Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Good Heart Week: Extending the Experiment, published by Ben Pace on April 2, 2022 on LessWrong.

Yesterday we launched Good Heart Tokens, and said they could be exchanged for 1 USD each.

Today I'm here to tell you: this is actually happening and it will last a week. You will get a payout if you give us a PayPal/ETH address or name a charity of your choosing.

Note that voting rings and fundraising are now out of scope, we will be removing and banning users who do that kind of thing starting now. More on this at the end of the post.

Also, we're tentatively changing posts to be worth 4x the Good Heart Tokens of comments.

Why is this experiment continuing?

Let me state the obvious: if this new system were to last for many months or years, I expect these financial rewards would change the site culture for the worse. It would select on pretty different motives for being here, and importantly select on different people who are doing the voting, and then the game would be up.

(Also I would spend a lot of my life catching people explicitly trying to game the system.)

However, while granting this, I suspect that in the short run giving LessWrong members and lurkers a stronger incentive than usual to write well-received stuff has the potential to be great for the site.

For instance, I think the effect yesterday on site regulars was pretty good. I'll quote AprilSR who said:

I am not very good at directing my monkey brain, so it helped a lot that my System 1 really anticipated getting money from spending time on LessWrong today.

...There’s probably better systems than “literally give out $1/karma” but it’s surprisingly effective at motivating me in particular in ways that other things which have been tried very much aren’t.

I think lots of people wrote good stuff, much more than a normal day. Personally my favorite thing that happened due to this yesterday was when people published a bunch of their drafts that had been sitting around, some of which I thought were excellent. I hope this will be a kick for many people to actually sit down and write that post they've had in their heads for a while.

(I certainly don't think money will be a motivator for all people, but I suspect it is true for enough that it will be worth it for us given the Lightcone Infrastructure team's value of money.)

I'm really interested to find out what happens over a week, I have a hope it will be pretty good, and the Lightcone Infrastructure team has the resources that makes the price worth it to us. So I invite you into this experiment with us :)

Info and Rules

Here's the basic info and rules:

Date: Good Heart Tokens will continue to be accrued until EOD Thursday April 7th (Pacific Time). I do not expect to extend it beyond then.

Scope: We are no longer continuing with "fun" uses of the karma system. Voting rings, fundraising posts, etc, are no longer within scope. Things like John Wentworth's and Aphyer's voting ring, and G Gordon Worley III's Donation Lottery were both playful and fine uses of the system on April 1st, but from now I'd like to ask these to stop.

Moderation: We'll bring mod powers against accounts that are abusing the system. We'll also do a pass over the votes at the end of the week to check for any suspicious behavior (while aiming to minimize any deanonymization).

Eligible: LW mods and employees of the Center for Applied Rationality are not eligible for prizes.

Votes: Reminder that only votes from pre-existing accounts are turned into Good Heart Tokens. (But new accounts can still earn tokens!) And of course self-votes are not counted.

Cap Change: We're lifting the 600 token cap to 1000. (If people start getting to 1000, we will consider raising it further, but no promises.)

Weight Change: We're tentatively changing it so that votes on posts are now worth 4x votes on comments. (This will...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: New Scaling Laws for Large Language Models, published by 1a3orn on April 1, 2022 on LessWrong.

On March 29th, DeepMind published a paper, "Training Compute-Optimal Large Language Models", that shows that essentially everyone -- OpenAI, DeepMind, Microsoft, etc. -- has been training large language models with a deeply suboptimal use of compute.

Following the new scaling laws that they propose for the optimal use of compute, DeepMind trains a new, 70-billion parameter model that outperforms much larger language models, including the 175-billion parameter GPT-3 and DeepMind's own 270-billion parameter "Gopher".

I'm going to walk through the background of the now-falsified scaling laws from prior to this paper; then I'm going to describe the new laws given by this paper, and why they weren't found earlier; and finally I'll briefly mention some possible implications of this paper.

