All of Liam Donovan's Comments + Replies

Let's See You Write That Corrigibility Tag

Would be very curious to hear thoughts from the people that voted "disagree" on this post

It's a shame we can't see the disagree number and the agree number, instead of their sum.

[$20K in Prizes] AI Safety Arguments Competition

Maybe you could measure how effectively people pass e.g. a multiple choice version of an Intellectual Turing Test (on how well they can emulate the viewpoint of people concerned by AI safety) after hearing the proposed explanations. 

[Edit: To be explicit, this would help further John's goals (as I understand them) because it ideally tests whether the AI safety viewpoint is being communicated in such a way that people can understand and operate the underlying mental models. This is better than testing how persuasive the arguments are because it's a) mo... (read more)

Don't die with dignity; instead play to your outs

I strongly prefer the "dying with dignity" mentality for 3 basic reasons:

  • as other people have mentioned, "playing to your outs" is too easy to misinterpret as conditioning on comfortable improbabilities no matter how much you try to draw the distinctions
  • relatedly, focusing on "playing to your outs" (especially if you do so for emotional reasons) may make it harder to stay grounded in accurate models of reality (that may mostly output "we will die soon")
  • Operating under the mindset that death is likely when AGI is still some ways around the corner and easy t
... (read more)
2022 ACX Predictions: Buy/Sell/Hold

Wait why are your predictions for Brazil so far from the market? As of right now, there are 180,000 shares of Bolsonaro on the orderbook under 50c on FTX (avg price of 44c if you buy them all).

I explicitly did not look and hadn't previously checked. Given that, this distance is small. With that information I would revise to be moderately above current fair.
Polymarket Covid-19 1/17/2022

Yeah it's definitely against poly's terms of service but not against US law (otherwise they wouldn't be complying with the prohibition on offering their services to US customers)

Polymarket Covid-19 1/17/2022

FWIW it is totally legal for Americans to trade on polymarket via a VPN or similar; it's just not legal for polymarket itself to offer services to people with US IP addresses

I'm not a lawyer, but their newest Terms of Service [] imply otherwise: Not sure how willing and able they would be to enforce such a regulation, but that's a different question. (Not legal advice!)
Yes, a huge one [] . "COVAX, the global program for purchasing and distributing COVID-19 vaccines, has struggled to secure enough vaccine doses since its inception.. Nearly 100 low-income nations are relying on the program for vaccines. COVAX was initially aiming to deliver 2 billion doses by the end of 2021, enough to vaccinate only the most high-risk groups in developing countries. However, its delivery forecast was wound back in September to only 1.425 billion doses by the end of the year. And by the end of November, less than 576 million doses had actually been delivered."
Occupational Infohazards

Yep, I wanted to experiment with a central example of a comment that should be in the "downvote/agree" quadrant, since that seemed like the least likely to occur naturally. It's nice to see the voting system is working as intended. 

This is a good example of a comment that is not a contribution to the discussion, but which lots of people here agree with, illustrating the power and flexibility of the two-axis voting system. Except, if you meant it as such, it's a valuable contribution to the discussion, and shouldn't be downvoted. But if it weren't downvoted it wouldn't properly show off the voting system, and would no longer be valuable. You've tied us in knots!

Zvi’s Thoughts on the Survival and Flourishing Fund (SFF)

I haven't done much research on this, but from a naive perspective, spending 4 billion dollars to move up vaccine access by a few months sounds incredibly unlikely to be a good idea? Is the idea that it is more effective than standard global health interventions in terms of QALYs or a similar metric, or that there's some other benefit that is incommensurable with other global health interventions? (This feels like asking the wrong question but maybe it will at least help me understand your perspective)

The idea is that the extra production capacity funded with that $4b doesn't just move up access a few months for rich countries, it also means poor countries get enough doses in months not years, and that there is capacity for making boosters, etc. (It's a one-time purchase to increase the speed of vaccines for the medium term future. In other words, it changes the derivative, not the level or the delivery date.)
Omicron Post #4

Wait, how do you get to 17%-25% chance of a crisis situation if there's only a 2.5% chance of omicron causing severe disease in vaccinated/previously infected people? Isn't that the vast majority of people in the US?

