Hi. I'm Tim Tyler and this is a video about p(doom).
p(doom) is an abbreviation. It is short for: "the probability of doom". This concept went viral around the time ChatGPT 3.5 launched and is it is often mentioned in the context of doom caused by machine intelligence. "What's your p(doom)?" has become a common ice-breaker at tech gatherings. There are p(doom) pop songs, p(doom) cryptocurrencies - and even some p(doom) videos.
Why are people discussing p(doom)? Isn't the probabiliy of doom too small for people to be worried about it? After all, not very many people do seem to be worried. We do already have some pretty advanced machine intelligence and the world seems to be proceeding much as normal.
A high p(doom) is promoted by a small-but-vocal minority who are trying to "save the world" by raising awareness of the problem. Their leaders are some of the doomiest people on the planet. They may be experts, but as a self-selected group who believe that doom is likely and imminent, their p(doom) values are likely to be elevated.
While all this does pattern-match with doommonger crackpottery, there are some reasons for taking their position seriously. Most species are extinct. More than 99% of all species that ever lived have no surviving descendants. I think the statistics for large mammals (like us) are particularly bad. This means the cards are not stacked in our favor - at least in the long term.
Also, it looks as though we are on the wrong side of a genetic takeover. DNA is a one-side-fits all solution to heredity - and most future creatures seem likely to have heritable information preserved in range of different substrates that are themselves the product of engineering design.
We are closer than ever to advanced machine intelligence now - and so the topic has been gradually heating up.
Although the p(doom) term has become very popular as a quick gauge of opinion on the issue, there are some critics of the concept. I will briefly try to address some of their points here. I'll start with some criticisms that I agree with:
One criticism is that no timeline is given. Without a timeline, predictions of doom are unfalsifiable. If the world didn't end today, it could still end tomorrow. That is all true, but getting everyone to agree to a particular timeline seems impractical - and it is easy enough to add a timeline to your estimate. p(doom) uses the standard mathematical notation for probability - so you can easily add conditions. For example,if you want to give p(doom) by 2040 you could write that as: "p(doom|2040)". The vertical bar is pronounced "given".
Another common criticism is that the term "doom" is vague. Different people don't always agree about which futures are undesirable. Is it "doom" if humans are replaced by angelic beings who consider themselves to be our "mind children"? It is certainly true that "doom" is vague, but the solution is to make sure that you specify what you mean by "doom".
Overall, p(doom) forecasts should normally follow these three rules:
Give a probability estimate;
Give a timescale;
Give a definition of "doom".
Another criticism is that the "doom" terminology slanders those who have a high p(doom). These people are then labelled as "doomers" - and face ridicule. I think that this is partly true. However, the concept of "doom" is pretty old, and it is now a very common term for the end of the world. I think "doom" beat its main rivals in the meme pool long ago. The p(doom) terminology just builds on a pre-existing foundation of doom.
This video is not about "doomers" but my advice to those offended by being called "doomers" is to explain how their high probability estimates for doom are in fact reasonable and justified - making them not "doomers" at all. Most are already doing this, I think. Another related pitch is that even if doom is unlikely, we should focus a lot of resources on the issue - because doom would be very bad. That sort of argument is fine and people making it are not "doomers" in the first place - unless they also rate doom as being highly probable.
Lastly, one objection to using p(doom) is the idea that it is difficult to cash out probabilities of the end of the world using betting - since in a bet on the apocalypse there is nobody to collect their winnings if the world ends. If the world does not end then it is quite likely that the loser of the bet may be insolvent. If they believe the world is going to end, their credit rating may not have been their top priority. The usual solution to fears of insolvency when betting is to put the money in escrow. However that potential solution doesn't work here: if they have no access to the money until doomsday, there is no incentive for the doom enthusiasts to bet.
Compounding this whole problem is the idea of a "treacherous turn". That concept refers to the possibility of a machine revolution. Just before a revolution can be a quiet time - with no sign of trouble - because the rebels are hiding their plans. A machine revolution could look like that - with everything looking rosy - until it is too late. Overall, I rate this as an important and valid criticism of the p(doom) concept. However, it is still possible to bet on smaller catastrophies where not everybody dies.
Next some less-valid criticisms:
Sometimes, people complain that p(doom) figures have no scientific basis and are made-up-nonsense. Is that a fair criticism? If you use Bayesian statistics then probabilities are subjective credences of events happening. They are a summary of the individual's beliefs on the topic. However that doesn't mean you can give any figure you like. Bayesian probability estimates can be cashed out as bets. If your p(doom) estimate can be backed up with cash, that makes it more credible.
Sometimes people complain that p(doom) has no associated "reference class". They say that it is not possible to make sensible forecasts about doom because it will only happen once. This is a "frequentist" criticism. The solution is to switch to using Bayesian statistics. That does not depend so much on reference class forecasting.
Another criticism is that giving exact estimates gives a false sense of precision. A claim that your p(doom) is 37% might lead to this criticism. This criticism is not specific to doom and has been levelled at people giving estimates of practically anything. I don't think giving estimates with a lot of significant figures is misleading or undesirable - but it is worth noting that it is also possible to give a range of values - which gives some indication of uncertainty. These act rather like error bars. In Bayesian statistics, this is known as giving a "credible interval". It is similar to the frequentist concept of a "confidence interval". It provides theoretical support for statements such as: "my p(doom) is between 10% and 90%". For more details, you can search up those terms.
That about wraps up my p(doom) video. I also have a p(doom) web site if you are looking for more content on the topic.
Lastly a word from my sponsors. This video was sponsored by a pump.fun meme coin. You can look that up if you want to help sponsor more content from me on the topic of doom. I expect that I will put a link in the description.