Independently of the consequences -- this paper is exciting! Machine learning researchers thought they knew laws about how to scale compute optimally, and the laws turned out to be wrong! It's a nice clear instance of science-functioning-in-ways-it-should in ML.

Background

In 2020 OpenAI proposed scaling laws which have since been used (at least implicitly) to guide the training of large models.

These scaling laws attempt to answer several questions. One of these questions is "Given a certain quantity of compute, how large of a model should I train in order to get the best possible performance?"

The answer isn't "as large a model as possible" because, for a fixed quantity of compute, a larger model must be trained on less data. So training a 1-million parameter model on 10 books takes about as many FLOPS as training a 10-million parameter model on one book.

In the case of very large language models like GPT-3, these alternatives look more like training a 20-billion parameter model on 40% of an archive of the Internet, or training a 200-billion parameter model on 4% of an archive of the Internet, or any of an infinite number of points along the same boundary.

Compute on this scale is not cheap -- so if you're going to be spending 10 million dollars per training run on a model scaled up to be 100x bigger than your toy version of the model, you want principles better than a feeling in your gut to guide how you allocate this compute between "amount of data the model sees" and "how big the model should be."

So if you get 10x more compute, how much bigger do you make your model? What about 100x more compute? Or 1000x more compute?

Well, the OpenAI paper answers the question. If you get 10x more compute, you increase your model size by about 5x and your data size by about 2x. Another 10x in compute, and model size is 25x bigger and data size is only 4x bigger.

Model size is almost everything.

Model Size Is (Almost) Everything

Subsequent researchers and institutions took this philosophy to heart, and focused mostly on figuring out how to engineer increasingly-large models, rather than training comparatively smaller models over more data. Thus, the many headlines of increasingly-larger models that we've seen coming from ML research institutions and AI accelerator startups.

See, for instance, the following chart from the new DeepMind paper.

Large Subsequent Models

Note the increase to half a trillion parameters, with identical quantities of training data.

And note that this understanding of the world has also been used to project forward future data requirements -- NVIDIA, for instance, talks about training a trillion parameter model with only 450 billion tokens. Everyone had decided model size was much more important than data size.

Falsification

The DeepMind paper re-approaches the issue of scaling laws.

It uses three separate methods to try to find the correct scaling law, but I'm going to zoom in on t...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Introducing a New EA Dating App: UpDating, published by Aris Richardson on April 1, 2022 on The Effective Altruism Forum.

While Swapcard has been a reliable software to use at EAGs, there are undeniably some features that could be improved in the app. It’s easy to miss notifications in a sea of messages, it’s difficult to propose one-on-one meetings on mobile, and most importantly, the app horribly misses the opportunity to capitalize on low-hanging fruit for improving the dating lives of EAs.

Thankfully, just in time for the start of EAGx Boston, I am proud to officially launch UpDating (promotional video linked). UpDating is an application funded by CEA and designed to be the optimal micromarriage optimizer for any EAG(x). It has all of your favorite features from Swapcard, plus our additional features:

A feature to reject one-on-ones by swiping left.

A new narrow filter for the attendee page to track relationship status.

A calculator that automatically tracks your micromarriages throughout the conference.

Attendees identifying as polyamorous now unlock the option to receive AI Matches in pairs.

Attendees will finally have the option to upload their Tinder profiles alongside their other social media links.

Future updates of the app, which are on track to roll out by EAGL, include:

A competitive micromarraige contest.

An option for past partners to leave reviews on attendee profiles and publicly display a rating out of 5 stars.

Pre-typed pick-up lines to send when requesting one-on-ones.

Find us at EAGx Boston this Sunday in Junior Ballroom hosting our workshop: How to Write a Blog Post that Gets You a Date!

Note: As we’re still in the early stages of this project, we welcome any comments with suggestions for additional features!

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Manafold Markets is out of mana 🤭, published by Austin Chen on April 1, 2022 on LessWrong.

Crossposted from Above The Fold.

The problem

As you may know, we’ve been busy raising our seed round in the last couple of weeks. One piece of feedback we often hear from investors is to be careful with how we issue our M$ (aka mana). We conducted an internal audit, which revealed that... Manafold Markets is out of mana!!

