Resolves to YES if we have public knowledge by 12/24/27 of a company or other entity having created something that roughly equates to the 'drop-in remote worker' described in the paper Situational Awareness, or is otherwise similarly or more capable. It need not be available to the public, if it is being used only internally.
Resolves to NO if this is not the case.
If Leopold Aschenbrenner says this did or did not happen I will abide by his observation. If he does not do so, I will use my best judgment.
1,000People are also trading
@JonasVollmer Here are two paragraphs from the essay which include descriptions of what Leopold imagines:
"An agent that joins your company, is onboarded like a new human hire, messages you and colleagues on Slack and uses your softwares, makes pull requests, and that, given big projects, can do the model-equivalent of a human going away for weeks to independently complete the project.
...
Again, critically, don’t just imagine an incredibly smart ChatGPT: unhobbling gains should mean that this looks more like a drop-in remote worker, an incredibly smart agent that can reason and plan and error-correct and knows everything about you and your company and can work on a problem independently for weeks."
Seems to require a very general worker. However, I don't think Leopold implies that such a worker needs to be a pareto improvement over virtual human workers in all domains:
"Still, in practice, I do expect somewhat of a long tail to get to truly 100% automation even for the job of an AI researcher/engineer; for example, we might first get systems that function almost as an engineer replacement, but still need some amount of human supervision."
@AdamK @ZviMowshowitz How will you resolve this if we "get systems that function almost as an engineer replacement, but still need some amount of human supervision"?
@ZviMowshowitz Suggestion for how to resolve this:
1. Resolves YES if it needs as much supervision as a human remote worker per amount of work done, NO if it requires substantially more supervision.
2. Needs to be a general remote worker, not just a software engineer. Should be able to do coding, writing, copyediting, bookkeeping, customer support, graphic design, project management.
People holding yes on this question, I'd be interested understanding the answers to any of the following:
1. How much would your forecast change if this question was resolved directly according to Zvi's judgement, rather than him deferring to Leopold? Leopold has a modest incentive to confirm his own predictions, and I'm curious how much that shifts forecasts.
2. Do you believe the timelines predictions in Situational Awareness are roughly correct or relatively high probability, beyond just the "drop-in remote worker" prediction? In the same page where Leopold makes this prediction he says: "We are on course for AGI by 2027. These AI systems will basically be able to automate basically all cognitive jobs (think: all jobs that could be done remotely)." It's unclear to me whether he and others view this as the same as the "drop-in remote worker" prediction or different and stronger. He also makes many other timeline related prediction for after the arrival of drop-in remote workers, and I'm curious if YES holders have similar beliefs on timelines as him.
3. Are you changing your life decisions based on the extent to which you believe that drop-in remote workers will arrive before 2027? For example, changing your career, working more or less, spending or saving more or less, or investing differently? Believing that automation of most cognitive jobs is near would justify more drastic actions than trying to win a modest amount of mana, I would think. Or is this question resolving YES very different from "automation of most cognitive jobs".
@MaxMorehead 1. I'd go down from like 60% to maybe 50%ish. 2. Yep, Leopold is right that this is sorta the standard view among people most in the know. 3. Yep, my life has been heavily influence by these beliefs since 2020, when I first came to have them.
@MaxMorehead I'm noticing a hole in this, same as a lot of AI bets. Simply - do we mean that say 50% of the knowledge worker economy has been replaced by 'drop in' agents, or do we mean
there is a service you can rent.
It's been demoed.
The underlying AI model is general and has no glaring weaknesses. Meaning it has at least a. online learning,
b. model will evaluate it's answer to a task before submitting,
c. mode has a memory representation for 3d space,
d. model can perceive video.
e. I want to say it also has robotics manipulation abilities since almost all human knowledge workers have built stuff with their hands. Even pure SWEs have actually used their phones at a tactile level, and can perceive things like juddering and a lack of responsiveness to swipes - things that indicate a bad user experience.
But even then, in 2027, if (3) is satisfied, it's pretty unlikely that the actual economy will have been converted yet. There will also be many knowledge work tasks where it is not a drop in replacement. You need to give the model feedback, a lot of it, much more than a new employee, for it to be competent at a task. There also will be jailbreaks and other ways to exploit a model, same way social engineering works on humans.
Probably also it won't be 'drop in' in that companies may need to structure their daily workflow in a way that helps the models make good choices. For example one obvious thing to do is have .json files that in detail explain what a particular job role is, with many example situations, and then you have a model employee assume that role. The .json files and other artifacts are kept on github.
There would be people working at a company that make patches to this github as bugs and mistakes are found. (HR SWE hybrids?)
It will take new startup companies outcompeting legacy firms, and acquisitions, to actually replace 50% of knowledge worker jobs. This could take years.
Typing this out, I think I disagree on 'drop in' replacements by 2027. I don't think that is happening. Voting no.
@GeraldMonroe The resolution criteria says "It need not be available to the public, if it is being used only internally.". That means a model with the potentiality of being a drop in model for the whole economy.
@Tossup Here are those markets: