This market made in response to this article. Quote:
"
One-hour AGI: Beating humans at problem sets/exams, composing short articles or blog posts, executing most tasks in white-collar jobs (e.g., diagnosing patients, providing legal opinions), conducting therapy, etc.
One-day AGI: Beating humans at negotiating business deals, developing new apps, running scientific experiments, reviewing scientific papers, summarizing books, etc.
One-month AGI: Beating humans at carrying out medium-term plans coherently (e.g., founding a startup), supervising large projects, becoming proficient in new fields, writing large software applications (e.g., a new operating system), making novel scientific discoveries, etc.
One-year AGI: These AIs would beat humans at basically everything. Mainly because most projects can be divided into sub-tasks that can be completed in shorter timeframes.
Although it is more formal than the definitions provided in the previous section, the (t,n)-AGI framework does not account for how many copies of the AI run simultaneously, or how much compute.
As of the third quarter of 2023, we can establish a rough equivalence “from informal initial experiments, our guess is that humans need about three minutes per problem to be overall as useful as GPT-4 when playing the role of trusted high-quality labor.”(source) So existing systems can roughly be believed to qualify as one-second AGIs, and are considered to be nearing the level of one-minute AGIs.
They might be a few years away from becoming one-hour AGIs.
"
I took objection to the article stating "might be a few years away". To me, that implies probably at least more than 2 years, perhaps more like 10 years. I think it will be less than 2 years. So here I am, putting my money where my mouth is.
Edit:
Based on questions asked, here are some more details:
This is AI (plus any tools the human is granted, e.g. web search, but excluding other AI) versus human (web search, code ide, and calculator, but no explicit AI).
Secret knowledge that the human has which the AI can't look up on the web wouldn't count. This is about ability rather than knowledge.
The human for the comparison should be an above-average, competent professional, but not a remarkable genius.
1,000
3.00
@NathanHelmBurger Whatever the benchmark, it seems unlikely to be representative. If truly AI could do 30 minute tasks comparably to typical humans, you'd see massive adoption in areas like consumer support. Yet we don't see that.
@EliezerYudkowsky good question. I was thinking above-average, competent professional, but not a remarkable genius. Does that clarify?
I am personally interested in gaming. One guy in Dead by Daylight won 1000 matches in a streak. That game involves a lot of fifty-fifties, where you have to just guess what opponent does based on his personality earlier in match.
Currently people are better at games than ai if learning iterations is the benchmark. Human needs 2-10 matches to understand chess, neural networks need to simulate millions of matches to grasp the basic concepts.
Opening a business and similar tasks are not interesting. Wisdom about those things is definitely in training data. But facing A NEW challenge, a new game, being able to learn on the go is interesting, is closer to human brains.
I have no doubt AGI will appear soon, but I look forward to the next step in the research - system with no memory which is dropped in the real world and is able to learn instead of being taught.
Thanks for sharing! Great market and that is for referring to the source! Btw, in order to win, does AI have to be better than humans without that AI or also with humans + that AI (making humans basically not only inferior but totally useless)? And how can you measure whether AI is able to do ANY task (I can expect that at least some tasks will require exclusive knowledge that only some specific people have, say about a corporate system or about a very niche topic).
@SimoneRomeo This is AI (plus any tools the human is granted, including web search but excluding other AI) versus human (web search, code ide, and calculator, but no explicit AI).
Secret knowledge that the human has which the AI can't look up on the web wouldn't count. This is about ability rather than knowledge.
The human for the comparison should be an above-average, competent professional, but not a remarkable genius.
@NathanHelmBurger I get the idea, but drawing the line between ability and knowledge is very difficult. I guess you'd count writing as an ability but you still have to study and practice. Same to be a lawyer or most of other tasks. As a comparison, if you took a totally smart human hunter gatherer and you gave them a computer, Devin could already be more competent than them right now