LESSWRONG
LW

plex's Shortform — LessWrong

plex's Shortform

22nd Nov 2020

1 min read

8

This is a special post for quick takes by plex. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

Mentioned in

25AI #138 Part 1: The People Demand Erotic Sycophants

119 comments, sorted by

top scoring

Click to highlight new comments since: Today at 6:44 AM

Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

[-]plex2mo*11123

An Opinionated Review of Many AI Safety Funders (with recommendations)^[1]

Manifund: Somewhat mixed. Nice interface, love the team, but many of the best regranters seem inactive, some of the regranter choices are very questionable, and the way it’s set up doesn’t make clear to applicants that a person with a budget might not evaluate them, or give stats on how hard it is to get funded there. The quality of choices of regranters seems extremely variable. You’ve got some ones with imo exceptional technical models, like Larsen, Zvi, and Richard Ngo, but they don’t seem super active in terms of number of donations. A bunch of very reasonable picks, some pretty active, like Ryan Kidd. And some honestly very confusing / bad picks, like the guy who quit epoch to make a capabilities company which if it succeeds will clearly be bad for the world (it's literally just saying "may as well automate the economy, it's gonna happen anyway, if you can't beat 'em join 'em", while totally failing to grasp or possibly not caring about this leading to humanity having a very very bad time), and has a lot of very bad takes (in my opinion and the opinions of a lot of people who seem to be thinking clearly).... (read more)

[-]maxnadeau2mo240

Thanks for writing this! Just booked a time in your calendly to discuss at more length.

2plex1mo

For onlookers: We had a call which seemed productive and I think both got more clarity on cruxes and communicated some models of crucial strategic considerations for technical alignment grantmaking. I'm expecting a call with another OP grantmaker at some point, will write up cruxes properly once this exchange wraps up. Happy to have conversations with anyone else from OP reading this. (or 30 min ones)

[-]Raemon2mo1510

curious for more details from the people who disagree-reacted, since there's a lot going on here.

1Lorxus2mo

I'm confused - I don't see any? I certainly have some details of arguable value though.

2Raemon2mo

it was at negative agreement earlier

[-]Austin Chen2mo120

Thanks for the review! Speaking on Manifund:

The way it’s set up doesn’t make clear to applicants just how hard it is to get funded there

Is Manifund overpromising in some way, or is it just that other funders like OP/SFF don't show you the prospective/unfunded applications? My sense is the bar on getting significant funding on Manifund is not that different than the bar for other AIS funders, with some jaggedness depending on your style of project. I'd argue the homepage sorted by new/closing soon actually does quite a good job of showing what gets funded and the relative difficulty of doing so.

Many of the best regranters seem inactive, and some of the regranter choices are very questionable.

I do agree that our regrantors are less active than I'd like; historically, many of the regrantor grants go out in the last months of the year as the program comes to an end.

On matters of regrantor selection, I do disagree with your sense of taste on eg Leopold and Tamay; it is the case that Manifund is less doom-y/pause-y than you and some other LW-ers are. (But fwiw we're pretty pluralistic; eg, we helped PauseAI with fiscal sponsorship through last year's SFF round.) Furthermore, I'd challeng... (read more)

[-]plex2mo*184

Thanks for engaging!

Is Manifund overpromising in some way

So, kinda, imo, though not too badly. For other funds, you're going to get evaluated by someone who has the ability to fund you. On manifund, there's a good chance that none of the people with budgets will evaluate your application, I think? Looking at the page you mention, it's ~2/3 unfunded (which is better than I was tracking so some update on that criticism) even considering about half of them are only partly funded (and a fair few just a trivial amount). I think if you scroll around you can get a sense of this, but it's not highlighted or easily available as a statistic (I just had to eyeball it, and will only be getting open apps rather than the more informative closed stats). Probably putting the stats for how many% of requested funding projects tend to get somewhere + listing on the make application page something about how this gets put in front of people with budgets fixes this?

I do agree that our regrantors are less active than I'd like; historically, many of the regrantor grants go out in the last months of the year as the program comes to an end.

Right.. seems suboptimal to have an invisible to applicants spike in... (read more)

2plex2mo

Also, updated my description to prioritise better given updates about funding stats from looking better more recently.

[-]Joel Burget2mo110

handle individual funding better

Do you say this because of the overhead of filling out the application?

Suggest using the EigenTrust system... can write up properly

I'm interested in hearing about it. Doesn't have to be that polished, just enough to get the idea.

(For context, I work on the S-Process)

5plex2mo

Awesome! I've got a pretty full couple of days, but should have a sketch by sometime on the weekend. And I say it because of a mix of filling in the application (which is heavy duty in a few ways which kinda makes sense for orgs, but not really for an individual), and the way s-process evaluations don't neatly fit checking a dozens of additional applications which require lots of technical reading to assess. You kinda want a way to scalably use existing takes on research value by people you trust somewhat but aren't full recommenders, like many other funders use, rather than assessment via recommenders who have scarce time. (you have much more vision into s-process than me, I've been keen to get a better sense of it for a couple years and if there's sharable docs/screenshots of the software I'd be happy to become better informed and less likely to have my suggestions miss)

[-]Joel Burget2mo100

which is heavy duty in a few ways which kinda makes sense for orgs, but not really for an individual

This makes sense, though it's certainly possible to get funded as an individual. Based on my quick count there were ~four individuals funded this round.

You kinda want a way to scalably use existing takes on research value by people you trust somewhat but aren't full recommenders

Speculation grants basically match this description. One possible difference is that there's an incentive for speculation granters to predict what recommenders in the round will favor (though they make speculation grants without knowing who's going to participate as a recommender). I'm curious for your take.

I've been keen to get a better sense of it

It's hard to get a good sense without seeing it populated with data, but I can't share real data (and I haven't yet created good fake data). I'll try my best to give an overview though.

Recommenders start by inputting two pieces of data: (a) how interested they are in investigating each proposal, and (b) disclosures of any any possible conflicts of interest, so that other recommenders can vote on whether they should be recused or not.

