You can use ChatGPT 3.5 for free with chat history turned off. This way your chats should not be used as training data.
Sune
Karma: 523
The corporate structure of OpenAI was set up as an answer to concerns (about AGI and control over AGIs) which were raised by rationalists. But I don’t think rationalists believed that this structure was a sufficient solution to the problem, anymore than non-rationalists believed it. The rationalists that I have been speaking to were generally mostly sceptical about OpenAI.
They were not loyal to the board, but it is not clear if they were loyal to The Charter since they were not given any concrete evidence of a conflict between Sam and the Charter.
I don’t understand how this is a meaningful attitude to your own private economy. But want to donate to someone who needs it more is also a way to spend your money. This would be charity, possibly EA.
I have noticed a separate disagreement about what capitalism means, between me and a family member.
I used to think of it as how you handle your private economy. If you are a capitalist, it means that when you have surplus, you save it up and use it (as capital) to improve your future, i.e. you invest it. The main alternative is to be a consumer, who simply spend it all.
My family member sees capitalism as something like big corporations that advertise and make you spend money on things you don’t need. She sees consumerism and capitalism as basically the same thing, while I see them as complete opposites.
Ok, looks like he was invited in to OpenAIs office for some reason at least https://twitter.com/sama/status/1726345564059832609
It seems the sources are supporters of Sam Altman. I have not seen any indication of this from the boards side.
It seems this was a surprise to almost everyone even at OpenAI, so I don’t think it is evidence that there isn’t much information flow between LW and OpenAI.
There seems to be an edit error after “If I just stepped forward privately, I tell the people I”. If this post wasn’t about the bystander effect, I would just have hoped someone else would have pointed it out!
Corollary: don’t trust yourself!
Most cryptocurrencies have slow transactions. For AI, who think and react much faster than humans the latency would be more of a problem, so I would expect AIs to find a better solution than current cryptocurrencies.
I don’t find it intuitive at all. It would be intuitive if you started by telling a story describing the situation and asked the LLM to continue the story, and you then sampled randomly from the continuations and counted how many of the continuations would lead to a positive resolution of the question. This should be well-calibrated, (assuming the details included in the prompt were representative and that there isn’t a bias of which types of ending the stories are in the training data for the LLM). But this is not what is happing. Instead the model outputs a token which is a number, and somehow that number happens to be well-calibrated. I guess that should mean that the prediction make in the training data are well-calibrated? That just seems very unlikely.
Two possible variations of the game that might be worth experimenting with:
Let the adversaries have access to a powerful chess engine. That might make it a better test for what malicious AIs are capable of.
Make the randomisation such that there might not be an honest C. For example, if there is 1⁄4 chance that no player C is honest, each adversary would still think that one of the other adversaries might be honest, so they would want to gain player A’s trust, and hence end up being helpful. I think the player Cs might improve player A’s chances of winning (compared to no advisors) even when all the adversarial.
I think the variations could work separately, but if you put them together, it would be too easy for the adversaries to agree on a strong-looking but losing move then all players Cs are adversaries.
Why select a deterministic game with complete information for this? I suspect games like poker or backgammon would be easier for the adversarial advisors to fool the player and that these games are a better model of the real world scenario.
This seems like the kind of research that can have a huge impact on capabilities, and much less and indirect impact on alignment/safety. What is your reason for doing it and publishing it?
How about “prediction sites”? Although that could include other things like 538. Not sure if you want to exclude them.
In case you didn’t see the author’s comment below: there is now a patreon button!
Sorry my last comments wasn’t very constructive. I was also confusing two different critisisms:
that some changes in predicted probabilities are due to the deadline getting closer and you need to make sure not to claim that as news, and
that deadlines are not the in headlines and not always in the graphs either.
