We need to start including cost in these markets.

o1-preview is incredible, but it's useless because it's ridiculously expensive and limited. If we keep basing progress off of models that are only usable by a select few, then it's not really progress.

If the model is price competitive with, for example, software engineers, it is really progress. and they seem to be? does it cost more than 50$ per hour of "thinking"? The cost can also go down a lot from here.

@Bayesian Even if it's cost-competitive, there are other issues.

1. It has to be open source. You can't build a business around a model owned by a company when the law allows that company to terminate service to you for any time for any reason. I stopped all use of E-Mail marketing firms for that reason, because they terminated service to me.

2. The model has to actually be available, and o1-preview is not. The world is clearly limited by manufacturing capacity and will be for some time. I would pay $1000/month to use the model but nobody will offer me that because the GPUs don't exist.

These models are clearly "AGI," without a doubt. But I've always thought that @EliezerYudkowsky 's ideas that the world is suddenly going to change overnight were ridiculous. AGI hasn't, and won't, change the world for some time because intelligence is not our limiting factor.

1. actually you can. Lots of software is not opensource and is used by businesses all of the time

2. How is o1-preview not available?

>- OpenAI o1 (or any direct iteration) accomplishes a comparable feat as the model that caused the linked market to resolve YES

What do you mean by "comparable feat"? Does it have to solve in the same time limit human competitors get?

@mongo Will resolve similarly to the linked market, so probably with time limit

So if a model called o2 gets it, this resolves no?

@ChrisPrichard It should still resolve Yes.
The original description included "future iteration". I changed it slightly to make it more explicit that the new iteration does not have to be called o1 (but has to be directly based on o1).

@Simon74fe Then it seems much more likely to me! Basically, will something that uses a lot of test time inference succeed. It doesn't need to be the first to do it to resolve yes, right? Like maybe some specialized model does it first and then o3 with gpt-5 backing it can also do it and this resolves yes.

Given the title says "o1" this is very confusing. And "directly based on" - what does that mean, if OpenAI builds GPT 5 which iterates on o1 does that count too?

Updated the description again and included it in the title. GPT-6 would most likely not count even if it uses some insights from o1, unless it is explicitly marketed as a direct iteration of o1.

Sorry everyone for the confusion. If you think you lost mana because of any update, please let me know and I will refund you.

@Simon74fe thanks that helps