Because there aren't enough hospital beds/nurses etc
My understanding is that a few months after vaccination the vaccines provide ~85% (?) protection against hospitalisation compared to an unvaccinated person. I think the 2.5% prediction is intended to indicate that Zvi thinks it's highly unlikely that this will drop to 20% protection from Omicron (or similarly low number). However even a drop to 70% (which seems more reasonable) would mean a doubling of hospitalisations per case. Even with this, my model of a crisis situation is more to do with a huge increase in infections than with increasing hospitalisation rates per case.
We know the vaccination rates (NPR [] ). And we have estimates for the the number of infected persons (Estimated COVID-19 Burden []). But do we have a trustworthy estimate for the intersection of those two groups (and with that an estimate for the total number of previously infected and/or vaccinated)? That would be important information, ideally statewise. The other important data point would be the intensive care capacity.
Omicron Variant Post #1: We’re F***ed, It’s Never Over

My uninformed impression is that an "adjuvant" is just something that stimulates increased immune response, which has historically been an additional chemical (such as an aluminum salt) added to vaccines. The mRNA vaccines do not contain these chemicals, but some people confusingly refer to the lipid nanoparticles that surround and protect the mRNA as adjuvants (because they also help increase immune response). I haven't seen any evidence that these lipid nanoparticles are the kind of adjuvants that might mitigate OAS, but that doesn't mean much because I'... (read more)

Immune response to the lipid nanoparticles themselves is very bad. It reduces the uptake of protein coding mRNA into the cells and diminishes vaccine efficacy. It also leads to most of the severe allergic reactions.
Adjuvants are for activating the immune system to respond to free-floating virus particles. If the virus actually infects a cell and produces proteins on the cell's surface, that is a different signal to the immune system and adjuvants are not needed. Thus adjuvants were not needed for traditional live/attenuated vaccines and 21st century mRNA and vector vaccines (AZ, etc), but were needed for 1980s recombinant protein vaccines (Novavax) and traditional dead/inactivated vaccines.
Yes. [] works fine.
What are fiction stories related to AI alignment?

Detonation -- an entertaining book that tries to flesh out a fast takeoff scenario and explicitly cites Bostrom and Yudkowsky. However, it also makes some extremely dubious choices; for example, the protagonist is a Marine hired to fight against the unfriendly AI, which doesn't seem like a very effective AI alignment strategy.

The link doesn't work.
My Journey to the Dark Side

Hopefully it's not too late to try to keep the focus on the fun puzzles, etc! There really does seem to be an alarming amount of craziness floating around LW, along with the constant weird attempts to explicitly model things that we evolved to understand instinctively (eg most aspects of social interaction). Reading that stuff slightly negatively affected my mental health despite thinking it was mostly silly -- to the extent it's taken seriously it seems like it could have more substantial negative effects. 

Open and Welcome Thread - April 2021

If anyone wants to get money onto Polymarket (real-money prediction market with no limits, no fees, and a wide variety of markets), I can facilitate that for free. Send me cash or crypto and I will send money directly to your poly account. As a heavy polymarket user I'm somewhat sad more LW people aren't taking the opportunity to bet on their beliefs, so I'd like to make it as easy as possible for anyone interested. 

Violating the EMH - Prediction Markets

What prediction markets are liquid enough for a fund to make sense? (unless you were just referring to the crypto trading part?)

Hello, You're right. Most of the time there are not many prediction markets I would want to take positions in that also have enough capacity. We would be willing to regularly bet millions of dollars on coups not occuring in the US, but these opportunities are unusual.
Scott Alexander 2021 Predictions: Buy/Sell/Hold

Risk neutral pricing is always a danger when trying to make inferences about real world probabilities based on market pricing, but it's usually a negligible one because participants in current prediction markets are generally speculators with no built-in exposure to the underlying asset, or ability to hedge against other markets.

 On the other hand, implied probabilities from options pricing can differ significantly from real world probability, because any participant in the options market can hedge their position against the underlying asset.