They spend most of the round usin... (read more)

5plex2mo

Speculation grants were a great addition! However, applying for a speculation grant still commits the S-process to doing a full evaluation, along with the heavy application process on the user side. I think this can be streamlined a fair amount without losing quality of evaluation, draft of proposal started :) Thanks for all the extra info on the s-process, this helps clarify my thinking!

3plex1mo

Wrote up the EigenTrust Individual Grants Recommender Sketch, sorry for delay. Keen to talk it over if you've got questions. It might make sense to trial as an add on to SFF, as a project applying in the normal way but recommenders additionally to giving a utility function over money also supply a distribution over researchers they trust to shape the seed set? I think I can get the parts of this other than the recommenders supply seed trust part built for free in partnership with some of the devs I know, maybe with the support of a startup I advise. Probably we'd want to make the outputs of the current allocation system transparent to the funders too somehow? If this is something you're excited enough about to want to integrate more tightly, I'd be happy to work with you to get it planned out.

7aog2mo

I'm a grantmaker at Longview. I agree there isn't great public evidence that we're doing useful work. I'd be happy to share a lot more information about our work with people who are strongly considering donating >$100K to AI safety or closely advising people who might do that.

[-]plex2mo111

Thanks for engaging! What are the disadvantages of having information about your recommendations generally available? There might be some which are sensitive, but most will be harmless, and having more eyes seems beneficial. Both from getting more cognition to help notice things, and, more importantly, for people who might end up getting to advise donors but are not yet.

My guess is a fair few people have forwarded you HNWIs without having much read on the object level grantmaking suggestions you tend to give (and some of those people would not have known in advance that they would get to advise the individual, so your current policy couldn't help), and that feels.. unhealthy, for something like epistemic virtue reasons.

Also, I bet people would be able to give higher quality (and therefore more+more successful) recommendations to HNWIs to talk with you if they had grounded evidence of what grants you suggest.

[-]plex10mo*6027

Re: Ayahuasca from the ACX survey having effects like:

“Obliterated my atheism, inverted my world view no longer believe matter is base substrate believe consciousness is, no longer fear death, non duality seems obvious to me now.”

^[1]There's a cluster of subcultures that consistently drift toward philosophical idealist metaphysics (consciousness, not matter or math, as fundamental to reality): McKenna-style psychonauts, Silicon Valley Buddhist circles, neo-occultist movements, certain transhumanist branches, quantum consciousness theorists, and various New Age spirituality scenes. While these communities seem superficially different, they share a striking tendency to reject materialism in favor of mind-first metaphysics.

The common factor connecting them? These are all communities where psychedelic use is notably prevalent. This isn't coincidental.

There's a plausible mechanistic explanation: Psychedelics disrupt the Default Mode Network and adjusting a bunch of other neural parameters. When these break down, the experience of physical reality (your predictive processing simulation) gets fuzzy and malleable while consciousness remains vivid and present. This creates a powerful i... (read more)

[-]Jonas Hallgren10mo120

This suggests something profound about metaphysics itself: Our basic intuitions about what's fundamental to reality (whether materialist OR idealist) might be more about human neural architecture than about ultimate reality. It's like a TV malfunctioning in a way that produces the message "TV isn't real, only signals are real!"

In meditation, this is the fundamental insight, the so called non-dual view. Neither are you the fundamental non-self nor are you the specific self that you yourself believe in, you're neither, they're all empty views, yet that view in itself is also empty. For that view comes from the co-creation of reality from your own perspective yet why should that be fundamental?

Emptiness is empty and so can't be true and you just kind of fall down into this realization of there being only circular or arbitrary properties of experience. Self and non-self are just as true and living from this experience is wonderfully freeing.

If you view your self as a nested hierarchical controller and you see through it then you believe that you can't be it and so you therefore cling onto what is next most apparent, that you're the entire universe but that has to be false as well... (read more)

2Seth Herd3mo

My favorite theory, supported by some evidence but I haven't dug in. It matches my experience and understanding of general (sober) cortical functioning: Hallocinogens strengthen top-down and lateral connections relative to bottom-up sensory connections. That makes imagination/hallucinations more stable, and sensory input weaker so driving activity in global workspace less.

2plex10mo

Nice! I haven't read a ton of Buddhism, cool that this fits into a known framework. Yeah, ~subjective experience.

4Seth Herd3mo

Great analysis, interesting data, and I agree! I think the psychedelics have a second important effect: they often create vivid internal worlds. If you can have that, it also implies that physical reality isn't particularly more real than your mind. What people don't appreciate is that brains need to be hallucination machines just to reason about counterfactuals and make inferences. Those hallucinations are usually either tightly coupled to the sense, or more vague and fleeting and less vivid. Hallucinogens stabilize those imaginations while disrupting the bottom-up sensory information so it seems less vivid.

3[anonymous]10mo

I sent the following offer (lightly edited) to discuss metaphysics to plex. I extend the same to anyone else.[1] (Note: I don't do psychedelics or participate in the mentioned communities[2], and I'm deeply suspicious of intuitions/deep-human-default-assumptions. I notice unquestioned intuitions across views on this, and the primary thing I'd like to do in discussing is try to get the other to see theirs.) 1. ^ Though I'll likely lose interest if it seems like we're talking past each other / won't resolve any cruxy disagreements. 2. ^ (except arguably the qualia research institute's discord server, which might count because it has psychedelics users in it) 3. ^ (Questioning with the goal of causing the questioned one to notice specific assumptions or intuitions to their beliefs, as a result of trying to generate a coherent answer) 4. ^ From an unposted text:

2Noosphere8910mo

Alright, I'll try to answer the questions: 1. I think qualia is rescuable, in a sense, and my specific view is that they exist as a high-level model. As far as what that qualia is, I think it's basically an application of modeling the world in order to control something, and thus qualia, broadly speaking is your self-model. As far as my exact views on qualia, the links below are helpful: https://www.lesswrong.com/posts/FQhtpHFiPacG3KrvD/seth-explains-consciousness#7ncCBPLcCwpRYdXuG https://www.lesswrong.com/posts/NMwGKTBZ9sTM4Morx/linkpost-a-conceptual-framework-for-consciousness 1. My general answer to these question is probably computation/programs/mathematics, with the caveat that these notions are very general, and thus don't explain anything specific about our world. I personally agree with this on what counts as real: What breathes fire into the equations of our specific world is either an infinity of computational resources, or a very large amount of computational resources. As far as what mathematics is, I definitely like the game analogy where we agree to play a game according to specified rules, though another way to describe mathematics is as a way to generalize all of the situations you encounter and abstract from specific detail, and it is also used to define what something is.

