basil.halperin
(Nothing useful to contribute but wanted to say this seems very nice!)
The fact that Alameda could and did “borrow” (much) more than any other account on FTX due to the allow_negative flag is consistent with Levine’s description, but I agree a fuller accounting of events would include this piece of information and the accusations you cite.
Matt Levine had an especially clear exposition recently I thought:
A simple version of the charges against Sam Bankman-Fried would be something like “people deposited money at his crypto exchange, FTX, and he stole it and gave it to his crypto trading firm, Alameda Research, which squandered it on dumb crypto trades and endorsement deals.”
But this story is not exactly right. There was not money sitting in customer accounts that was then transferred to Alameda accounts and squandered. FTX was a futures exchange; it did not keep money in a box for customers. The money in your FTX account was just money that FTX owed you. Nor did Alameda need to steal the money; FTX was a leveraged futures exchange, and traders like Alameda could, in the ordinary course of business, borrow money from FTX based on their crypto positions. The problem, ultimately, at FTX, was that it owed customers a lot of money, but it couldn’t pay them, because Alameda owed FTX a lot of money, but it couldn’t pay it, because it had squandered the money on dumb crypto trades and endorsement deals.
This distinction seems nitpicky, but it is important. The story in the first paragraph is obviously illegal, but the story in the second paragraph might not be. A story like “we owed a lot of customers money, but our biggest customer owed us money, and the market moved against that customer and it defaulted on its obligations to us, so we couldn’t pay our other customers,” can be legitimate, an embarrassing accident but not fraud. (As I keep saying, it kind of happened in the legitimate regulated market for nickel futures last year.)
Instead, to prove fraud, prosecutors need to prove some lies. The evidence of fraud is not just “people put money at FTX and the money ended up with Alameda, which squandered it”: That could happen legally. The evidence of fraud has to be something more like “people put money at FTX and it ended up with Alameda in ways that FTX said it wouldn’t.” “Alameda owed us money, and defaulted” is not in itself evidence of fraud. But if FTX was going around saying things like “we have made it impossible for Alameda to owe us money and default,” then that turned out to be a lie. And that is fraud.
I have said this before, and I think some people have found it annoying. People want this case to be simple: The money was stolen, all of this stuff about leveraged futures trading is a hand-waving distraction. But I think it is clear, from the trial so far, that the prosecutors know what they have to prove. They are doing what I would: They are trying to prove that Bankman-Fried was going around lying to people about how much customers (including Alameda) could owe to FTX, and how carefully FTX managed the risk of any customer defaulting on its obligations.
Edit March 2024: bookmarking a useful Ben West comment on the same question
Thanks for this, this is a topic I am very interested in—to the the killer feature missing in Speechify is the ability to highlight and sync those highlights. Or more broadly, annotating in a multimodal way is difficult.
I instead use Goodreader, where you can have e.g. a Dropbox folder of all your PDFs synced across desktop and mobile; and you can annotate those PDFs while listening, then sync to Dropbox.
The downside of Goodreader is that the voice is pretty bad, and also that you can’t reflow the text to make it easier to read on mobile while in audio mode.
PS: The Readwise Reader app seems to be working towards an excellent all-in-one reader with TTS capability, but the existing version I’ve found to be a little too slow / some other issues. But it’s still in development and seems very promising.
Any links on the referenced strategic review? Thanks!
If you haven’t seen, a long analysis of this:
AGI and the EMH: markets are not expecting aligned or unaligned AI in the next 30 years
TLDR: bet on real rates rising
Against using stock prices to forecast AI timelines
Some cautionary points ¯\_(ツ)_/¯
Forgive me if I’m just being dumb, but—does anyone know if there is a way in settings to revert to the old font/CSS? I’m seeing a change that (for me) makes things harder to read/navigate.
(Nice, thanks for sharing)
Levered ETFs exhibit path dependency, or “volatility drag”, because they reset their leverage daily, which means you can’t calculate the return without knowing what the interest rate does in between the 3% rise
The entire section is based on a first-order approximation, as explicitly noted in the post (which is also why we set aside e.g. the important issue of convexity). This point is of course correct!