We did some brainstorming on how to address this unexpected shortfall. Some ideas:

Charge new users to join Manafold instead of handing out M$ 1,000

Institute demurrage, a negative interest rate on unspent mana

Start an OnlyFans and post exclusive photos of our mascot, Foldy McFoldface

The solution

In the end, we decided to go after one source of nearly limitless mana: tax the rich traders who sit atop our leaderboards!

We were inspired by Radical Markets to set up a Harberger Tax. This means that all our leaderboard slots are now for sale. Instead of trying to make good predictions in our markets in the hopes of turning a profit, you can just buy that leaderboard position you’ve always wanted! Check it out:

How this works

Every leaderboard slot has an "assessed value": what the current holder thinks their slot is worth.

Slot holders pay a continuous fee of 25% per hour to Manafold.

At any time, you can pay the assessed value of a slot to buy it from the current holder.

The slot is now yours! You can customize the message, or reassess it to a new value.

Try it out now, at!

Where did Manafold's mana go?

Honestly, we're as puzzled as you are. Leading theories include:

Leaky abstractions in our manabase

One too many floating-point rounding errors

Our newest user Robert');DROP TABLE Balances;--

We'd be happy to pay a bounty to anyone who can help us solve this riddle! Preferably in USD, as we have a lot more of that now than mana.

In other news

It’s been quite a busy day on the internet; who could have predicted all these announcements?

Move over, GPT-3; here come blog posts generated via prediction market

Our largest funder is now spinning up an Effective Altruism reality TV show

Don’t miss the April 2222 edition of the Forecasting newsletter

LessWrong is replacing Karma with Good Heart Tokens

(Fiction) Prediction markets tackle corruption in China

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: MIRI announces new "Death With Dignity" strategy, published by Eliezer Yudkowsky on April 2, 2022 on LessWrong.

tl;dr: It's obvious at this point that humanity isn't going to solve the alignment problem, or even try very hard, or even go out with much of a fight. Since survival is unattainable, we should shift the focus of our efforts to helping humanity die with with slightly more dignity.

Well, let's be frank here. MIRI didn't solve AGI alignment and at least knows that it didn't. Paul Christiano's incredibly complicated schemes have no chance of working in real life before DeepMind destroys the world. Chris Olah's transparency work, at current rates of progress, will at best let somebody at DeepMind give a highly speculative warning about how the current set of enormous inscrutable tensors, inside a system that was recompiled three weeks ago and has now been training by gradient descent for 20 days, might possibly be planning to start trying to deceive its operators.

Management will then ask what they're supposed to do about that.

Whoever detected the warning sign will say that there isn't anything known they can do about that. Just because you can see the system might be planning to kill you, doesn't mean that there's any known way to build a system that won't do that. Management will then decide not to shut down the project - because it's not certain that the intention was really there or that the AGI will really follow through, because other AGI projects are hard on their heels, because if all those gloomy prophecies are true then there's nothing anybody can do about it anyways. Pretty soon that troublesome error signal will vanish.

When Earth's prospects are that far underwater in the basement of the logistic success curve, it may be hard to feel motivated about continuing to fight, since doubling our chances of survival will only take them from 0% to 0%.

That's why I would suggest reframing the problem - especially on an emotional level - to helping humanity die with dignity, or rather, since even this goal is realistically unattainable at this point, die with slightly more dignity than would otherwise be counterfactually obtained.

Consider the world if Chris Olah had never existed. It's then much more likely that nobody will even try and fail to adapt Olah's methodologies to try and read complicated facts about internal intentions and future plans, out of whatever enormous inscrutable tensors are being integrated a million times per second, inside of whatever recently designed system finished training 48 hours ago, in a vast GPU farm that's already helpfully connected to the Internet.

It is more dignified for humanity - a better look on our tombstone - if we die after the management of the AGI project was heroically warned of the dangers but came up with totally reasonable reasons to go ahead anyways.