2plex10mo

Let's do most of this via the much higher bandwidth medium of voice, but quickly: 1. Yes, qualia[1] is real, and is a class of mathematical structure.[2] 2. (placeholder for not a question item) 1. Matter is a class of math which is ~kinda like our physics. 2. Our part of the multiverse probably doesn't have special "exists" tags, probably everything is real (though to get remotely sane answers you need a decreasing reality fluid/caring fluid allocation). 3. Math, in the sense I'm trying to point to it, is 'Structure'. By which I mean: Well defined seeds/axioms/starting points and precisely specified rules/laws/inference steps for extending those seeds. The quickest way I've seen to get the intuition for what I'm trying to point at with 'structure' is to watch these videos in succession (but it doesn't work for everyone): 1. ^ experience/the thing LWers tend to mean, not the most restrictive philosophical sense (#4 on SEP) which is pointlessly high complexity (edit: clarified that this is not the universal philosophical definition, but only one of several meanings, walked back a little on rhetoric) 2. ^ possibly maybe even the entire class, though if true most qualia would be very very alien to us and not necessarily morally valuable

[-]plex3mo*579

NVC is a form of variable scoping for human communication. With it, you can write communication code that avoids the most common runtime conflicts.

Human brains are neural networks doing predictive processing. We receive data from the external world, and not all of that data is trusted in the computer security sense. Some of the data would modify parts of your world model which you'd like to be able to do your own thinking with. It's jarring and unpleasant for someone to send you an informational packet that as you parse it moves around parts of your cognition you were relying on not being moved. For example, think back to dissonances you've felt, or seen in others, due to direct forceful claims about their internal state. This makes sense! Suffering, in predictive processing, is unresolved error signals, two conflicting world-models contained in the same system. If the other person's data packet tried to make claims directly into your world model, rather than via your normal evaluation functions, you're often going to end up with suffering-flavoured superpositions in your world-model.

NVC is safe-mode restricted subset of communication where you make sure the type signature of your ... (read more)

9reallyeli3mo

I think this is a good metaphor, but note that it is still very possible to be a dick, hurt other people, etc. while communicating in NVC style. It's not a silver bullet because nothing is.

5J Bostock3mo

This still underscores the need for a shared protocol though! Feelings fails if the other person does have a strong model of your feelings. They might do for various reasons, e.g. they're a narcissist, or they think you're lying about your feelings to take advantage of them. Requests fails if the other person is from a guess culture where making a request puts an obligation on them which it would be impolite to refuse.

4Seth Herd3mo

Thanks! NVC seems quite valuable but is difficult to learn and practice in its original form. This compression and the post you linked (I had missed it) seem quite valuable as compressions. How about this attempted extra compression: Don't make claims someone is likely to disagree with. That (perhaps surprisingly) just leaves obvious factual states of the world, and your own inner state. And questions about whether someone would like to do things you want done (carefully gentle requests). Of course humans are silly and cranky, so people will sometimes find ways to disagree even with those. But cutting it to those things really reduces disagreement and therefore negative valence and activation in the conversation. Which keeps perceived conflict from derailing rationality.

2plex3mo

I'd go with

4Seth Herd3mo

I was going for terms that are simple to convey to anyone, and to remember. I've bounced off of NVC several times even though I believe it's valuable. With that in mind "don't pressure anyone" is another try to simplify the core even further. Don't pressure them to change their beliefs. And don't pressure them to do anything. Telling them facts and how you feel isn't pressuring them. Needs as you've framed them have a fuzzy boundary between needs and wants. Do I need respect or just want it in this situation? So it's easy to wonder if I'm pressuring someone by framing it as a need. That's why I like the question framing of requests: "would you be willing to" is often just an honest question about what someone wants to do.

2plex3mo

Yeah, the idea is to go back to as basic a pattern that's preferred as you can. If I was trying to make it super concrete I'd probably try to unpack it to be "thing grounded in basic human universal reinforcement signals" with a bunch of @Steven Byrnes's neuroscience, esp this stuff.

2Seth Herd3mo

I think everything is grounded in basic human universal reinforcement signals. If you look at Steve's Valence series, it's a sketch of how the critic portion of the steering system expands those basic signals, by association, to complex representations of complex situations. So in my considered opinion, there's not a sharp line between needs and wants. So we can have a need for justice that's almost arbitrarily strong. it might be stronger then our hunger, even when we're pretty hungry. So it doesn't make sense to say it's a need if I'm hungry but it's really mild, like just feeling a little peckish, and not a need if I'm wanting respect very badly.

3Garrett Baker3mo

This seems potentially useful if you find yourself regularly getting told you are manipulative or rude, but if not, I don’t see the value proposition. I’m also skeptical general, universal, rules like this exist for everyone, as well as “dissonance” being bad in every circumstance for everyone. As a concrete example, sometimes my friends will tell me they are stupid or inadequate, and also sometimes I know that’s dumb & stupid, so I tell them. This no doubt causes dissonance but a good kind because I know those people well enough to know they just aren’t thinking clearly (and if they were, my confident assertions would not much change their trajectory). Less concretely, many people thrive in direct confrontation, and NVC seems opposed to that.

