent systems, where e.g. different models with different strengths and weaknesses could give feedback to each other to keep each other on track, etc. And we're just starting to explore the possibilities of such systems and how to make the most out of them. And of course the pause would not prevent training of very powerful specialized systems. Especially with the development of more powerful models blocked, there would be a huge shift in incentives to make the most use out of the above possibilities. Which means that there is a a high probability that maybe even pretty quickly we would start seeing all kinds of agentic AI systems that are optimized for various fields, and which can perform the vast majority of tasks in those fields with sufficient reliability that much fewer humans are needed, to supervise them or to perform some subset of tasks which the AI still can't handle well enough. Predicting which fields will be affected and how much is hard, so this should concern almost everyone. Still, a pause should give more time before this starts, and the change would be slower, giving more time for societies to adapt.

primary-violetOP•6/4/24, 12:40 PM

16. Potentially useful subproject: Refuting individuals. Of course we should avoid stuff that looks like ad hominems, or painting individuals as bad guys, for lots of reasons, but a series of articles of the type: "yeah, this guy is very smart, insightful and right about a lot of things, but here's why he is wrong about x". I got this idea from listening to a debate between Roman Yampolskiy and Robin Hanson, and Hanson is a good example of this class of problems in convincing people about risk / inadequacy of certain safety measures. But there's probably loads of people that match this kind of pattern but in different fields, and very different ideas or political leanings. He is just an example that I know. There are multiple significant aspects of this: (1) People might think that they are not competent to evaluate the issue themselves, so they defer to others, whom they think they can trust. But that person might not be an expert in this area, or, they might have a good track record of predicting or modeling things in other fields, and a very convincing theory of why they can do that, it might not be applicable here, so we should explain why that approach breaks down, or how a certain assumption that person makes is wrong, and how that leads them to wrong conclusions, despite sound methods. (2) Hanson is also an example of a successful contrarian. He has a lot of good ideas which don't seem widely accepted by mainstream economists, but which many smart people find obviously true or useful in retrospect, and consider lack of widespread adoption a sign of societal insanity. For example, futarchy, which is the use of prediction markets to make policy decisions. Of course there are failure modes, such as use of predictions as assassination contracts, and a full blown futarchy could backfire in unexpected ways, but instead of increasing research into it and how to counter failure modes, US just blanket banned all prediction markets, which is mind blowingly stupid...

primary-violetOP•6/4/24, 12:40 PM

... (3) Contrarianism in general appeals to a lot of smart people, as they have personal experiences and read about smart people and smart ideas not being taken seriously. And this can lead to a pattern of being too skeptical about mainstream opinions among experts, and attaching to contrarians in the field, or commenting on the field in question. I have no idea how much this kind of pattern might contribute to people dismissing risks, but it might be worth looking in to.

primary-violetOP•6/4/24, 12:45 PM

Link relevant to improving conversation level argumentation: https://aisafety.info/ , HT @Max Winga

AISafety.info

AI Safety FAQ

primary-violetOP•6/4/24, 12:48 PM

Oops, out of order, in reply to Recruiter team

@Max Winga :

Yes, I think @hirvinen does a great job breaking down important updates to communication that we need. I think one additional area we need to focus on is who/what people think of when they think of AI safety. Getting a bit abstract, many people, when you ask them what they think of a particular movement, will have some imagined person that is the spokesperson for that movement. Right now, I fear that the wall of propaganda and e/acc "doom cult" language is setting that default image of us in people's minds as disconnected techbro's holding up "the end is near" cardboard signs. This is made worse by people who post overly emotional responses to AI leaders (I'm thinking of this one person who responded to Altman on Twitter with full caps saying he's "actually the most evil person" or something along those lines.) I fully believe that the stakes are truly all or nothing, but we need to court the public with rational calm discussions about the serious nature of AI risk.

And my reply to it:

Ah, yes, that is a very good point. I was just getting into the topic of "why people who aren't worried about x risk should also support a pause", but I had only thought about it in terms of getting more people to support it, not about how it would be beneficial to have pausing be supported by people who don't believe in superintelligence.

dead-brown•6/4/24, 5:41 PM

@DuncanHaldane I think that this would be excellent.

dead-brown•6/4/24, 5:41 PM

I'll link your comments so far here

dead-brown•6/4/24, 5:41 PM

DuncanHaldane — Today at 11:38 AM

Could use more orientation here. From that positioning of your most important problem I notice three avenues of action:
Go talk to cybersecurity advisors and figure out what messaging lands, with who, and why. Secondary - see if the existing cybersecurity advisors will start pushing parts of your message for you.
See if transparency and more data on Anthropic's lobbying activities can be used as a channel to alienate their safety conscious talent and potential talent. What is Anthropic saying to policy makers OTR? Dario's showing was really weak on safety, so there might be meat here.
Educate AI engineers on how to make resignations that are legible to congress, and cause the big red button push. (I think someone is already on this part, but you can probably amplify).