"In all t... (read more)

Right, but then the underlying asset is telling you something and if you disagree with that, then you can trade the underlying asset. There's nothing special about options here. The difference comes from the fact that the underlying asset can have a return. (In the same way that a bond have a price different from par doesn't (necessarily) mean that the market is forecasting default - they are discounting the value of a future cash flow). The evidence would be something akin to "the historic sharpe for risk assets is <1" so the order magnitude of risk premia is "small enough" relative to the volatility. I don't think there is anything misleading about taking the market prices, constructing a bet and presenting that as a market probability, any more than taking showing betting odds and saying that's the betting market probability. Sure, there might be some subtleties depending on the market (eg long-shot bias, fees, etc), but fundamentally that's the price the market is offering. If you disagree, BET. I agree with this, all I'm saying is that the degree to which those assets fail to be a martingale is small relative to their volatility. I assume all those people are long crypto, which fundamentally means they disagree with the underling price and are long... I don't see any inconsistency between that and what I'm saying. I would be more interested if you could find me someone who thinks both that * option prices are wrong * they shouldn't have a position in options * they shouldn't have a position in the underlying because of some kind of risk-neutral vs real-world probability considerations.
Scott Alexander 2021 Predictions: Buy/Sell/Hold

I'm happy to; no commission needed. If anyone else wants to get money from fiat into polymarket easily with no fee, just let me know

Update: we did this, I bought shares, we'll see how it goes.
Want to sell me USDC on there in exchange for paypal, so I can bet? (I'll gladly pay a 2% "commission" for, say, $200 in USDC.)
Scott Alexander 2021 Predictions: Buy/Sell/Hold

This isn't really a meaningful explanation for why risk neutral vs real world is meaningless? To me "the credence I have that something happens" is actually a meaningful, important number that is by definition different from the risk neutral price. You can argue that all market probabilities may deviate from real-world probabilities in some way, but that doesn't make real-world probability meaningless! 

I'm now confused as to what you mean by "real world" in this context? Zvi is giving a credence for the event (p_zvi). The market is offering a bet which implies some probability for the event (p_market). All I am noting is p_zvi is different from p_market. I don't think there's anything special about the fact that options are involved here. (Unless I'm the one inferring you were specifically talking about options when you talk about "risk neutral measures". All market probabilities are in some sense in risk neutral probabilities. If you're complaint is about me talking about market probabilities then I guess this post wasn't really for you?) EDIT: To be more concrete about this, the places where "risk neutral" vs "real world" probabilities end up mattering is places where there is a concrete risk premium. (ie what the options market implies about stocks in 1y's time doesn't account for the fact that "stocks tend to go up over time"). In all the examples we're talking about, those risk premiums are tiny relative to the numbers involved so they don't make a significant difference to how we should be calculating the "market implied" odds.
Scott Alexander 2021 Predictions: Market Prices has a market for "California Governor on 12/31/2021", which should function as a great proxy for the chances of Newsom being recalled. They're quoting .0625@? for the chance of Newsom not being governor on 12/31 (they accept bets on Newsom remaining governor, but not on Newsom being removed as governor, thus the "?")

Scott Alexander 2021 Predictions: Buy/Sell/Hold

You think there's only an 80% chance the olympics happen? This is a bit of a tangent but I'd love to hear why, since I have 100k+ shares of OLY2021 on FTX and haven't heard a convincing argument for the odds being below 90%.

re options, presumably Zvi is using the real world measure, which you can'treally infer from options prices without making a lot of dubious assumptions. Can you elaborate on how/why Zvi is "off the market forecast"?

I think there's ~80-85% chance the Olympics happen on time. I think there's a ~90% chance that the Olympics go ahead this year. I think the case against them going ahead this year is roughly: * Current state of COVID in Japan, potential for it getting worse * "Cancellation is possible" statements from government * Public opinion is against the games I don't think it's very likely, but do I think there's a ~15-20% chance that COVID flares up in Japan in the next 3 months in a particularly bad way? Doesn't seem crazy to me. (FWIW, I don't think betting on the Olympics at the FTX odds is a bad bet, I just don't think it's a sure thing) Risk neutral vs Real world measure isn't really a meaningful distinction the way you think it is. You can construct a binary bet in terms of options, and the price is the market price for that bet and that's the market probability. It's no different than betting on any other event. If you don't like market pricing, then sure, ignore everything I've written here, but don't think "risk neutral measure" is some magic phrase which lets you ignore the options market. If you think the odds are different, you can always place that bet. I'm not saying Zvi is wildly wrong. Indeed he says he wouldn't trade with anything in 40-60% (and the market being at 60% means he's technically not "off" it), but I given it's close to what he'd consider trading, I think that's an interesting difference worth noting.
Scott Alexander 2021 Predictions: Buy/Sell/Hold

Cumulative cases are still much lower, right? I assumed that "case count" meant cumulative cases, but it sounds like you're interpreting it as daily cases, which would explain our difference. Given that Scott only gives 50% chance it seems much more likely that he meant cumulative cases too imho

Scott Alexander 2021 Predictions: Buy/Sell/Hold

Can you elaborate on why India's cases being higher than the US is an instabuy to 80%? I haven't given it any thought but it seems like that would require a pretty big increase in cases there?

