What is/was your total monetary compensation at both jobs? Or if you don't want want to say absolute numbers, what is the relative change in compensation?

You said that you believe that the AI bubble is going to burst in the next few months. Could you phrase that as a testable prediction? For example, you're 80% sure that in 6 months, the stock market price of Nvidia will be 80% of what it is today. This would help me understand your prediction better. Different people have different views on what "pretty certain" and "burst" means. I just want a clarification. I don't want to argue about the prediction.

You asked people to try resolving the problem on their own. So here is my answer before I read further. I was motivated to find a reason for the doomsday math to be incorrect by your prompt.

I felt the mistake was the hidden ambiguity of this statement:

Assuming that you are a random person among all the people destined to be born.

How do we randomly sample the person? Imagine you had to code this as a simulation in Python. Which of the following algorithms does your program follow?

1. Pick uniformly randomly between Short and Long. If Short, then pick a number between 1 and 100 billion. If Long, then pick a number between 1 and 100 trillion.
2. Create a list ["Short-1", "Short-2", ..., "Short-100-billion", "Long-1", "Long-2", ..., "Long-100-trillion"]. Pick a uniform random sample from this list.

The math you present matches the first algorithm. The math is different with the second algorithm. With the second algorithm we have:

• P(sample beginning with Short) = 1/1001
• P(sample beginning with Long) = 1000/1001
• P(sample ending with 6 | sample beginning with Short) = 1/10
• P(sample ending with 6 | sample beginning with Long) = 1/10000
• P(sample ending with 6) = P(sample ending with 6 | sample beginning with Short) * P(sample beginning with Short) + P(sample ending with 6 | sample beginning with Long) * P(sample beginning with Long) ≈ 0.0002
• P(sample beginning with Short | sample ending with 6) = P(sample ending with 6 | sample beginning with Short) * P(sample beginning with Short) / P(sample ending with 6) = 0.5

This matches the following experiment:

1. Run the program to pick a random sample.
2. If your sample does not end with 6, then restart.
3. You have a sample ending with 6. Check whether it begins with Short or Long.

You observe Short and Long with equal probability of 0.5 .

Update?

Invariably, within 10 minutes, I receive messages from coworkers congratulating me on my boldness. You asked him? That question? So brave!

How do you respond to those messages?

Could you make it more clear how you would change the original Harris market to avoid the problem? I'm not sure what exactly you mean by "base question and a conjunctive market". I figure you're talking about the "other" option in the market you link, which prevents N/A. But how do you add that to the Harris market? Is it just adding one more option "Another person not listed here is nominee and Harris wins"?

I read your post but didn't understand it immediately. After thinking about it, I wrote what I feel the main points are in my own words. Maybe this makes it more clear to others too. Or maybe the OP will comment that I misunderstood something.

The "Harris market" can be seen as (is functionally equivalent to) a collection of markets. Each market has the following form:

Would Harris win, if {person} were her Vice Presidential nominee in 2024?

Each market resolves as follows:

• Yes, if {person} is the nominee and Harris wins.
• No, if {person} is the nominee and Harris loses.
• N/A, if {person} not the nominee.

Each market is independent. I mean this in the sense that there is no manifold mechanic that ties them together, except that they are visually grouped under "the Harris market".

There is a problem with this kind of market. When people see a high probability of a person in the market, they mistake it as meaning that the person is a likely choice or a good choice.

This is a mistake because it does not follow from the market's incentives. Imagine someone that will obviously not be nominee, like "Batman". You know that Batman will not be nominee, so you know that this market will resolve to N/A. This makes bets on Batman meaningless. You can bet Yes or No, it doesn't matter. The market will resolve to N/A and you get your money back. Some people do this because they find it fun to see that Batman has a high Yes probability.

Batman can't be nominee because he doesn't exist, but you have the same problem for obviously terrible candidates. Imagine the person most hated by all potential Harris voters, "the Joker". (Assume Joker was a real person.) The true probability of Harris winning with Joker is tiny. Yet, in the market, there is little incentive to correctly predict this. Everyone knows Joker is a terrible choice, so Harris is not going to choose him as nominee, so the market will resolve N/A.

The author of the Substack you quote makes this mistake. They did not consider what the market incentivizes.

They also made a separate mistake that applies to all prediction markets, not just misleading conditional markets. They did not take into account how much money it takes to intentionally influence the market. It is easy (cheap) for people to push small markets (markets with little money invested) in their desired direction.

I agree and I am putting my money where my mouth is.

I will play this game under the rules linked in the OP with me as the gatekeeper and anyone as the AI. I will bet at odds 10:1 (favorable to you) that I will not let the AI out. The minimum bet is my 10 k USD against your 1 k USD and the maximum bet my 100 k USD against your 10 k USD. We will use a mutually accepted LW community member as referee.

If you believe you have at least a 10% chance of making me let the AI out, you should take this bet. I predict no one will take me up on it.

I speculate that the reason that gatekeepers let the AI out is that they do not take the experiment seriously enough. They don't really care about not letting the AI out. Maybe they want the outcome to be that the AI has been let out so that people take AI safety more seriously. I'm not saying an actual superintelligence couldn't do it, but no currently available intelligence can (with me as the gatekeeper).

The bet has been lost. The merge wasn't allowed. https://archive.ph/MDYWF

The share price of Spirit Airlines (NYSE: save) dropped to 50%.

I consider it unlikely that my interactions with ChatGPT 3.5 (that led to my original post) played a significant role as training data that helped make ChatGPT 4 so much better.

Why do you consider it unlikely?

If I was good at memorization (model parameter size) but bad at reasoning then your original post showing up in my training data would help me.

I'm also curious how much of the improvement comes from improvements to the hidden and secret prompt that Openai adds to your Chatgpt interactions. Unfortunately we can't test that.