(The whole series of essays has been fantastic.) 

Have you put any thought into whether two of your points can be combined?

1. The constant hazard rate model: there is a predictable difference between T_"50% success" horizon vs. T_"X% success" horizon.
2. Costs matter, and costs are plausibly rising fast.

In particular, can one use the constant hazard rate model -- and other information? -- to go from "data on costs to achieve 50% success" to extrapolate "cost to achieve 99% (e.g.) success". 

I spent a bit of time thinking about this, but I think there's a missing ingredient: for example,

• If you can do a 70-minute task, at 50% reliability, for $100
• Then (per constant hazard rate) you can do a 1-minute task, at 99% reliability, for $X

The difficulty is: what can we say about $X?

• Presumably it is upper bounded by $100, the cost of doing a 70-minute task at 50% reliability
• Presumably it is lower bounded by "cost of doing a 1-minute task at 50% reliability"

==> basically, as you wrote, would love more data from METR on costs...!

(Details: I had ChatGPT attempt to digitize the cost curve for GPT-5 from METR, and then generated the upper and lower bounds as described above)

We're looking at real interest rates, though! So the scenario you envision of "5% nominal rates and 25% deflation" would still show up as 30% real rates. So the monetary policy reaction function is really totally irrelevant in this sense.

PS: your 2011 blog post on a related topic has sat in my head for over a decade ("Piecing together optimal monetary policy in the singularity is a side project of mine and though it sounds silly possibly very important"... I am also interested in this topic 😊)

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?

(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.