About 2): I don’t actually think this is much of a problem, if you ensure that the headline is not misleading and that the information about deadlines is easily available. However if the headline does not contain a deadline, and the deadline is relevant, I would not write any percentages in it. Putin has a 100% chance of dying, just like the rest of us, so it doesn’t make sense to say he has 90% chance of staying in power. In that case, I would prefer the headline to just state the direction the probability is moving in, e.g. “Putin hold on power in Russia is as stable as 3 month ago” or something like that.
To avoid writing “by 2024” in all headlines, maybe you could create subsections of the site by deadline. It would be a good user experince if you could feel like you are scrolling further into the future, starting with predictions for start of 2024, then 2025, then 2030. Of course this requires that there are several predictions for each deadline.
About 1), I think you should only include predictions if they cannot be explained by the nearing deadline.
For some questions this is not a problem at all, e.g. who is going to win an election.
For questions about whether something happens within a given timeframe, the best solution would be if prediction sites started making predictions with constant timeframe (e.g. 1 year) instead of constants deadline. I made a feature request about this to metaculous. They said they liked the idea, but they did not provide any prediction of the probability it would be implemented!
An alternative is to ask for a probability distribution for when something is going to happen. Such questions already exists on metaculous. Then you can see of expected remaining time or if time until median prediction is increasing, or something similar.
For questions with a fixed deadline, if the predicted probability of something happening is increasing, you can still conclude that the event is getting more likely.
For questions with fixed deadline and declining probabilities, it is harder to conclude anything. The very naive prediction would be linear decline, so p(t)/t is constant, where t denote time left and p(t) the probability at that time. E.g. with one month left t=1/12. A slightly less naive solution would be to model the event having constant probability at any time given that is hasn’t already happened. In this case, the constant would be log(1-p(t))/t.
In this model, if the probability is declining faster, meaning that |log(1-p(t))/t| is decreasing, I would stay that is a signal that the probability of the event is getting lower.
If p(t) is declining, but slow enough that |log(1-p(t))/t| is increasing, I would not conclude anything based on that, at least not on metaculus, because people forget to update their predictions sufficiently. I’m not familiar with other prediction sites, maybe it works better when there is money at stakes.
However, this model does not apply for all questions. It would be a useful model for e.g. “will there be a big cyber attack of a given type by a given date” that happens without warning, but for other questions like “Will Putin loose power be a given date”, we might expect some further indication of it happening before it happens, so we would expect the probability to go to 0 faster. For such questions questions, I don’t think you should ever conclude that the underlying event is getting less likely for a single fixed-deadline question.
So to conclude: if predicted probability of some event happing before the deadline is going up, it is fair to report it as the probability going up. If the prediction is going down you should only in rare cases conclude on the. The rare case is if you think the event would happen without much notice and |log(1-p(t))/t| is decreasing.
I think this is a great project! Have you considered adding a donation button or using Patreon to allow readers to support the project?
I do have one big issue with the current way the information is presented: one of the most important things to take into account when making and interpreting predictions is the timeframe of the question. For example, if your are asking about the probability that Putin losses power, they the probability would likely be twice as high if you consider a 2 year timeframe compared to a 1 year time frame, assuming the probability per month does not change much.
Currently, the first 5 top-level headlines all ignores the timeframe, making the headlines meaningless: “Putin >90% likely to remain in power, no change” “Odds of ceasefire ~15%, recovering from ~12% low” “Russia ~85% likely to gain territory, up from ~78%” “Crimea land bridge ~30% chance of being cut, no change” “Escalation: ~5% risk of NATO clash with Russia, down from ~9%”
The last one is particularly misleading. It compares the probabilities from start April to the probabilities now (start June). But one of the markets have deadline on June 12th and the other prediction have deadline July 1, so it is not surprising that the probability is down!
In order to conclude that the risk is decreasing, the question should have a moving deadline. I’m not aware of any prediction sites that allows questions with a moving timeframe, although it would be a great feature to have.
I have a question that tricks GPT-4, but if I post it I’m afraid it’s going to end up in the training data for GPT-5. I might post it once there is a GPT-n that solves it.