See latest Covid post, it's already much higher.
Tales from Prediction Markets

Because mr co2 guy was clearly making a -EV bet that happened to pay off this time :)

Speculations Concerning the First Free-ish Prediction Market

FWIW FTX allows you to bet on its prediction markets on margin with a tokenized version of the S&P 500 as collateral, which accomplishes exactly what you want to accomplish here

The EMH is False - Specific Strong Evidence

markets that easily allow you to make tens of thousands of dollars with nearly no risk. 

Back during November-January, FTX had a contract called TRUMPFEB that gave  ~risk free ~15% returns in 2 months, up to millions of dollars (by betting against Trump to be president in Feburary). Right now the FTX OLY2021 market comes pretty darn close -- you can bet hundreds of thousands of dollars on the Olympics happening at 76c. There is obviously the risk of the Olympics not happening, but I haven't seen a good case for that risk being under 10%, making this a fantastic trade in expectation. 

I agree and recommend this trade here:
Violating the EMH - Prediction Markets

Just commenting to say this is a great post and I'm surprised it hasn't gotten more engagement (maybe it's so good there's nothing else to say)

Exploiting Crypto Prediction Markets for Fun and Profit

The fees are always 2% of the transaction value; i.e. numShares*avgPrice. The trick I described lets you substitute (1-avgPrice) when that would be cheaper . I should've been much clearer about this initially; I forgot that most people here probably don't know the polymarket fee structure. 

Exploiting Crypto Prediction Markets for Fun and Profit

You still pay fees, they're just lower. Most sharps on polymarket do indeed do this

How high are the fees in that case?
Exploiting Crypto Prediction Markets for Fun and Profit

I don't know about Omen, but on Polymarket you can mostly avoid this issue by minting a complete set of shares and selling the cheaper side. In your example this would only cost you $0.0004 in fees to acquire shares at $0.98. The easiest way I know of to mint complete shares is to add and then immediately remove liquidity

What are the limits of this technique to avoid fees of polymarket? Why doesn't everyone do it and pay no fees?
Bet On Biden

But he's been a professional politics gambler for years now, which seems like much stronger evidence for evaluating his calibration than the results of one election cycle?

That does not seem like very strong evidence his bets on the 2020 election were good. He did better than people who refused to bet at all. But he did worse than people who just bet on Biden. Many people figured out the Biden bet in the poker and gambling communities. You should certainly downgrade your opinion of RJ relative to people who did a lot better.
The Pyramid And The Garden

But that line has an infinitesimal chance of intersecting with any point as significant as the Great Pyramids? Scott's argument about Uganda and Tanzania seems wrong for the same reasons 

Agreed, there is an infinitesimal chance of intersecting something with such large significance as the pyramids. My point was to show that by leaving 1 of only 2 chosen variables open AND having a variable that can be set to many different values along a continuum, you significantly increase optionality for how you could retrospectively jump through hoops to assign importance to what is just a coincidence. One further thought would be this example chose a single hypothesis a conspiracy theorist may try to prove, when in reality they will be looking for these 'proofs' in many different places. If they don't find the connection in one place they'll just keep looking.
Ongoing free money at PredictIt

I'm probably too late here but if anyone's wondering, PI will send you a 1099-MISC with $8.57 of income, and that's what you have to pay taxes on. (well, they only send the form if you make over $600, but that doesn't change your tax liability).

 The $85 you deposit definitely does not count as income, regardless of whether you itemize deductions. Regarding the discussion below, the IRS does not treat PI winnings as gambling income, but rather as 1099-MISC (other) income.

(source: have paid taxes on PredictIt winnings)

DanielFilan's Shortform Feed

Shor is very open about the fact that his views are to the left of 90%+ of the electorate, and that his goal is to maximize the power of people that share his views despite their general unpopularity. 