Or, failing that, if people made a heroic effort to do something that could maybe possibly have worked to generate a warning like that but couldn't actually in real life because the latest tensors were in a slightly different format and there was no time to readapt the methodology. Compared to the much less dignified-looking situation if there's no warning and nobody even tried to figure out how to generate one.

Or take MIRI. Are we sad that it looks like this Earth is going to fail? Yes. Are we sad that we tried to do anything about that? No, because it would be so much sadder, when it all ended, to face our ends wondering if maybe solving alignment would have just been as easy as buckling down and making a serious effort on it - not knowing if that would've just worked, if we'd only tried, because nobody had ever even tried at all. It wasn't subjectively overdetermined that the (real) problems would be too hard for us, before we made the only attempt at solvi...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Time-Time Tradeoffs, published by Akash on April 1, 2022 on The Effective Altruism Forum.

Epistemic Status: An idea I thought of a few days ago and wrote up quickly.

I often hear EAs making Time-Money tradeoffs. The general idea is that money can often be exchanged for time (e.g., buying an Uber to save 30 minutes), and time can be exchanged for money (e.g., spending an hour to earn $X).

For example, let’s suppose Alice values her time at $50 per hour. Alice is willing to turn her money into ways of freeing up their time (by paying other people to do work for her for less than $50 per hour, for paying for services that save her an hour for less than $50, etc.) or by knowing how much money she’d take to lose time (by only working only pay $50+, etc.).

I recently thought of another kind of tradeoff that I’ve been finding useful: Time-Time tradeoffs. The idea here is that present time can often be exchanged for future time (and vice-versa).

The idea here is that not all hours are equally valuable. A standard case might involve thinking about which hours of your day tend to be most productive. Let’s suppose that Alice “peaks” around 12PM: she tends to be 2X more productive from 12PM-1PM than any other hour of her day.

But Alice also benefits from eating lunch at some point in the afternoon. By default, Alice had been eating lunch at 12PM each day. But Alice realizes that she can push her lunch break back to 2PM. She is effectively making a trade: 12PM Alice gains an hour in exchange for an hour from 2PM Alice.

Time-time trades also relate to time-money trades. For example, if Alice values a “normal hour” of her time at $100/hour, she should be willing to pay $200 to save the hour from 12-1PM.

Additionally, time-time trades don’t have to be 1:1. I noticed this when I was at EAGxOxford. A marginal hour of my time spent on community-building (networking, coming up with ideas, etc) at EAGxOxford was at least 3X as valuable as a marginal hour of “regular” community-building. In order to maximize especially impactful time, it would be worthwhile to sacrifice a greater number of hours of less impactful time. For example, if Alice's 12-1PM is twice as effective as her usual hour, she should be willing to trade 90 minutes of “non-peak Alice time” in order to gain 60 minutes of “peak Alice time.”

After thinking about Time-Time trades, I intend to spend more time completing tasks before or after conferences, even if it takes more time overall. This allows “EAG Akash” to borrow time from “Past Akash.”

I think there are also examples in which “EAG Akash” would want to borrow time from "Future Akash." For example, suppose it’s the last day of EAG:

Last-day-of-EAG Akash: Sorry, but I think I should stay up until 2AM and wake up at 7AM in order to maximize the number of great conversations that I have.

Next day Akash: But that’s going to make me tired, so it’ll take me twice as long for me to write a blog post and send follow-up emails!

Last-day-of-EAG Akash: True. But although it’ll make you ½ as productive for 10 hours, I’d say my time is 4X as valuable as yours, and I’d save 4 hours. So you lose 5 productivity-adjusted hours, but I gain 16! As an added bonus, the conversations might help me come up with blog post ideas that are 2X as good as the ones I would write about without these conversations.

Next day Akash: Ah, okay. I’ll take one for the team!

The examples are meant to be purely illustrative. For some people, it will not make sense to get 5 hours of sleep or to move their lunch to 2PM. I encourage you to consider how time-time tradeoffs might be useful in your life.