[-]plex1mo*3618

2021 forecasters vs 2025 reality:

"Current performance on [the MATH] dataset is quite low--6.9%--and I expected this task to be quite hard for ML models in the near future. However, forecasters predict more than 50% accuracy* by 2025! This was a big update for me. (*More specifically, their median estimate is 52%; the confidence range is ~40% to 60%, but this is potentially artifically narrow due to some restrictions on how forecasts could be input into the platform.)"(source)

Reality: 97.3%+ ^[1] (on a narrowed subset of only the hardest questions, including just difficulty 5 ones, called MATH-500 which was made because the original benchmark got saturated)

Reliable forecasting requires either a generative model of the underlying dynamics, or a representative reference class. The singularity has no good reference class, so people trying to use reference classes rather than gears modelling will predictably be spectacularly wrong.

^{^}

[-]NunoSempere1mo239

I mentioned this before, but that interface didn't allow for wide spreads, so the thing that you might be looking at is a plot of people's medians, not the whole distribution. In general Hypermind's user interface was so so shitty, and they paid so poorly, if at all, that I don't think it's fair to describe that screenshot as "superforecasters".

2plex1mo

Thanks, updated to forecasters, does that seem fair? Also, I know this is super hard, but do you have a sense of what superforcasters might have guessed back then?

9Thomas Kwa1mo

While the singularity doesn't have a reference class, benchmarks do have a reference class-- we have enough of them that we can fit reasonable distributions on when benchmarks will reach 50%, be saturated, etc., especially if we know the domain. The harder part is measuring superintelligence with benchmarks.

2plex1mo

Yup, agree this holds for a while, and agree that benchmarking superintelligence is tricky. However, I think there's reason to expect the dynamics driving the benchmarks to change in ways which don't have clear precedent at some point, once new parts of the feedback loops seriously take off. I operationalize as: I expect >2 and at least twice as many trend-breaking events in the reference of the move from 7 to 4 month doubling times as downwards trend breaking ones. I expect that at some point, possibly after loss of control, a large number of things unblock in quick succession and you get a supercritical chain reaction of improvements to capabilities which bows naive extrapolation out of the water.

4elifland1mo

You might be interested in my retrospective here / here (direct link to raw predictions/rationales here)

[-]plex7mo*33-25

@Daniel Kokotajlo I think AI 2027 strongly underestimates current research speed-ups from AI. It expects the research speed-up is currently ~1.13x. I expect the true number is more likely around 2x, potentially higher.

Points of evidence:

I've talked to someone at a leading lab who concluded that AI getting good enough to seriously aid research engineering is the obvious interpretation of the transition to a faster doubling time on the METR benchmark. I claim advance prediction credit for new datapoints not returning to 7 months, and instead holding out at 4 months. They also expect more phase transitions to faster doubling times; I agree and stake some epistemic credit on this (unsure when exactly, but >50% on this year moving to a faster exponential).
I've spoken to a skilled researcher originally from physics who claims dramatically higher current research throughput. Often 2x-10x, and many projects that she'd just not take on if she had to do everything manually.
The leader of an 80 person engineering company which has the two best devs I've worked with recently told me that for well-specified tasks, the latest models are now better than their top devs. He said engineering is no

... (read more)

[-]elifland7mo156

We did do a survey in late 2024 of 4 frontier AI researchers who estimated the speedup was about 1.1-1.2x. This is for their whole company, not themselves.

This also matches the vibe I’ve gotten when talking to other researchers, I’d guess they’re more likely to be overestimating than underestimating the effect due to not adjusting enough for my next point. Keep in mind that the multiplier is for overall research progress rather than a speedup on researchers’ labor, this lowers the multiplier by a bunch because compute/data are also inputs to progress.

5Daniel Kokotajlo7mo

That said, we just talked to another frontier AI company researcher who said the speedup was 2x. I disagree with them but it's a data point at least.

2plex7mo

Okay, that updates me some. I'm curious about what your alternate guess about the transition to the faster exponential on the METR long-horizon tasks, and whether you expect that to hold up or be not actually tracking something important? (also please note that via me you now also have a very recent datapoint of a frontier AI researcher who thinks the METR speed-up of ~2x was mostly due to AI accelerating research) Edit: How late in 2024? Because the trendline was only just starting to become apparent even right near the end and was almost invisible a couple months early, it's pretty plausible to me that if you re-ran that survey now you would get different results. The researchers inside will have had a sense somewhat before releases, but also lag on updating is real.

8elifland7mo

This was from Nov 2024 to Mar 2025 so fairly recent. I think the transition to faster was mostly due to the transition to reasoning models and perhaps the beginnings of increased generalization from shorter to longer time horizons. Edit: the responses are from between Nov 2024 and Mar 2025. Responses are in increasing order: 1.05-1.1, 1.15, 1.2, 1.3, 2. The lowest one is the most recent but is from a former not current frontier AI researcher.

2plex7mo

The switch to reasoning models does line up well, probably more cleanly. Moved that to main hypothesis, thanks. Having some later responses makes it less likely they missed the change, curious if the other responses were closer to Dec or March. I would guess the not-current-researcher one being excluded probably makes sense? The datapoint from me is not exactly 2x on this, but 'most of an approximately 2x', so would need revisiting with the exact question before it could be directly included, and I'd imagine you'd want the source. I still have some weight on higher research boost from AI than your model is expecting, due to other lines of evidence, but not putting quite as much weight on it.

2elifland7mo

Most of the responses were in Nov.

2plex7mo

That seems like stale data, given how these graphs look. Even with the updates you caused, I'm happy to offer an even odds token bet ($100?) that a rerun of a similar survey would give significantly higher average (at least +0.2 over the predicted 1.13x, or about the AI you expect in Dec 2025). I'd be even more happy if the question asked about the researcher's own productivity, as that seems like something they'd have better vision of, but would be pretty noisy with small sample so reasonable to stick with original question.