Yeah, I think I'm more surprised by Galef's tone than Shor's.
Real-Life Examples of Prediction Systems Interfering with the Real World (Predict-O-Matic Problems)

For anyone interested, the keyword to read about things like this in the economics literature is "performativity"

Thanks to both; this is a great example; I might add it to the main text
Thanks - this is super-helpful! And after looking briefly, a citation for the above example is here [] .
Bet On Biden

FWIW I'm 99% sure RJ made money from this election, and I'm 50% sure he made over 90k. Why would you update against someone who has consistently made enormous amounts of money betting on his beliefs?


Edit: looks like I was a bit overoptimistic about his profits but he supposedly did make a decent amount

(keep in mind he lies all the time so this is only noisy evidence)

Many people including myself got way better percentage and absolute returns on the election. He was way too optimistic.
Kelly Bet or Update?

reasonable expectation of negative outcomes if you lost your bankroll

One of the necessary conditions for the Kelly Criterion to be optimal is that losing your bankroll is infinitely bad, so a reasonable expectation of negative outcomes isn't enough to deviate from Kelly.

Kelly Bet or Update?

Isn't the point of "bet or update" is that you should be either updating on your counterparty's credences or taking a bet that your counterparty thinks is +EV? Here, the player is updating upon observing the host point to the door, not on the bet itself. 

 After the player has updated on the host pointing to the door, you can require the player to take the offered bet or update as normal. Assuming the host is offering the bet as a function of his credences*, the player should update from P(gold) = 1/3 to P(gold) ~= 0, because the player knows that... (read more)

Anti-EMH Evidence (and a plea for help)

Can you explain how the leverage system works on FTX for TRUMPFEB? The calculator on the site seems to produce bizzare results. 

Pascal's Mugging: Tiny Probabilities of Vast Utilities

You've actually met the first mugger so he's more likely to exist than a hypothetical counter-mugger you've never seen any evidence for

Kelly Bet or Update?

yeah maybe the best framing is fractional kelly is doing 2 things -- updating your beliefs based on fact that the market is offering you these odds, then betting full kelly based on your new posterior belief. 

Liam Donovan's Shortform

updated as of 11/2  around when 538 forecast frozen

Kelly Bet or Update?

I'm not sure of the math OTTOMH but isn't fractional Kelly equivalent to having some credence that your model is correct and some credence that the market is correct? 

In your example, it seems like you're assuming that the event has a 0% chance of occuring in the worlds where your model is wrong, but that doesn't make a lot of sense -- the worst case from a betting perspective is that the true odds are equal to the odds that the market assigns, because that means that any bet you make will be -EV.

Here's an interesting case study for fractional Kelly in real world betting scenarios 

I'm not sure of the math OTTOMH but isn't fractional Kelly equivalent to having some credence that your model is correct and some credence that the market is correct? 

Ah, seems promising!

So, with my credence in my own model being x, and my credence in the market being y,  becomes: 

So, yeah, makes sense.

Bet On Biden

I mean yes, clearly if you believe the Economist and 538 models are reasonable, then the TX bet isn't +EV. My point is that the models are clearly unreasonable for the reasons listed above, and the bet is actually extremely +EV. 

There are three arguments (1) polls underestimating Dems in Southern states, and (2) benchmarking against 2018 senate, and (3) some low-quality Tweets. It's weird to hold a lot of stock in (2), given noise from candidate selection and other variables. If you place a lot of weight on (1), the actually sane bet would be Biden in AZ. It's rated 2nd and 4th most likely to go dem by Cohn and Wasserman [] respectively. Biden for AZ: 77% likely (Economist), priced at 54% on Election Betting Odds.
Liam Donovan's Shortform

Since I've been talking about the flaws with the 538 model and betting markets I figured I should post my own predictions. I'm pretty confident these will beat both 538 and the betting markets by a substantial margin (using a metric like log loss)

1Liam Donovan2y
updated as of 11/2 around when 538 forecast frozen
PredictIt: Presidential Market is Increasingly Wrong

FWIW,  understanding is that the 538 model is meant to account for the chance of catastrophe but not the chance of electoral interference; i.e. it's meant to model the result of a hypothetical fair election held on Nov. 3. 

Bet On Biden

Neither model takes into account any of these factors, and indeed their discrepancy with the markets are much smaller. 

65% just comes from my estimate of the fair value of a TX biden contract, which comes from...whatever internal process generates probabilistic credences. 

Bet On Biden

There's plausible arguments for Trump being favored (he is currently ahead in the polls, after all), but I can't think of any good arguments for him being at  ~75c. 

EDIT: Trump and Biden are now tied in the 538 TX polling average

Load More