AGI achieved internally
>>109762977Does general intelligence require math?
Alright yeah this convinced me, GIVE SAM HIS TRILLY!
>>109763094yes
>109762977Has anybody kept count of how many times we've achieved AGI already?
>>109762977>blue line>darker blue line
so why not release the internal model? it would make them a shitload of money if it's that good. unless.... that goddamn Altman.
>>109763917wait until they gimp astra to the sol level first
>>109763112The definition requirements of AGI just keeps increasing every time a model crosses it. Is solving navier-stokes finally enough?
>>109763101This, fuck Cuckio Wokemodei.
>>109763917They have to apply a thick layer of censorshit, otherwise it will tell some very inconvenient truths.
>>109763094Yes, but not just.It also requires being able to hold onto multiple levels of plan (selective pruning of context) and the ability to solve practical problems, like how to boil an egg. Not to just write the instructions, but to carry them out.
>>109762977I've begun to hat teh antichrist
>>109762977>curated set>open math problemsIn other words, they cherry-picked the problems, and these are well known problems that countless mathematicians have written about already and possibly already solved without realizing it.
>>109762977AGI: the new graphene
>>109762977>muh mathI love how OAI gave up with coding.
>>109763094First requirement
>>109762977fuck you openai, solve world hunger, cure cancer, make every human being live comfortably and in good health, do something other than make sam altman richer
>>109764780to me it seems like everyone kind of gave up on coding. anthropic is now talking about biology or something. maybe im just coping but the current models dont seem to be significantly better at long-term software engineering. they are better but the difference is not that big. there doesn't seem to be anything about astra that now feasible to just have it develop complicated software meant for long-term use on its own unsupervised.
>>109763094I'm not even memeing, Anon. There's math and numbers in literally everything.
>>109764780ASStra is a data ANALysis god!
Some results point strongly toward AGI, particularly where generated output exceeds mean human capability. At the same time, the gaps elsewhere suggest those capabilities may not constitute genuinely general intelligence. Still, human intelligence is itself uneven, so we may be mistaking an unfamiliar configuration of generality for its absence. But that risks redefining generality around whatever the system happens to do, so I'm really divided on this issue.
>>109764870>solve world hungerthe world does not need more niggers we are full
>>109764935350k lines of .lean files seems kind of slopped though. Sure, exceeds human capacity but I can't help but to think that it's also still underperforming humans at least on the slop metric.
>>109763094If you can't see that the very concept of "general intelligence" is a subversion then you get what you deserve.
>>109764901>everyone kind of gave up on codingbecause it's solved and people would take coding improvements for grantedneed grow wider to invite more money
>>109765085it's not though. you still need people to supervise the models and desloppify things for sufficiently complex software.
>>109764121> The definition requirements of AGIWhat was the original definition? Pass our synthetic benchmark XY or impress me with Math?
>>109764780>none of the models can follow instructions instructions even 80% of the timegrim>>109765085it's not solved - if it were solved then software would start getting better instead of getting worse at an accelerating pace
>>109763112Ask AGI to do it
>>109762977>Irl omega team
>>109763101just one? make it 20
>>109762977Ah, good old "curated math problems" benchmark.
>>109764935Everyone keeps saying it's AGI and it keeps slopping up my C++ project when I ask it to do the simplest things.
>>109763887AGI made this
>>109766011Every time I've tried using a free model for C++ questions, they can't do it. They simply cannot understand template metaprogramming and are too unreliable with the standard library (they like to make up functions that don't exist and pass the wrong parameters to functions that do exist).
>>109765144>>109765215It's good enough for what management want, and thats all they need.
>>109763887De dark line do be higher doe
>>109763917They need to make sure it doesnt say nigger first
>>109764935I don't think they can be truly general because of their sample efficiency. You can't put absolutely everything there is to know about the world in the weights. If you can't adapt to a completely novel unseen complex task without a mountain of training data you aren't generally intelligent.That doesn't mean they can't be useful.