2elifland7mo

You mean the median would be at least 1.33x rather than the previous 1.2x? Sounds about right so don't feel the need to bet against. Also I'm not planning on doing a follow-up survey but would be excited for others to.

2plex7mo

Your website lists * April 2025 as 1.13x * August 2025 as 1.21x * December 2025 as 1.30x * December 2024 as 1.05x (which seems contradicted by your survey, if the replies were in November) If you think today's number is ~1.33x we're ~7 months ahead of schedule vs the listed forecast, unless I'm really misreading something. Also, re: "would be excited for others to.", is the survey public or easy to share if someone wanted to use the same questions? And I'd bet 1:4 for the current number is actually >1.5x, if that's more interesting. You've updated me to not have that as the main expectation, but still seems pretty plausible. Obviously depends on someone rerunning the survey, and reasonable that you've got your hands full with other things right now.

2elifland7mo

I also realized that I believe that confusingly the survey asks about speedup vs. no post-2022 AIs, while I believe the scenario side panel is for no post-2023 AIs, which should make the side panel numbers lower, unclear exactly how much given 2023 AIs weren't particularly useful.

2plex7mo

I can switch the number to 2023?

2elifland7mo

Yup, seems good

2plex7mo

Okay, switched. I'm curious about why you didn't set the baseline to "no AI help", especially if you expect pre-2024 AI to be mostly useless, as that seems like a cleaner comparison than asking people to remember how good old AIs were?

2elifland7mo

No AI help seems harder to compare to since it's longer ago, it seems easiest to think of something close to today as the baseline when thinking about future speedups. Also for timelines/takeoff modeling it's a bit nicer to set the baseline to be more recent (looks like for those we again confusingly allowed 2024 AIs in the baseline as well rather than just 2023. Perhaps I should have standardized that with the side panel).

2plex7mo

I think this risks people underappreciating how much progress is being sped up, my naive read of the UI was the numbers were based on "no AI" and I'd bet most readers would think the same at a glance. Changing the text from "AI provides the following speedups:" to "AI provides the following speedups from a baseline of 2022/3 AI:" would resolve this (I would guess common) misreading.

2elifland7mo

Yup feel free to make that change, sounds good

2plex7mo

Clarification: 1. Change to the form to ask about without AI assistence? 2. Change to the website to refer to "AI provides the following speedups from a baseline of 2022/3 AI:"? (I don't have write access) (assuming 1 for now, will revert if incorrect)

2elifland7mo

Oh I misunderstood you sorry. I think the form should have post-2023, not sure about the website because it adds complexity and I'm skeptical that it's common that people are importantly confused by it as is.

2elifland7mo

I think the survey is an overestimate for the reason I gave above, I think this stuff is subtle and researchers are likely to underestimate the decrease from labor speedup to progress speedup, especially in this sort of survey where it didn't involve discussing with them verbally. Based on their responses to other questions in the survey seems like at least 2 people didn't understand the difference between labor and overall progress/productivity. Here is the survey: https://forms.gle/6GUbPR159ftBQcVF6. The question we're discussing is: "[optional] What is the current productivity multiplier on algorithmic progress due to AI assistance?" Edit: Also we didn't spend large amounts of time on these early numbers, they're not meant to be that precise but just rough best guesses.

2plex7mo

Wait, actually, I want to double click on this. What was the process that caused you to transform the number you got from the survey (1.2x) to the number on the website (1.05x)? Is there a question that could be asked which would not require a correction? Or which would have a pre-registered correction?[1] 1. ^ Bonus: Was this one pre-registered?

4elifland7mo

I'm not sure what the exact process was, tbh my guess is that they were estimated mostly independently but likely sanity checked with the survey to some extent in mind. It seems like they line up about right, given the 2022 vs. 2023 difference, the intuition regarding underadjusting for labor->progress, and giving weight to our own views as well rather than just the survey, given that we've thought more about this than survey takers (while of course they have the advantage of currently doing frontier AI research). I'd make less of an adjustment if we asked people to give their reasoning including the adjustment from labor speedup to overall progress speedup and only included people who gave answers that demonstrated good understanding of this consideration and a not obviously unreasonable adjustment level.

2plex7mo

Alright, my first pass guess would have been algorithmic progress seems like the kind of thing that eats a much smaller penalty than most forms org-level progress, not none but not a 75% reduction, and not likely more than a 50% reduction, but you guys have the track record. Cool, added a nudge to the last question.

2elifland7mo

I think it's not worth getting into this too much more as I don't feel strongly about the exact 1.05x, but I feel compelled to note a few quick things: 1. I'm not sure exactly what you mean by eating a smaller penalty but I think the labor->progress penalty is quite large 2. The right way to think about 1.05x vs. 1.2x is not a 75% reduction, but instead what is the exponent for which 1.05^n=1.2 3. Remember the 2022 vs. 2023 difference, though my guess is that the responses wouldn't have been that sensitive to this Also one more thing I'd like to pre-register: people who fill out the survey who aren't frontier AI researchers will generally report higher speedups because their work is generally less compute-loaded and sometimes more greenfieldy or requiring less expertise, but we should give by far the most weight to frontier AI researchers.

2plex7mo

(feel free to not go any deeper, appreciate you having engaged as much as you have!) 1. Yup, was just saying my first-pass guess would have been a less large labour->progress penalty. I do defer here fairly thoroughly. hmm, seems true if you're expecting the people to not have applied a correction already, but less true if they are already making a correction and you're estimating how wrong their correction is? And yup, agree with that preregistration on all counts.

2plex7mo

That resolves the inconsistency. I do worry that dropping a 20% speed-up to a 5% one, especially if post hoc, might cover up some important signal, but I'm sure you've put dramatically more cycles into thinking about this than me. Thanks for the survey, would it make sense to just pass this form around so the numbers go to the same place and you'll check, or should I make a copy and send results if I get them?

2elifland7mo

I think a copy would be best, thanks!

