I'm defining AGI as an algorithm that can do any human white collar job. If things get complicated I might resolve if a single algorithm can perform better than humans on all the most common AI benchmarks in text, vision, and audio.
The common LLM benchmarks: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
The most common vision tasks:
https://scale.com/blog/best-10-public-datasets-object-detection
And the Audio Tasks / Datasets mentioned here:
https://huggingface.co/blog/audio-datasets
If a single algorithm can do better than a human on all of these tasks, I think it is almost certain that it can do most any desk job. And I would resolve this question with the creator of that algorithm.
1,000
3.00Interesting update. The US government may be trying to fund an AGI Manhattan Project : https://x.com/deanwball/status/1858893954982232549
@ElliotDavies yes, and as a result, the "US government may be trying to fund an AGI Manhattan Project".
Don't star me unless your correctly identifying a mistake. I'm intentionally giving you the assist here for learning purposes.
@ChrisCanal It's always good to write with as much precision as possible - I believe your statement could mislead others into thinking this came out of the Biden administration. This would be a much stronger claim than a policy recommendation by a commission.
@ChrisCanal I definitely expected a higher caliber of conversation and epistemics when I originally joined manifold lmao
What if AGI achieves a solution that uses different LLMs as required, e.g. the notdiamond.ai metamodel, which currently outperforms all LLMs in benchmarks? Would something like this be considered AGI?
@Linch I don’t think writing a best seller tells us enough about general capabilities. But if the same AI can do many things, like exceed human performance on the benchmarks I mentioned, then it is more indicative of AGI, than only being good at writing.
@ChrisCanal if no one creates AGI by 2030 what will you do? Extend resolution time, resolve Other, or N/A?
@parhizj It hadn't crossed my mind as possibility, but it feels like I should extend. I really really don't think this will come into play though. Why do you think it should be NA @ElliotDavies ?
@ChrisCanal well resolving "other" would be a misresolution (it would clearly imply some other org had created agi.
You could extend, but you should probably do that sooner rather than now, because resolution dates are typically considered an extension of the resolution criteria. Modifying resolution dates modifies the relative odds
@MatthewLichti It's quite possible that market has already turned positive, but it's difficult to test because of finetuning
I was just informed that Google Brain and DeepMind merged a few months back. So I think if this team is first I will resolve DeepMind to be the winner, so no one buy google brain. Sorry. https://www.deepmind.com/blog/announcing-google-deepmind
