This is civilization ending stuff that's mostly flying under the radar while they have us at libs' necks and vice versa

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477051)

won't someone please make it all stop

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477054)

this is the faulty premise that is causing you to unreasonably panic. it doesn't have a "model" at all in the sense of a human explanatory model. it just turns out that you can get incredibly, shockingly, breathtakingly far in creating the appearance of having such a model by doing predict-the-next-most-likely-word math hijinks. nifty and interesting and a tool with useful applications, but just a tool

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477820)

it is, at its root, matrix multiplication where the dimensions of the matrix are huge (hence why it's run on GPUs, which are the tits at matrix multiplication) and the result is predicting the next word in a sentence based on the foregoing words in the sentence plus the prompt plus the corpus of all text it has read

it doesn't reason and it doesn't understand; hell, it doesn't even apply grammar rules

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47479499)

more importantly nobody understands how the human mind actually works, which makes it all the more silly when people claim ai "is just a tool". what exactly makes us more than tools?

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47480780)

please never speak on this subject ever again

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47479498)

A main one being "human redundancy" and this was considered to be unpalatable to the consumer market in the late 70s or early 80s or whenever. Cost-wise I think they found it was feasible for denser markets albeit with a slimmer menu

These days that same "human redundancy" is viewed strictly as a benefit due to WAAC powergaming bankers and inhuman SV technomancers setting the standards for this globalized cannibalistic race to the bottom we call American capitalism

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477113)

I guess they're at it again

It's not difficult to imagine and I'm not some restaurant mgmt mfe either. Minimize the points of contact for the customer and you don't have to worry about vandalism or whatever third world problems we have in polyglot New Babylon now. The machinery has been around for decades and afaik was rolled out in Asia before it was (considered) here

The only thing that was missing was mgmt will to ram it through PR-wise

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477133)

Every 100th customer gets a roman candle fired into their windshield

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477135)

AI could literally be mass exterminating humans with swarms of nanobot robots and normies would still be arguing with each other about the latest netflix episode of random goyslop or heckin sportsball game result or petty office gossip

they do not care about anything but themselves and their own lives. in fact they are not even physiologically able to care about anything else

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477094)

“It will create new jobs! Dur durr!”

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477102)

And the New York Pravda will hail this as a great achievement because Biden created 16000 jobs this quarter (by liquidating 320000 lower salary ones)

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477144)

I am absolutely terrified. But I will fight on, like John Henry.

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47477848)

the actual answer is that normies are stupid and self-absorbed as fuck

see my poast above

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47478134)

dude, talk to a teacher at the high school level; they'll tell you their kids don't even know how to use a fucking keyboard anymore, and you want them to buy a specialty graphics card and install it in a PC? and that's "accessible" to you?

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47479514)

partially because there's a large chunk of humans that just take for granted that true AGI is impossible. the most honest of these people is the 90 IQ religious normie who will tell you that humans have souls and machines don't, so obviously there's no real threat from AI becoming smart. the smarter and more subtle version of this is people who in principle accept metaphysical naturalism, reductionism, etc but still insist there's some magic spark inside human brains that will take centuries for science to unravel, if ever. i can't prove they are wrong but they are very obviously engaging in motivated reasoning with this shit.

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47478200)

There's actually huge reasons for believing this complex reasoning is so far off it's never gonna happen. xo is just retarded like all the foomers

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47479496)

1. Uses the “it’s just a next token predictor” argument. Ignores what optimal token prediction looks like at the limit. Also says they are just matrix multiplications with non-linearities. Like this says anything meaningful about whether they can emulate biological neural computation. You could be a reductionist idiot with brains too if you didn’t know what they did.

2. Never sets fixed goal posts for actual cognition or reasoning. Look how quickly people who had all sorts of handy examples of GPT-3 failures as proof of this ignored the fixes with GPT-4 and started complaining about something else.

3. Highlights LLM failures while ignoring their flexibility in comparison to most humans. I would trust GPT-4 more to get an SAT question right (not from its training corpus) than a typical human. People are retarded in all sorts of ways LLMs aren’t. Maybe they can’t reason either (I actually like this argument in a way - most people seem like limited pattern matchers who can only make basic inferences from their limited world model. But I think deep sort of reasoning can’t be far away considering intelligent people have the same basic biological substrate with a few genetic differences)

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47481128)

“ Other skeptics are critical of the technology itself, with AI analyst Gary Marcus saying that hallucinations (where AI systems such as ChatGPT make up or mangle facts) are not a simple problem to get rid of.

Marcus believes this intractable problem could stop AI from delivering the financial returns its advocates expect.

Marcus told the Financial Times, ‘There is a fantasy that if you add more data, it will work. But you cannot succeed in crushing the problem with data.’

Marcus pointed to previously highly hyped AI technologies, such as Facebook’s AI assistant M or IBM’s Watson, which promised world-changing results but were under-delivering.

Marcus wrote on his Substack, ‘Hallucinations are in their silicon blood, a byproduct of the way they compress their inputs, losing track of factual relations in the process. I first pointed out this risk in 2001, in the fifth chapter of my book The Algebraic Mind, and the problem has persisted ever since. To blithely assume that the problem will soon disappear is to ignore 20 years of history.”

Marcus says that he believes the problem of AI hallucinations will eventually be ironed out, but no one knows whether the technological breakthroughs to solve the problem will arrive in months, years, or decades.”

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47482187)

Hallucinations are largely the result of having imperfect recall of the data and token prediction discouraging the model from expressing uncertainty. I think retrieval based systems that can directly access the data during inference will greatly increase model accuracy while increasing capacity for algorithm learning. Better or more RLHF or something like it so the model can learn to express uncertainty about its knowledge. It's hard to know for sure what will work at this point but the hallucinations have already declined significantly from base GPT-3.

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47482538)

https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47496444)

The other point I would make is that the presence of earlier LLM output in the training corpus can probably be undone by RLHF or training harder on higher quality data. LLMs are trained on all sorts of shit human text and learn to emulate it as well, but the final RLHF tuning steers them away from it. I don’t think this will prevent LLMs (or diffusion models) from improving.

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47498537)

https://www.poerkan.com/product/african-superman-sex-pill/

(http://www.autoadmit.com/thread.php?thread_id=5501700&forum_id=2#47481973)