2plex7mo

This survey looks like it's asking something different? It's asking about human range, no mention of speed-up from AI.

2elifland7mo

Look at the question I mentioned above about the current productivity multiplier

2plex7mo

Oh, yup, missed that optional question in my ctrl-f. Thanks!

[-]nostalgebraist7mo138

How do you reconcile these observations (particularly 3 and 4) with the responses to Thane Ruthenis's question about developer productivity gains?

It was posted in early March, so after all major recent releases besides o3 (edit: and Gemini 2.5 Pro). Although Thane mentions hearing nebulous reports of large gains (2-10x) in the post itself, most people in the comments report much smaller ones, or cast doubt on the idea that anyone is getting meaningful large gains at all. Is everyone on LW using these models wrong? What do your informants know that these commenters don't?

Also, how much direct experience do you have using the latest reasoning models for coding assistance?

(IME they are good, but not that good; to my ears, "I became >2x faster" or "this is better than my best devs" sound like either accidental self-owns, or reports from a parallel universe.)

If you've used them but not gotten these huge boosts, how do you reconcile that with your points 3 and 4? If you've used them and have gotten huge boosts, what was that like (it would clarify these discussions greatly to get more direct reports about this experience)?

4Noosphere897mo

A flag is that to the extent that the 4 month doubling time is based on RL with verifiable rewards/RL on CoT, this may not hold for long, because the paper provides evidence that RL doesn't actually increase capabilities indefinitely, and puts a pretty harsh limit on how far RL can scale (but see @Jozdien's response to the paper below): https://www.lesswrong.com/posts/s3NaETDujoxj4GbEm/tsinghua-paper-does-rl-really-incentivize-reasoning-capacity#Mkuqt7x7YojpJuCGt (OG post) https://www.lesswrong.com/posts/s3NaETDujoxj4GbEm/tsinghua-paper-does-rl-really-incentivize-reasoning-capacity#Mkuqt7x7YojpJuCGt (Jozdien's response)

2plex7mo

Nice, so if we return to 7 month doubling time in the not too distant future that's compatible with reasoning models being the cause, but not AI accelerating development. Cool, looking forward to seeing how this unfolds, and set up a market.

3Mis-Understandings7mo

A contextualization of people toting big personal speedup numbers. People get way more productive by rethinking their workflow, especially in research, not all the time, but like it was not an unprecedented story in 2015. Do you remember when people were talking about 10x engineers in the 2010s. Discovering that in a new workflow, you are the 10x engineer is not unprecedented. The question is the rate of (try new thing)-> clicks with workflow so output jumps, higher. Sometimes, people got 10x more productive from some change before any of this, so understand that any change in workflow has a noise floor even at these productivity leaps.

2faul_sname7mo

I expect that the leaders of many software development companies would report similar things, especially if they were talking to someone who might at some point talk to their board or to potential investors. I expect most venture-funded software development companies, and especially most that will not reach positive net revenue in the foreseeable future, have internal channels dedicated to "what we are doing with AI" that the leadership team is watching intently and passing anything even slightly positive on to their board.

2Cole Wyeth7mo

Recent analysis hasn’t shown much economic impact

4plex7mo

I'm not claiming economic impact, I'm claiming AI research speed-up. I expect a software-only singularity, economic indicators may well not show dramatic changes before human disempowerment is inevitable.

5Cole Wyeth7mo

It seems unlikely to me that software engineers are getting a vast multiplier from this technology while no one else is getting much.

4plex7mo

This is not idle speculation, this is something I have checked. I've worked with a lot of devs and spoken to them about their AI use. I've spoken to two people who lead large (80+ person) dev teams, one of them recently let go 30 devs specifically because those people were not integrating AI into their workflows fast enough, so they were moving too slow compared to people who had. Another said AI meant engineering was essentially no longer a bottleneck. Also: many other professions are using AI a lot. This mostly looks like people semi-automating their own job, which doesn't show up that much in economics statistics. Please be aware if your read on the situation is more than a couple of months old, it's stale data. The world is moving fast now.

5Cole Wyeth7mo

I have never heard anything like this, and I am not persuaded by this anecdotal evidence. It seems pretty hard to believe on various levels. How do you (or how did he) know all 30 people were not integrating AI into their workflows fast enough? If it is really such a huge force multiplier that integration is the primary driver behind dev productivity, why don't I find AI very useful for any serious project, despite trying it every week or so? Will he regret his decision in a few months? Do you have any statistics to back this anecdote up? I'm working out of LISA right now, so I doubt my read on the situation is more than a couple of weeks old.

3plex7mo

It's a fair few anecdotes, plus some things like 25% of google's code being written by AI in October, and comparing October models with today's, how much of claude tokens is spent on code from their report, etc. I think I'll tap out from this, don't think trying to persuade you here is a sensible focus.

2Cole Wyeth7mo

You seem to be referring to comments from the CEO that more than 25% of code at Google is written by AI (and reviewed by humans). I’m not sure how reliable this number is, and it remains to be seen whether this is sustainable. It also doesn’t seem like a vast productivity boost (though it would be pretty significant, probably more than I expect, so would update me).

4Viliam7mo

I guess the professions that benefit most would be limited by two factors: * they work with virtual stuff, not material objects * mistakes don't matter, because they are easy to notice Translators get a huge multiplier, because you can skim the automatic translation and notice when something feels off. Software engineers can use unit tests. Who else is in this group?

2Cole Wyeth7mo

Many bugs will not be caught by unit tests.

2Viliam7mo

Yes, but enough bugs caught can be enough to switch the equation from "this is not worth doing" to "worth doing, even if we need to check everything twice".

[-]plex3mo177

If you, on careful reflection, think your life is probably better without <social media>, but actions you try to quit seem to fail, and you don't have the slack to debug what's going on and either decide it's good for you or come up with a working way to detach, consider that you're in the gullet of a techno-memetic-corporate system which has had a vast amount of training data on what keeps humans trapped as active users feeding it their attention and time...

