>computer vision now understands ~7.5M unique concepts, up from about a dozen from the mid 2000s
>even understands humor to some extent
>can also segment images perfectly
>makes it 100x easier to train even better models as it improves the ease of making training/testing data
Is it over for us meatbags? we don't even need to look at images anymore, computers can do it far better and faster
https://github.com/mbzuai-oryx/groundingLMM
>person about to hit sports ball
nobody's perfect
Accelerate
It's still missing the I part of AI and I don't understand how these models can ever truly lead to actual intelligence.
A 3 year old can do all of those things and he doesn't need to be trained on the entirety of the internet to do it.
>actual intelligence
How do you define "actual intelligence"?
>A 3 year old can do all of those things
Can a 3 year old fix bugs in your code, or write a function to solve a task?
>doesn't need to be trained on the entirety of the internet to do it.
Yes, this means that a more sample-efficient training method is possible, but it's probably compute-constrained. If we had 100x the compute of GPT4 available (comparable to the complexity of a brain), we could try finding such a sample-efficient training method, but right now it doesn't make sense to invest resources looking for a method like that.
>Can a 3 year old fix bugs in your code, or write a function to solve a task?
Sure, talking to a rubber duck can do the same.
I just don't think anyone is retarded enough to drag their kid into the work place and I'm not gonna bring my own
If you think that the output of a rubber duck is the same as the output of a language model, then you could probably be replaced by a rubber duck.
Does it understand foxgirls?
>improves the ease of making training/testing data
no it doesn't
the quality is still shit
it's ok to be afraid, the world needs midwits too
you can literally farm a training set off youtube thumbnails just by running them through this for zero effort.
good question, it will know "girl" and probably "fox ear" or at least "ear" plus "fur" and since it can segment it can localize them and you can map "ear+fur/fox" + "girl" + vicinity = foxgirl
the model you trained will be shit and not deployable in production
>samples from a dataset
That doesn't mean the AI actually knew how to label things correctly, right? Those are just training examples.
someone's got to hook those CV models into a LLM and make it play gta5, frame by frame
even if it can tell a cop from a hooker from a regular person, and even a rifle from a pistol, it doesn't know what to do though
you could give it some basic rules but the input for that's probably pretty hard compared to say, a 2d game or an RTS/MOBA
>Understands
Nope. It can classify 7.5m but it doesn't understand anything.
Classification Is Understanding.
Lol. Lmao, even.
You're Afraid. It's OK to be Afraid, Anon.
>人
I clasify this symbol as "bill". When ever you show me this symbol I will correctly tell you that it's bill. You can ask me to show you "bill" and I will draw it correctly 100% of the time. I have no idea what bill is, or what it represents.
Classification is not understanding.
There is no such thing as true or false, only Majorities and pluralities exist. We choose to say that a rock is a rock because most people say a rock is a rock, the consensus is there.
>Reality is a social construct
>Its all a power play, bro
Holy fuck, an actual post modernist.
>Inb4 anon insists that he's not.
Rabbit or duck?
>2+2=4
True or false?
do either of you want to offer a definition of "understanding", to make the question meaningful, or are you just going to trade insults?
see
99%+ of the time I don't have to look at the image myself to know if a thing is present in them, and i even don't have to look myself to see where it is
that is understanding to me
if you get it right, you understand it
if you think a hat is a rock you dont' understand it
Functionalism is the only reasonable belief.
>still arguing about if LLMs are intelligent or not when they outclass 80% of people 80% of the time
a pointless distinction, they are functionally the equivalent of a legal "Reasonable Person" except with severe Korsakoff syndrome and a penchant for insane californian identity politics.
I'm not as fast, but probably still better at it.
I would go as far as...I can actually tell gorillas and African-Americans tell apart.
YOU JUST KNOW
>midwit wants to debate
loool