For the purposes of this discord, what do I do with these ideas?

Pprimary-violet 15. Almost all of humanity is at risk of economic marginalization, even if we pa...

primary-violetOP•6/4/24, 7:29 PM

Nvidia just made a big spiel about this; Paraphrasing; "We make it easy for you to chain together a bunch of specialized AI models to create digital humans. For example easy replacement of customer service workers, like nurses that don't do prescriptive or diagnostic work, retail, financial services, insurance. You bring in a bunch of experts(the different AIs) and build a team. You don't even have to figure out how to connect them, you just give the task to an agent." https://youtu.be/PM6dRb8ttzQ?t=11m55s

YouTubeWes Roth

NVIDIA CEO Jensen Huang Reveals AI Future: "NIMS" Digital Humans, W...

Learn AI With Me:
https://www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.

#ai #openai #llm

dead-brown•6/4/24, 7:33 PM

Can we make a short for it?

Ddead-brown Can we make a short for it?

primary-violetOP•6/4/24, 7:49 PM

I don't have relevant experience, but we probably have here a lot of people who do. This bit is just a few minutes in a much longer video that mainly focuses on other things. There might be better videos covering that presentation as a basis for a short.

There is also a bit later a section where he talks about a virtual cycle. While he doesn't mention the term, what he means is "we are the drivers of singularity"...

dead-brown•6/4/24, 7:50 PM

We need a social media effort to be on tiktok. We really need it.

primary-violetOP•6/4/24, 8:31 PM

17. Tools to help crafting messages to various groups, such as influencers, interest group representatives, again, not just about existential risk. ( see Ideas for improving communications and Ideas for improving communications). A few different approaches:
- Just write a bunch of templates. Easy to get started, but not flexible, and more work in the long term. Also, people might not be happy with the existing templates, and get stuck tweaking them.
- A simple app that would allow you to customize the message by selecting which topics to include, and each topic might even have a few alternative wordings from which to pick. With a bit of forethought, not much more work than static templates. Could easily be implemented in a spreadsheet as message fragments with tick boxes next to them. With a spreadsheet approach, it would also be easy for more people to collaborate on the texts.
- next level: Same as before, but instead of message fragments, you click which topics you want covered, how much to focus on each topics, which aspects of those topics should be covered, etc. And the output is a prompt that you can give to an LLM of your choice.

colossal-harlequin•6/5/24, 2:29 AM

Thanks @hirvinen that's a great list. I have some additions to suggest...

colossal-harlequin•6/5/24, 2:29 AM

A1. Despite its "doomer" associations, cope provocation and necessarily somewhat high inferential distance cost, emphasizing "this can kill you and everybody you care about" still has bleeding obvious alpha we may have gotten too used to, so to be stupidly explicit: once sufficiently believed, it is more motivating than any other risk can be. This is a superpower. It can dwarf all other pushback and habits.

colossal-harlequin•6/5/24, 2:33 AM

A2. Another framing of all other more concrete risks: examples of those risks manifesting are themselves the mostly missing warning shots for x-risk: if this unprecedented tech makes surprising A, B and C happen, it can make X happen too.

colossal-harlequin•6/5/24, 2:38 AM

A3. We can and should usefully a/b test our choices re type and content of messages. With experiments establishing both framework and tech platform have initial costs but together they provide game changing value.

Ccolossal-harlequin A1. Despite its "doomer" associations, cope provocation and necessarily somewhat...

primary-violetOP•6/5/24, 9:10 AM

Related to making the point about x risk:
A point which Connor Leahy made in some interview. I'm paraphrasing this from memory, so don't attribute this as a quote from him:

AGI is an extremely vague term. When you say that artifical general intelligence means that it can do anything that humans can do, people will think of something more like a friendly robot buddy that might know a lot, but is not really that much smarter than you. But what these huge Silicon Valley tech companies mean by it is an AI that can do original science, including making itself smarter, which would quickly lead to creation of a godlike intelligence, but also that they have no idea how to control something like that. Whether or not you think they can succeed, that's what they believe, and they have gone on record saying all of these things, and they also believe that they could succeed in just a few years. And that's when you get a response "Hell no, that's crazy! Where is the government?!"

primary-violetOP•6/6/24, 10:41 PM

Do we have people from MIRI in this discord, or people who are in direct contact with them? Some of these things are relevant to their new strategy to focus on communication, and they stated that they are going to do related stuff, like introductory materials, and want data on which arguments and materials worked etc.