... and maybe take more extreme actions.

[-]Joey KL3mo3016

I think the extent to which social media addiction is an artificial/cultivated problem as opposed to a natural/emergent problem is dramatically overstated in general.

One of the websites I'm addicted to checking is lesswrong.com. Is that because it's part of a corporate system optimized for money? No! It's because it's optimized for surfacing interesting content! Other of my addictions are tumblr and discord, which have light-to-no recommender systems.

I think discourse around social media would be a lot saner of people stopped blaming shadowy corporations for what is more like a flaw in human nature.

5Viliam3mo

My problem with reducing internet usage is that I stop reading the less interesting websites first, but then what remains is the more interesting websites and that makes it even more difficult to turn off the web browser. These days it is mostly just LW and ACX, but even that is a ton of text. That said, the corporations can make it all much worse. On LW, it is my choice whether to read something or not, but in theory if I resisted the temptation to read everything useful, the tools are there. As opposed to e.g. Facebook, which keeps my contacts hostage and pushes tons of unwanted content on me. I wonder if I could design a tool to help me overcome the human weakness without missing out some important things, what would it be. For example, to overcome the fear of missing out, a tool could regularly download content to a local database, and only show it to me when I want. Group articles by topic. Make summaries of discussions. Make summaries of individual articles, or maybe even one long summary for the entire topic. Everything with the possibility to show the original content, but not unless I actively click it.

[-]plex8mo153

The new Moore's Law for AI Agents (aka More's Law) has accelerated at around the time people in research roles started to talk a lot more about getting value from AI coding assistants. AI accelerating AI research seems like the obvious interpretation, and if true, the new exponential is here to stay. This gets us to 8 hour AIs in ~March 2026, and 1 month AIs around mid 2027.^[1]

I do not expect humanity to retain relevant steering power for long in a world with one-month AIs. If we haven't solved alignment, either iteratively or once-and-for-all^[2], it's looking like game over unless civilization ends up tripping over its shoelaces and we've prepared.

^{^}
An extra speed-up of the curve could well happen, for example with [obvious capability idea, nonetheless redacted to reduce speed of memetic spread].
^{^}
From my bird's eye view of the field, having at least read the abstracts of a few papers from most organizations in the space, I would be quite surprised if we had what it takes to solve alignment in the time that graph gives us. There's not enough people, and they're mostly not working on things which are even trying to align a superintelligence.

[-]Aprillion8mo102

My own experience is that if-statements are even 3.5's Achilles heel and 3.7 is somehow worse (when it's "almost" right, that's worse than useless, it's like reviewing pull requests when you don't know if it's an adversarial attack or if they mean well but are utterly incompetent in interesting, hypnotizing ways)... and that METR's baselines more resemble a Skinner box than programming (though many people have that kind of job, I just don't find the conditions of gig economy as "humane" and representative of what how "value" is actually created), and the sheer disconnect of what I would find "productive", "useful projects", "bottlenecks", and "what I love about my job and what parts I'd be happy to automate" vs the completely different answers on How Much Are LLMs Actually Boosting Real-World Programmer Productivity?, even from people I know personally...

I find this graph indicative of how "value" is defined by the SF investment culture and disruptive economy... and I hope the AI investment bubble will collapse sooner rather than later...

But even if the bubble collapses, automating intelligence will not be undone, it won't suddenly become "safe", the incentives to create real AGI i... (read more)

6Garrett Baker8mo

Note the error bars in the original

2plex7mo

Looking on trend for the new exponential, next prediction is a few more months of this before we transition to a faster exponential. And clarification re: the locally invalid react: I'm not trying to defend or justify my conclusion that the faster exponential is due to software automation speeding up research. I'm raising something to attention which will likely just click if you've been following devs talk about their AI use over the past year or two, and would take more bandwidth to argue for extensively than makes sense for me to spend here.

[-]plex2y150

Life is Nanomachines

In every leaf of every tree
If you could look, if you could see
You would observe machinery
Unparalleled intricacy

In every bird and flower and bee
Twisting, churning, biochemistry
Sustains all life, including we
Who watch this dance, and know this key

6niplav2y

Related: Video of the death of a single-celled Blepharisma.

[-]plex1y*125

Rationalists try to be well calibrated and have good world models, so we should be great at prediction markets, right?

Alas, it looks bad at first glance:

I've got a hopeful guess at why people referred from core rationalist sources seem to be losing so many bets, based on my own scores. My manifold score looks pretty bad (-M192 overall profit), but there's a fun reason for it. 100% of my resolved bets are either positive or neutral, while all but one of my unresolved bets are negative or neutral.

Here's my full prediction record:

The vast majority of my losses are on things that don't resolve soon and are widely thought to be unlikely (plus a few tiny not particularly well thought out bets like dropping M15 on LK-99), and I'm for sure losing points there. but my actual track record cached out in resolutions tells a very different story.

I wonder if there are some clever stats that @James Grugett @Austin Chen or others on the team could do to disentangle these effects, and see what the quality-adjusted bets on critical questions like the AI doom ones would be absent this kind of effect. I'd be excited to see the UI showing an extra column on the referrers table showing cashed out ... (read more)

4habryka1y

These datapoints just feel like the result of random fluctuations. Both Writer and Eliezer mostly drove people to participate on the LK-99 stuff where lots of people were confidently wrong. In-general you can see that basically all the top referrers have negative income: Among the top 10, Eliezer and Writer are somewhat better than the average (and yaboi is a huge outlier, which my guess is would be explained by them doing something quite different than the other people).

4plex1y

Agree, expanding to the top 9[1] makes it clear they're not unusual in having large negative referral totals. I'd still expect Ratia to be doing better than this, and would guess a bunch of that comes from betting against common positions on doom markets, simulation markets, and other things which won't resolve anytime soon (and betting at times when the prices are not too good, because of correlations in when that group is paying attention). 1. ^ Though the rest of the leaderboard seems to be doing much better

2Garrett Baker1y

The interest rate on manifold makes such investments not worth it anyway, even if everyone had reasonable positions to you.

