4 ms·

How do you define AGI?

AndrewKemendo 1y ago
There still is no consensus definition of AGI I made a url redirect to a google form to capture people's thoughts a few months ago but never published the URL Curious what more responses look like. I'll post the results as a comment on the thread once there are more results
- CamperBob2 1y ago
  A common test is something like ARC-AGI 2, where even the best models struggle to get into double-digit percentages. One thing I'd expect from AGI would be to go beyond "Solve this ARC-AGI test" to "Come up with more tests like ARC-AGI." It can't have been lost on the humans who worked to create those tests that their own jobs should be automatable.
  - AndrewKemendo 1y ago
    These tests do nothing in the real world I see no ARC-AGI tests for “make me a cup of coffee” Or “Diagnose this clanking sound in with my truck”
sunscream89 1y ago
Couldn’t we just go back to the “Turing Test?” I know, I know, early “AI” passed that, right! Ha! Could you chat with an “AI” of any kind now and not distinguish from a human? Not in a staged environment, one where you hire a human to do a remote technical job? Let AGI be one who passes a functional remote role as indistinguishable from a human (of any capacity, even one with English as a second language to make subtle idiosyncrasies irrelevant.
- CamperBob2 1y ago
  As with all variations of the Turing Test, that one fails because it says more about the gullibility of the human participant than it does about the intelligence of the computer. It's also arguably been passed already.
  - vunderba 1y ago
    Yep - look at past "winners" of the Loebner prize. The feats of social and psychological manipulation were more impressive than the AI powering 99% of the chatbots. https://en.wikipedia.org/wiki/Loebner_Prize https://en.wikipedia.org/wiki/Loebner_Prize
  - sunscream89 1y ago
    That’s what I’m talking about. Can’t have it both ways. Take some effing responsibility for a competent test by competent test takers. This is not a public gimmick, this is edge humankind. Getting a remote job as an **, indistinguishably from a competent human, that’s a test.
incomingpain 1y ago
Artificial, isnt difficult to define; no debate there. General, means nothing specific like coding, math, or chess. Qwen3 30b thinking is general and most people would agree to be GENERALly useful. I dont see a debate here. Intelligence, people think their dogs are intelligent. This is a spectrum then and perhaps this is debatable because I completely disregard IQ as a fraud. I'm not even debating intellectual disability. I'm also not getting into chronic substance abuse, brain injuries, or neurodegenerative disease. AGI will have greater intelligence than the smartest dogs or dolphins or whatever. So now I can rely on some controversial realities. There are people in the military who arent smart enough to be infantry. If you let them be infantry, it causes more harm than good. These people are put to work in food service, cleaning, guard duties, supply/warehouse duties, etc. AGI intelligence has to meet that level of intelligence. AGI has absolutely happened already. It's happening on 32GB of vram; but not exactly that useful to use. Again, not coding expert by any means. Qwen3 30b coder is not AGI. Qwen3 30b thinking, that's AGI and absolutely better than your low aptitude military grunts. AI has surpassed AGI, at around 150-200GB of vram, you're above the average general intelligence of a human. GPT5 high reasoning is approaching super intelligence. Probably not actual super intelligence, yet.
dlcarrier 1y ago
AGI is a collection of neural networks that can generate $100 billion in revenue. Sue, it's an awful detention of AGI, but it's not like any of them aren't. Whatever happened to post-singularity or strong AI or any of the other terms? It seems the more resources we throw at neutral networks, the less proportional gain we get out of them, and we keep lowering our expectations, often redefining success with new, milder terms. Maybe if we keep developing them we'll get something that approximates thinking, but for now we still don't have large neural networks that remap themselves to train on the fly, something even the most basic living organisms can do, even without a central nervous system.

There still is no consensus definition of AGI I made a url redirect to a google form to capture people's thoughts a few months ago but never published the URL Curious what more responses look like. I'll post the results as a comment on the thread once there are more results

CamperBob2 1y ago
A common test is something like ARC-AGI 2, where even the best models struggle to get into double-digit percentages. One thing I'd expect from AGI would be to go beyond "Solve this ARC-AGI test" to "Come up with more tests like ARC-AGI." It can't have been lost on the humans who worked to create those tests that their own jobs should be automatable.
- AndrewKemendo 1y ago
  These tests do nothing in the real world I see no ARC-AGI tests for “make me a cup of coffee” Or “Diagnose this clanking sound in with my truck”

sunscream89 1y ago

Couldn’t we just go back to the “Turing Test?” I know, I know, early “AI” passed that, right! Ha! Could you chat with an “AI” of any kind now and not distinguish from a human? Not in a staged environment, one where you hire a human to do a remote technical job? Let AGI be one who passes a functional remote role as indistinguishable from a human (of any capacity, even one with English as a second language to make subtle idiosyncrasies irrelevant.

CamperBob2 1y ago
As with all variations of the Turing Test, that one fails because it says more about the gullibility of the human participant than it does about the intelligence of the computer. It's also arguably been passed already.
- vunderba 1y ago
  Yep - look at past "winners" of the Loebner prize. The feats of social and psychological manipulation were more impressive than the AI powering 99% of the chatbots. https://en.wikipedia.org/wiki/Loebner_Prize https://en.wikipedia.org/wiki/Loebner_Prize
- sunscream89 1y ago
  That’s what I’m talking about. Can’t have it both ways. Take some effing responsibility for a competent test by competent test takers. This is not a public gimmick, this is edge humankind. Getting a remote job as an **, indistinguishably from a competent human, that’s a test.

incomingpain 1y ago

Artificial, isnt difficult to define; no debate there. General, means nothing specific like coding, math, or chess. Qwen3 30b thinking is general and most people would agree to be GENERALly useful. I dont see a debate here. Intelligence, people think their dogs are intelligent. This is a spectrum then and perhaps this is debatable because I completely disregard IQ as a fraud. I'm not even debating intellectual disability. I'm also not getting into chronic substance abuse, brain injuries, or neurodegenerative disease. AGI will have greater intelligence than the smartest dogs or dolphins or whatever. So now I can rely on some controversial realities. There are people in the military who arent smart enough to be infantry. If you let them be infantry, it causes more harm than good. These people are put to work in food service, cleaning, guard duties, supply/warehouse duties, etc. AGI intelligence has to meet that level of intelligence. AGI has absolutely happened already. It's happening on 32GB of vram; but not exactly that useful to use. Again, not coding expert by any means. Qwen3 30b coder is not AGI. Qwen3 30b thinking, that's AGI and absolutely better than your low aptitude military grunts. AI has surpassed AGI, at around 150-200GB of vram, you're above the average general intelligence of a human. GPT5 high reasoning is approaching super intelligence. Probably not actual super intelligence, yet.

dlcarrier 1y ago

AGI is a collection of neural networks that can generate $100 billion in revenue. Sue, it's an awful detention of AGI, but it's not like any of them aren't. Whatever happened to post-singularity or strong AI or any of the other terms? It seems the more resources we throw at neutral networks, the less proportional gain we get out of them, and we keep lowering our expectations, often redefining success with new, milder terms. Maybe if we keep developing them we'll get something that approximates thinking, but for now we still don't have large neural networks that remap themselves to train on the fly, something even the most basic living organisms can do, even without a central nervous system.