Pprimary-violet 11. Guides for how to address those who think the risk is real, but for various ...

primary-violetOP•6/7/24, 1:42 PM

I just realized, that (12), "How to respond to opposition against mitigation measures" is probably way more important than I previously thought:

If you think "preventing possible AI doom requires giving up too many certain / probable good things x, y and z", that creates a really strong emotional barrier / disincentive to accepting even "AI doom is a serious risk" let alone "risk of AI doom requires drastic action".

Because of that, it might be a good idea in at least some discussions to preemptively present the idea:
* "we think there are many ways in which we could prevent AI doom without giving up good things x, y and z, such as..."
* and even better: "and they could also help with these other risks too, also without giving up all good things"

dead-brown•6/7/24, 4:42 PM

What are the common good things? e.g. one thought I had is "health care" and we can show that even with current models, we can get substantial benefit from AI models.

primary-violetOP•6/7/24, 5:30 PM

Do you mean "good things that AI could offer, which some people might think we would lose due to risk mitigation measures"? That is a good question. And the answers, and even more so, the answers to "how we might get them despite mitigation measures. " But that discussion should not happen in this channel, imo. And we can't even create threads here... Maybe under #policy-

, or maybe under #general-

¯\_(ツ)_/¯

Pprimary-violet Do we have people from MIRI in this discord, or people who are in direct contact...

dual-salmon•6/7/24, 8:14 PM

I got in touch via email with Gretta at MIRI. She is in this server already but keeps the notifications off because it's a firehose of info. I also know Nate Soare at MIRI. I don't believe he is in this discord.

I suggested to Gretta she might want to use our discord to help beta test some "rock" style content, and she seemed receptive to the idea. So there does seem to be potential for collaboration.

Ddual-salmon I got in touch via email with Gretta at MIRI. She is in this server already but ...

primary-violetOP•6/8/24, 10:58 AM

I sent her a DM on LW about potential for collaboration if we could manage to get some of these projects going, or volunteers to help with their projects.

dual-salmon•6/9/24, 2:55 PM

Hacker News was pitched to me as a place it might be high impact to try to publish in

colossal-harlequin•6/9/24, 3:59 PM

I want us to run all kinds of tests including AB experiments, across all media. It's something I have experience of in online merchandising at least. Many big tech successes originate in what we call weblabs at Amazon.

The indefatigable @Jonathan Bostock is going to simply get on in there and start testing conversational gambits - which is great - but I'd love for us to also build something bigger and more general for testing our and MIRI's content.

Ccolossal-harlequin I want us to run all kinds of tests including AB experiments, across all media. ...

brilliant-lime•6/9/24, 4:02 PM

Point one is find a reliable way to talk to a person for ~1 minute about AI risk. Then I can start testing messaging strategies.

primary-violet•6/10/24, 9:40 PM

Hi @hirvinen, you lured me in here with your message on LW. I quickly skimmed the contents of this channel but there's a lot here and I'm not actually sure which part is the proposal. If you can send just the proposal in a google doc I can take a look. gretta@intelligence.org.

Thanks!

Pprimary-violet Hi @hirvinen, you lured me in here with your message on LW. I quickly skimmed th...

primary-violetOP•6/11/24, 8:57 AM

Hi @grettazon,

(also mailing this)
No detailed collaboration project proposal yet. Summary of which of these I think would be most important for MIRI to keep in mind, considering your strategy follows. Some are not exactly central to focusing on x risk from SI, but are important for interviews and debates.

1. Finding best ways to poke holes in faulty assumptions. https://aisafety.info/ is kind of on this already.
2. Making a case for "why you should want to pause/halt even though you personally / not all subject matter experts are convinced of danger"
3. Making a case for "we can take drastic against rogue AGI projects / malicious use without giving up good things like beneficial weaker AI systems or open source"
4. Best ways to frame your arguments.
5. Professional presentation coaching.
6. PauseAI might be a source of volunteers that could help you in your projects, possibly allowing much larger project scope.
7. People here might be starting projects whose results could help you, such as A/B testing of framing / arguments. Consider keeping an eye out for that and/or making suggestions for which kind of community projects you think would be useful. I can also try to let you know if a project relevant to you gets a lot of traction.