A few guiding questions:

What hours of the day are you most productive? (e.g., 12-1PM)

Where do you notice progress happening the most? (e.g., at an EAGx; at a co-working space)

Are there ways you can trade less e...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Announcing Impact Island: A New EA Reality TV Show, published by Linch on April 1, 2022 on LessWrong.

Impact Island is the hip new reality TV show that all of your cool, chic, and longtermist impact-oriented friends would love to binge watch. Think "Survivor" meets "Shark Tank" meets "Open Philanthropy."

Premise

The premise is simple:36 young humble aspiring effective altruists are invited to a private Caribbean Island. Each aspiring EA is expected to do high-impact projects during their stay here. They will also gossip and continuously publicly rate each other's projects, impact, and likability. Every week, an anonymous team of grantmakers rank all participants, and whoever accomplished the least morally impactful work that week will be kicked off the island.

At the end of the season (36 episodes), the single remaining most impactful participant will be given 10 million dollars of seed funding for their new org.

How will you select participants?

Participants will be chosen via a formal selection process combining past track records, good looks, EA promisingness, movie chemistry, reference letters, and the coherent extrapolated gut feels of leading EA grantmakers. Of course, all participants are expected to be polyamorous.

How much is this project? Do you need more funding?

Our current budget is $25 million dollars. We have already secured funding from reputable EA sources, such as the FTX Future Fund, Open Philanthropy, Longview Philanthropy's Media Engagement arm, actor Joseph Gordon-Levitt, and real estate mogul with political and reality TV experience Donald Trump.

However, 25 million is only enough money to tide us over for season 1 (36 episodes). As we (the producers) believe strongly in patient longtermism, we would ideally like to run for 3 quadrillion seasons. Thus, we have extremely large room for more funding, which we do not expect to fill any time soon, making us the premier community building EA megaproject.

Do you plan to seek additional sources of funding?

Yes, we believe in being self-financing as quickly as possible, so we plan to mint an NFT of every episode.

What is the expected impact of this project?

We think the expected impact of the project is so high it is literally incalculable. If we're forced to put numbers on this, we think we can get at least 10 million viewers, and conservatively 1% of those people (1 billion viewers) will counterfactually become EAs. This will be the largest and most successful community building megaproject EA has ever seen, and even if our impact calculation is substantially off (say by 25%), 750 million highly engaged EAs is still an impressive number.

In addition to the community building aspect, we also think that the direct impact of having 36 EAs do highly impactful projects can be quite high, and having them all continuously rate each other can help solve the vetting bottleneck in EA.

Further, we think the NFTs funding schema should quickly become self-financing, meaning that this project will pay for itself in no time (less than 3 geological eons in our techno-economic analysis).

How does this project fair in terms of importance, neglectedness, and tractability?

Importance

In addition to becoming a future cultural milestone (incalculable value), we also think this project has very high community building benefits, as well as direct impact.

Neglectedness

To the best of our knowledge, no other impact-oriented reality TV show in the Carribean exists, suggesting high neglectedness.

Tractability

We're very sure of the tractability of this project, having read skimmed multiple blog posts on how to run reality TV shows.

How will you make sure the TV show is fun to watch?

We think watching people doing high-impact altruistic activities in front of their computers is high utility (and thus fun) enough!

But for discern...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: New Product MVP: LightWrong, published by E. Garrett on April 1, 2022 on LessWrong.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Kosher Hot Dogs, published by johnswentworth on April 1, 2022 on LessWrong.

If ever there were an industry that no one in their right mind would trust, it's hot dogs. First rule of hot dogs: you do not want to know what's in it. But in this unsavory industry, Hebrew National stands out. The slogan on the company's site summarizes their advantage well:

When your hot dog's kosher, that's a hot dog you can trust.