A related point: The US stock market has averaged 10% annual returns over a century. If your style of reasoning worked, we should instead buy a 3x levered S&P 500 ETF, get 30% return per year, compounding to 1278% return over a decade, handily beating out 162%.
This calculation, like that of many other commenters, estimates the total return. What matters is risk-adjusted return (a la Sharpe ratio). If you think the market is literally wrong with certainty, then the bet could be literally risk-free (“infinite Sharpe”, speaking loosely). If you aren’t 100% certain, then you have a finite risk-adjusted return, but still high—how high depends on your confidence level (etc).
Equities, on the other hand, have risk!
We welcome other criticisms to discuss, but comments like your first line are not helpful!
Thanks for these comments!
The short answer here is: yes agreed, the level of real interest rates certainly seems consistent with “market has some probability on TAI and some [possibly smaller] probability on a second dark age”.
Whether that’s a possibility worth putting weight on—speaking for myself, I’m happy to leave that up to readers.
(ie: seems unlikely to me! What would the story there be? Extremely rapid diminishing returns to innovation from the current margin, or faster-than-expected fertility declines?)
As you say, perhaps the possibility of the stagnation/degrowth scenario would have other implications for other asset prices, which could be informative for assessing likelihood.
Yes, to emphasize, the post is meant to define the situation under consideration as: “something close to a 10x increase in growth; or death”. We’re interested in this scenario only because it’s the modal scenario in the particular world of LW/EA/AI safety.
The logic of the argument does not apply as forcefully to “smaller” changes (which could potentially still be quite large), and would not apply at all if AI did not increase growth (ie did not decrease marginal utility of consumption)!
To summarise, the effect on equities seems ambiguous to you, but it’s clearly negative on bonds, so investors would likely tilt towards equities.
“Negative for bonds” does not imply “shift investment from bonds to stocks”, though. It could mean “shift toward short bonds” or “shift investment from bonds, to just invest less overall”.
In addition, the sharpe ratio of the optimal portfolio is decreased (since one of the main asset classes is worse)
I would push back on this too, for a related reason—the optimal portfolio can include “go short bonds”, which might now have a higher expected return.
I think the standard asset pricing logic would be: there is one optimal portfolio, and you want to lever that up or down depending on your risk tolerance and how risky that portfolio is. So, whether you ‘take less total exposure to risky assets’ depends on whether the argument here updates your view on how ‘risky’ the future is (Tyler Cowen has argued this, I’m not sure it’s super clear cut though).
Here’s another way of putting things, that I’ll post here for reference:
Suppose I think Google is undervalued, because it is going to have a $1T dividend in 2030, and the market doesn’t realize this.
1. I buy Google today at some cheap price.
2. Possibility 1: before 2030, the market “corrects” and realizes that it was undervaluing Google. The stock price rises, and I receive capital gains.
3. Possibility 2: the market does not “correct” before 2030. I still get the big dividend in 2030, and was able to get it for a cheap price in 2023.
---
The above seems exactly analogous to the case with existential risk.
---
Suppose I think bonds are overvalued, because in 2030 the world is going to blow up.
1. I short real rates today.
2. Possibility 1: before 2030, the market “corrects” and realizes that it was overvaluing bonds. Rates rise, and I receive capital gains.
3. Possibility 2: the market does not “correct” before 2030. I still was able to take out a cheap loan in 2023 (i.e. by selling short bonds), and don’t have to pay it off in 2030 when the world ends.
1. Very interesting, thanks, I think this is the first or second most interesting comment we’ve gotten.
2. I see that you are suggesting this as a possibility, rather than a likelihood, but I’ll note at least for other readers that—I would bet against this occurring, given central banks’ somewhat successful record at maintaining stable inflation and desire to avoid deflation. But it’s possible!
3. Also, I don’t know if inflation-linked bonds in the other countries we sample—UK/Canada/Australia—have the deflation floor. Maybe they avoid this issue.
4. Long-term inflation swaps (or better yet, options) could test this hypothesis! i.e. by showing the market’s expectation of future inflation (or the full [risk-neutral] distribution, with options).