Thanks for taking the time. What I posted in this channel is not yet very highly organized, because I started only recently to take action, and my time is limited, sorry... At this stage, it is a list of what I think are important strategic considerations, arguments for them, identifying failure modes and some ideas how to address them. Posted here hoping that it could help others focus their efforts better. I might write detailed project proposals for some later, but might not have the time.

AISafety.info

AI Safety FAQ

colossal-harlequin•6/11/24, 1:06 PM

@grettazon as per my LW comment on the Execution section: I get the need for a high bar, but my guess is MIRI might try to grow ten times faster than the post indicates. More dakka: more and better content. We aren't short of green field opportunity. If the Pause AI community could provide necessary funding or writing/testing resources, would you be open to dialing the effort up like that?

Pprimary-violet Hi @grettazon, (also mailing this) No detailed collaboration project proposal y...

primary-violet•6/11/24, 1:25 PM

Thanks, replying to this in email is better for me so I'll handle it there.

Pprimary-violet Thanks, replying to this in email is better for me so I'll handle it there.

primary-violetOP•6/11/24, 1:26 PM

You don't need to reply, unless you want to ask for clarification or something, or would like help with something specific(I can't commit to much, but asking is okay.)

Ccolossal-harlequin @grettazon as per my LW comment on the Execution section: I get the need for a h...

primary-violet•6/11/24, 1:27 PM

Scaling a team is hard, especially if you want to keep the work high quality and cohesive. It's even harder when it's primarily made of volunteer labor. I think you and I have pretty different intuitions about how to proceed here. I do appreciate you taking the time to engage on it.

I think we may come to a point where we're able to harness a large quantity of volunteer labor but we're not there yet. I will keep the PauseAI community in mind.

Pprimary-violet Scaling a team is hard, especially if you want to keep the work high quality and...

primary-violetOP•6/11/24, 1:31 PM

Underrating management is easy. Hiring a (part time) manager might be a better way to start scaling rather than more content producers / devs / whatever in general. (No, don't take this as something you should feel the need to reply to, and it is not a comment on how you do things, just a general notion.)

primary-violet•6/11/24, 1:35 PM

Oh, believe me, I'm well aware of that and working hard to make it happen!

primary-violet•6/11/24, 1:35 PM

(I have experience managing and scaling large teams)

Pprimary-violet Scaling a team is hard, especially if you want to keep the work high quality and...

colossal-harlequin•6/11/24, 4:31 PM

Understood. My intuition for building some AB testing mechanism and employing it at scale comes from Amazon - who I just quit in favor of Pause AI. I'll continue thinking on it. Anyone at MIRI with related ideas is welcome to engage with me.

worthy-azure•6/24/24, 4:31 PM

Tagging @Jonathan Bostock and @Joseph Miller in this project thread.

Reviewing the notes from the PauseAI retreat, I saw you were both interested in a project around testing & optimizing messaging for PauseAI. Seems like a really great initiative.

How has this been going? I thought you might want to get plugged into this thread and use it as a space to collaborate with others.

Let me know if you have any questions or if there's anything I can do to be useful.

brilliant-lime•6/26/24, 1:30 PM

I've asked MIRI and CEA if they've done any work on this. MIRI haven't but are about to, CEA didn't get back. I think Joseph is working on a paper for the next month and a half, so doesn't have much free time. I'm going to try and analyse a bunch of online discussions regarding AI to see which arguments get the most positive responses. I would also like to try just talking to lots of people in series and analyse that data, but I don't know an optimal way to do that. Could go out on the streets but I would like another person. Could try text-based studies on MTurk but I expect that's exactly what MIRI is doing and MTurk isn't the best representative population sample.

primary-violet•6/26/24, 3:03 PM

(MIRI here) Text-based studies on MTurk doesn't actually sound all that promising to me; it'd be fine as one element in a portfolio of testing but I don't expect to get the audiences we think are most important participating in MTurk.

primary-violet•6/26/24, 3:04 PM

I too have wanted to "try just talking to lots of people" and that project has been awaiting staffing for a while; we might get there soon. Our concept there is to do hour-long structured interviews but it takes a skilled interviewer to get meaningful data and we have not had luck with hiring a firm to help us with that; they tend to insist on doing the interviews themselves, and we don't think they'll do a good job.

primary-violet•6/26/24, 3:05 PM

Overall it's hard, it takes really quite a lot of thought and effort to get right. Work in progress.

Ideas for improving communications

About

How you can help

Similar Threads

Similar Threads

Similar Threads