And indeed, people do trust Hebrew National's hot dogs.This is more remarkable than it might seem at first glance. Consider the problem from the perspective of a hot dog company. This company wants to carve out a niche: they will produce high-quality hot dogs, and sell them at a correspondingly higher price. Many people will happily pay extra to know that their hot dogs do not contain ground-up lucky charms, bits of fur, or the occasional lost dog. But how can the company communicate their quality to prospective consumers? How can they convince the public of the superior quality of their hot dogs? What claim could they make which an unscrupulous competitor could not copy?This is the problem known in game theory as signalling: one party wants to communicate their superior quality to another party, but in order to do so they must send a signal which their unscrupulous competitors cannot easily copy. A certification body can solve this problem.

Consumer Reports, for example, provides unbiased analysis of a wide range of products. Unfortunately, this solution is subject to attack in the real world by exploiting the limited information capacity of consumers. Any company can (and does) invent arbitrary metrics by which their product performs best. A less cynical interpretation is that each company stakes out a niche, claiming that their product is the best for X. If you want X, you buy that company's product. Just within hot dogs, we have Ballpark's "Angus", Applegate's "Organic", Nathan's "Bigger than the Bun", Oscar Meyer's "Selects", and several brands of "Premium Jumbo". Many of these brands have multiple lines of hot dog servicing different niches. Consumers with limited attention to devote will ignore most of these, and most are useless anyway.

Thus the true signals of quality are drowned out by the noise of their competitors.If a company is to charge extra for a truly superior product, then they need a more dramatic way to signal quality. Hebrew National does this by invoking kosher rules. Implicitly, the entire weight of the Jewish religion backs the quality of their product. The kosher rules force the company to produce high-quality hot dogs, and lets them communicate their high quality even in the noisy environment of the modern supermarket.But what does that even mean? I don't actually know the kosher rules. I remember a few bits and pieces... no hooved animals, separate meat and dairy, something about which cuts of meat are acceptable... but I don't know most of it. Yet I'm willing to accept kosher standards as an assertion of quality, at least in the unpalatable hot dog industry.I do know that kosher rules are generally intended to ensure food quality. They are a 3000 year old FDA regulatory equivalent.

They are interpreted by an active rabbinical community, which makes sure that the word and spirit of the rules are properly applied to new foods and new food processing technology. The result is a regulatory framework which is roughly understood and highly trusted by laypeople, even though most of us do not have a detailed knowledge of the rules! Just as important, I know that the whole community of people who observe, certify and maintain kosher rules is highly trustworthy. They consider themselves in service to God. Abusing the kosher rules or certification would be not just unethical, but a direct transgression against God. I can trust that the rules a...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Good Heart Donation Lottery, published by G Gordon Worley III on April 1, 2022 on LessWrong.

Greetings, all!

Good Heart Tokens are clearly the mechanism we need to do more Good in the world. To that end, I'm starting a daily donation lottery.

Here's how it works.

First, strong up vote this post. We need to try to get it to at least $600 worth of upvotes to maximize the payout. (though see update 1; there's no limit)

Next, give an answer to this "question" specifying the charity you'd like to see donated to.

Then, upvote answers. This will generate a weighted distribution from which I will randomly pick the winner of the lottery. So, for example, if there's 1 answer for MIRI that gets 69 Good Heart Tokens and 1 answer for AMF that gets 420 Good Heart Tokens, then I'll do a random weighted drawing where MIRI has a weight of 69 and AMF has a weight of 420. The winner gets the pot of tokens generated by upvoting this post.

You get to do what you like with the tokens on your answer. I think in good conscience you should donate them to the winner (announced in an update to the post the next day), however I have no way to enforce that.

No duplicate answers for the same charity. If you accidentally post a duplicate please delete it. If you don't, whenever I notice it I'll keep only the one that has the earliest time stamp.

Good luck!

Update 1: After posting this I quickly realized there's no $600 cap. From the payments page: "If you receive more than $600 in a year, you'll need to be entered into Center for Applied Rationality's payment system. CFAR will contact you via your LessWrong email address about next steps. (Make sure it's an email that you check regularly)". So up vote as much as you can, the sky is the limit!

To that end I'll be posting some seed answers to vote on since I'm not the bottleneck. All tokens earned on my answers will be added to the donation pool.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.