(duplicating from LW)
Thanks for these comments. In short, to all of your questions, the answer is “yes”. Some specific comments:
1. This is perhaps already clear, but it might be worth emphasizing that the economic logic is: real rates are particularly use for forecasting, since the sign of the effect is rather unambiguous for the TAI scenario; but it’s possible the expected returns could be higher for trading on other bets, if you’re willing to make stronger assumptions (e.g. “compute will be important”).
2. Re: equities, the appendix post (especially #4 there) summarizes how we’re thinking about this. To spell out a bit more:
An approximation for stock pricing is the Gordon growth formula, , where
P is stock price (i.e. market cap)
D is some initial level of dividends
r is the real rate
g is the growth rate of dividends over time
For the equity market as a whole, a natural approximation is that the growth rate of dividends equals the growth rate of the economy. And as we pointed out in section I, a first-order approximation for the Euler equation under certainty (“the Ramsey rule”) is
Combining the Ramsey rule and the Gordon growth formula, we have
How to interpret this? As a benchmark, suppose theta=1. That’s log utility (which I think is the benchmark used in a lot of EA, e.g. at OpenPhil, and has some support in the literature). Then you have P=D/rho. That is, price is future profits discounted by your rate of time preference—raising or lower the growth rate doesn’t affect the stock price at all, because it ‘cancels out’ in a specific way.
So, that denominator is picking up the ‘Merton optimality’ that you mention. And I guess the reason I wrote all of this out was to reply to this:
Perhaps with equities you might expect both returns and the interest rate to rise by 3%, which would cancel out
Yes! But also they might not cancel out. It could go either way depending on theta ¯\_(ツ)_/¯. To my knowledge it’s an active area of debate (‘financial economists think theta < 1, macroeconomists think > 1’).
If you really want to nerd out, Cochrane has extended wordy discussion here and Steinsson has long slides here (theta is the inverse of the elasticity of intertemporal substitution).
This is perhaps more than you asked for, and yet I’m not sure if this answered exactly what you were asking. Let me know if not!
I’ll just pop back in here briefly to say that (1) I have learned a lot from your writing over the years, (2) I have to say I still cannot see how I misinterpreted your comment, and (3) I genuinely appreciate your engagement with the post, even if I think your summary misses the contribution in a fundamentally important way (as I tried to elaborate elsewhere in the thread).
Thanks for this interesting exercise. The one caveat I’d note is that the multiplier you use is based on annual revenue—if the remittances from OpenAI to MSFT occur over a number of years, we would need to divide the $1T number that you calculate by that number of years.
(PS: amazing tiktoks)
Thanks Seth, we’ll read your paper carefully. I’ll just highlight that really the purpose of the analysis above is to engage specifically with the extreme scenario you mention at the end
what the effect would be today of a giant anticipated increase in automation in 30 years, or of everyone dying with certainty in 30 years [...] we could try if anyone thinks it would be interesting
Also note we briefly allude to demographic trends, but in the (blog post!) analysis here, we want to ignore them because they seem plausibly swamped by the huge growth/mortality scenarios under consideration. As a quick BOTEC:
We use
If—as you suggest—we model demographic trends in reduced form as a decrease in rho, then
At most the demographic effects could shrink our estimates of the increase in the real rate by one percentage point. Of course, that’s just in this simple rep agent model (TIABPNAJA; econometrica isn’t going to accept this!)
(PS: some of your other papers which I’ve already read before, I’ve found useful to read!)
1. We would welcome engagement from you regarding our argument that stock prices are not useful for forecasting timelines (the sign is ambiguous and effect noisy).
2. You offer what is effectively a full general argument against market prices ever being swayed by anything—a bit more on this point here. Price changes do not need to be driven by volume! (cf the no-trade theorem, for the conceptual idea)
3. I’m not sure if this is exactly your point about prediction markets (or if you really want to talk about total capital, on which see again #2), but:
Sovereign debt markets are orders of magnitude larger than PredictIt or other political prediction markets. These are not markets where individual traders are capped to $600 max positions and shorting is limited (or whatever the precise regulations are)! Finding easy trades in these markets is …not easy.
Raimondo and the Department of Commerce seem to have been remarkably effective on AI/China issues during the Biden administration. Is there any detailed reporting on how governance became (seemingly) so good there?