Fermi–Dirac Distribution

504Joined Apr 2022


Note that these numbers include illiquid (pre-exit) valuations.

2021 and early 2022 was a huge bull market — many private companies raised money in insane valuations that aren’t realistic today. Suppose they didn’t raise money again or go public in mid/late 2022. In that case, it’s likely that their valuations were last updated around the bull market's peak, in which case they're almost certainly systematically inflated. 

As mentioned in that section, we clearly didn’t prioritise or eliminate this risk - we’re reflecting on how much that was a mistake, versus a good decision (to focus on other significant risks). 

Hm, if you don’t mind the question — what are other significant risks you are thinking about? IMO, the FTX fraud scheme is a big deal if anything deserves to be called a big deal, from what I’ve read in the devastating legal documents about the situation, and given how SBF was framed as a poster child of EA and was a significant funder.

The only more significant risk I can think of is EA funding dangerous AI capabilities research via Anthropic, but that isn’t even unrelated to FTX (since FTX slash Alameda slash their leadership was Anthropic’s main funder). Also, my guess is that mitigating AI risk is not within the Community Health team’s scope. 

TL;DR provided by ABC News:

Ellison pleaded guilty to conspiracy to commit wire fraud on customers of FTX, wire fraud on customers of FTX, conspiracy to commit wire fraud on lenders of Alameda Research and wire fraud on lenders of Alameda Research, conspiracy to commit commodities fraud, conspiracy to commit securities fraud and conspiracy to commit money laundering, according to the court documents.

Together, the seven counts carry a maximum sentence of 110 years in prison.

Wang pleaded guilty to four counts: conspiracy to commit wire fraud on customers of FTX, wire fraud on customers of FTX, conspiracy to commit commodities fraud and conspiracy to commit securities fraud, the court documents stated. He faces up to 50 years in prison.

Did anyone else find the title of that post ("CEA Ops is now EV Ops") a bit confusing? 

When I first saw it, I thought, "oh, it seems like a team/sub-org within the CEA changed its name. Cool" and didn't think much else about it. It didn't strike me as important enough for me to click the title and read the post. 

It was only months later, by reading an EA Forum comment in another post, that I noticed that CEA UK (the umbrella org) had changed its name to Effective Ventures. When I read the comment, I spent some time trying to find an official announcement, and it took me a while to realize that it was that post. 

I'm confused about why that title was chosen over titles that made it clear that it was the CEA umbrella org that changed its name, rather than only a specific team within it. 

Damian Williams, the United States Attorney for the Southern District of New York, said the following today:

To any person, entity or political campaign that has received stolen customer money, we ask that you work with us to return that money to the innocent victims.

I know almost nothing about law, so it's unclear to me what this means, but it seems relevant to the topic of this post (and to EA in general), so I'm commenting this here in hopes that someone with more knowledge can shed some light. 

I want to point out something that I find confusing. 

the expected number of lives saved from giving a bednet is infinite.

This can't be true unless your credence that you're killing an infinite number of lives by buying a bednet is exactly zero, right? Otherwise -- if your credence is, say, -- then the expected number of lives saved is undefined. Am I thinking about this correctly?

when we assume the utility of an action can be modelled as a normal distribution, we are allowing for the possibility of negative and positive infinite utility. However, the expected value of the action is still finite (and equal to the mean of the distribution).

The first sentence here is not true. The formula below is the PDF of a normal distribution:

The limit of  as  approaches either  or  is zero. 

Moreover, if the first sentence I quoted from your comment were true, there would be no way for the second sentence to be true. This is the definition of expected utility:

Where  is the utility  of an outcome and  is its probability. 

If you have an unbounded utility function, and you have any probability greater than zero (say, ) that the outcome of your action has infinitely positive utility, and a similarly nonzero probability (say, ) that it has infinitely negative utility, then the formula for expected utility simplifies to

which is undefined.

Also, if you're biting the bullet on expectational total hedonistic utilitarianism, infinities will dominate everything, and you should ignore anything that doesn't have infinite EV.

I think the problem is a bit worse than this? 

If your decision procedure is “maximize the EV of an unbounded utility function,” you basically cannot make any decisions. After all, for any action you could take, there is an extremely low but still nonzero chance that the action is infinitely good, and a similarly low-but-nonzero chance that it is infinitely bad. Infinity minus infinity is undefined. So all actions have an undefined expected value. 

text-davinci-003 (which is effectively ChatGPT)

This is probably a stupid question, but: do we actually know if ChatGPT uses text-davinci-003?

When I talk to ChatGPT with the Network tab of Chrome DevTools open, filter for the name "conversation," and look at any request payload, I see that it has the key-value pair 

model: "text-davinci-002-render"

Which seems to indicate that it might not be using text-davinci-003.

Load More