>>106929490
whomst?

>>106929490
Buy an ad

AGI? We want Commer AI!

>>106929490
Can I trust Karpathy? Is he worth listening to these days?

>>106929628
Take a guess: Karpathy tunnel syndrome.

yes it's bearish that he's predicting that in 10 years the ai companies that exist now will replace literally all knowledge work and that progress towards this will come in the form of constant improvements

It can't be predicted. Either you have it or you don't know how to make it yet. Nobody can say if or when someone else will figure it out.

>>106930084
10 years and we will have cold fusion and infinite energy bruh

>>106929490
What do these dimwits even think AGI is? Chatgpt trained on even more stolen content so it can do ERP as more characters more accurately? AGI will never happen unless there is an entirely new way of making ai that can LEARN like a living being can. That is to say, without needing to memorize 7 petabytes of information to regurgitate half working code

>>106930263
i half listened to this yesterday and i came away with (and yes some of this is contradictory):
1. he doesn't like how existing models are fed mountains of data that they can memorise and regurgitate
2. he thinks the memorisation is imperfect likening it to a book you read months ago
3. he thinks we'll need more data
4. he thinks the agi models will be much smaller than existing models
5. he hates RL - he wants more than a simple true/false reward function. there are some recent papers he thinks are promising
6. he thinks we need ^ breakthrough and 5 or 6 more like it
7. he thinks using llm judges will just produce adverserial networks.
8. he thinks the existing models produce slop results esp when presented with novel scenarios

>>106929490
adssad

>>106929490
Why does this guy sound so consistently retarded and naive, if he's supposedly some AI genius? I get that overselling AI is in his interest and just a case of lying, but his takes on politics and other things are midwit tier too. Is he just another Lex Fridman, but with better opsec?

>>106930322
Didn't ask. I don't care what a spastic thinks.

>>106930322
Who came up with the "decade away". I really doubt all that will be so quick

>>106930365
karpathy. why? intuition from working in the field.
you see a lot of that these days - guys confidently predicting that we'll see x number of breakthroughs not knowing what those are and assuming they'll just keep coming.

i haven't listened to too much dwarkesh but i think he's wishywashy on his prediction but i think it's in 10-30 year range or smth. he's just parroting things that click with him though.

>>106929490
>still a decade away
Delusionally optimistic

>>106930322
>he doesn't like how existing models are fed mountains of data that they can memorise and regurgitate
Hadn't really thought about it, but yea, the current models use the most Indian way possible to 'learn', and they'll never be any more capable than the average jeet worker.

Until they have the capacity to look at a collection of statements and start understanding the commonalities, intent, patterns, and logic behind them, it's just a parrot squawking back the things it's heard before. Western whites have nothing to fear.

>>106929490
I've been saying same shit for a while, after working with every AI tool there is. The veneer of intelligence falls apart very quickly once you task it with a problem not in its explicit training set domain. A generally intelligent human could figure it out, the best models out there, fall apart completely and produce gibberish.

There is no intelligence in the models, its just a mirror of human knowledge applied with statistical probabilities.

File: 30 year old commer.png (98 KB, 780x537)

98 KB PNG

>>106929615
Did someone say commer?

>>106930426
>I've been saying same shit for a while
lol, why even bother? It was obvious af right from the start. Cunts like Karpathy know this, but keep peddling the slop for normies, so the eventual rugpull can be more profitable.

A few days ago I tried using ChatGPT (latest version on the website) to write a really simple C++ class. Very simple. It showed some code that had comments that described it doing what I wanted, but the actual code was not valid C++. I pointed out exactly what was wrong and it wrote new code, that had even more comments, like

// 100% verified to compile and run

and it also wasn't valid C++ and of course wouldn't compile.
I gave it a few more chances and each time it would produce more elaborate code that was more invalid than the times before.
It's obviously just a plausible bullshit generator. That's what we've been saying for years, and evidently it's not something they can fix or improve on because that's fundamentally how they work.

>>106930359
brown person
>>106930322
thanks for sharing

File: 1639277814193.jpg (7 KB, 245x206)

7 KB JPG

>>106932966
>brown person

>>106930420
You're just now developing that insight and felt like it was novel enough to be worth sharing? lmao.

>>106930322
>More data
They already data mined the entire internet.
Which alien race are we going to data mine a second internet from?

>>106930322
>adverserial networks.
>computer boards getting all pissy due to running code

>>106934054
Synthetic data. Augmented data. It can help a network to learn the existing data better, but won't necessarily make it smarter, if that makes sense.

Just follow the blog of Gary Marcus for some realistic reasoning about AI
Simon Willison is not bad either

muh agi

>>106929490
His definition of AGI is "can automatize office jobs with some human supervision", not human level at everything. Once you keep that in mind it makes more sense

>>106930084
He's predicting what they have already declared.
AGI will be a trillion in revenue not god in a box.

What does this mean