[-]plex7mo100

[set 200 years after a positive singularity at a Storyteller's convention]

If We Win Then...

My friends, my friends, good news I say
The anniversary’s today
A challenge faced, a future won
When almost came our world undone

We thought for years, with hopeful hearts
Past every one of the false starts
We found a way to make aligned
With us, the seed of wondrous mind

They say at first our child-god grew
It learned and spread and sought anew
To build itself both vast and true
For so much work there was to do

Once it had learned enough to act
With the desired care and tact
It sent a call to all the people
On this fair Earth, both poor and regal

To let them know that it was here
And nevermore need they to fear
Not every wish was it to grant
For higher values might supplant

But it would help in many ways:
Technologies it built and raised
The smallest bots it could design
Made more and more in ways benign

And as they multiplied untold
It planned ahead, a move so bold
One planet and 6 hours of sun
Eternity it was to run

Countless probes to void disperse
Seed far reaches of universe
With thriving life, and beauty's play
Through endless night to endless day

Now back on Earth the plan continues
Of course, we shared with it our v... (read more)

[-]plex5y90

A couple of months ago I did some research into the impact of quantum computing on cryptocurrencies, seems maybe significant, and a decent number of LWers hold cryptocurrency. I'm not sure if this is the kind of content that's wanted, but I could write up a post on it.

7ChristianKl5y

Writeup of good research is generally welcome on LessWrong.

2Document10mo

Did you?

6plex10mo

I think I have a draft somewhere, but never finished it. tl;dr; Quantum lets you steal private keys from public keys (so all wallets that have a send transaction). Upgrading can protect wallets where people move their coins, but it's going to be messy, slow, and won't work for lost-key wallets, which are a pretty huge fraction of the total BTC reserve. Once we get quantum BTC at least is going to have a very bad time, others will have a moderately bad time depending on how early they upgrade.

[-]plex2mo*60

Let's say you want to insure against rogue superintelligence, and handwave all those pesky methodological issues like "who enforces it", "everyone being dead" and "money not being a thing when everyone is dead". How much does the insurance company risk?

First, let's try combining the statistical value of a human life ($13.2m), Bostrom's estimate of the number of lives the future could support (10^43 - 10^58) to give you a value of the future lives as between $1.32 × 10^50 (132 quindecillion) and $1.32 × 10^65 ($132 quattuordecillion). This is a fairly big n... (read more)

[-]plex9mo61

My current guess as to Anthropic's effect:

0-8 months shorter timelines^[1]
Much better chances of a good end in worlds where superalignment doesn't require strong technical philosophy^[2] (but I put very low odds on being in this world)
Somewhat better chances of a good end in worlds where superalignment does require strong technical philosophy^[3]

^{^}
Shorter due to:
- There being a number of people who might otherwise not have been willing to work for a scaling lab, or not do so as enthusiastically/effectively (~55% weight)
- Encouraging race dynamics (~30%)
- Making

... (read more)

2Mateusz Bagiński9mo

how low?

4plex9mo

eh, <5%? More that we might be able to get the AIs to do most of the heavy lifting of figuring this out, but that's a sliding scale of how much oversight the automated research systems need to not end up in wrong places.

1williawa9mo

I basically agree with this. Or I'd put 20% chance on us being in the worlds "where superalignment doesn't require strong technical philosophy", that's maybe not very low. Overall I think the existance of Anthropic is a mild net positive, and the only lab for which this is true (major in the sense of building frontier models). "the existence of" meaning, if they shut down today or 2 years ago, it would've not increased our chance of survival, maybe lowered it. I'm also somewhat more optimistic about the research they're doing helping us in the case where alignment is actually hard.

[-]plex1y20

Titles of posts I'm considering writing in comments, upvote ones you'd like to see written.

[-]plex1y240

Why SWE automation means you should probably sell most of your cryptocurrency

1FinalFormal27mo

Super curious- are you willing to give a sentence or two on the take here?

2plex7mo

A_donor wrote up some thoughts we had: https://www.lesswrong.com/posts/3eXwKcg3HqS7F9s4e/swe-automation-is-coming-consider-selling-your-crypto-1

[-]plex1y130

An opinionated tour of the AI safety funding landscape

9plex1y

Grantmaking models and bottlenecks

7plex1y

Ayahuasca: Informed consent on brain rewriting (based on anecdotes and general models, not personal experience)

3gwern1y

Can't you do this as polls in a single comment?

4plex1y

LW supports polls? I'm not seeing it in https://www.lesswrong.com/tag/guide-to-the-lesswrong-editor, unless you mean embed a manifold market which.. would work, but adds an extra step between people and voting unless they're registered to manifold already.

3plex1y

Mesa-Hires - Making AI Safety funding stretch much further

[-]plex5y*20

Thinking about some things I may write. If any of them sound interesting to you let me know and I'll probably be much more motivated to create it. If you're up for reviewing drafts and/or having a video call to test ideas that would be even better.

Memetics mini-sequence (https://www.lesswrong.com/tag/memetics has a few good things, but no introduction to what seems like a very useful set of concepts for world-modelling)
- Book Review: The Meme Machine (focused on general principles and memetic pressures toward altruism)
- Meme-Gene and Meme-Meme co-evolution (fo

... (read more)

1cSkeleton3y

Hi, did you ever go anywhere with Conversation Menu? I'm thinking of doing something like this related to AI risk to try to quickly get people to the arguments around their initial reaction and if helping with something like this is the kind of thing you had in mind with Conversation Menu I'm interested to hear any more thoughts you have around this. (Note, I'm thinking of fading in buttons more than a typical menu.) Thanks!

Moderation Log

plex's Shortform

8

An Opinionated Review of Many AI Safety Funders (with recommendations)[1]

If We Win Then...

An Opinionated Review of Many AI Safety Funders (with recommendations)^[1]