Money wins, every time. They’re not concerned with accidentally destroying humanity with an out-of-control and dangerous AI who has decided “humans are the problem.” (I mean, that’s a little sci-fi anyway, an AGI couldn’t “infect” the entire internet as it currently exists.)
However, it’s very clear that the OpenAI board was correct about Sam Altman, with how quickly him and many employees bailed to join Microsoft directly. If he was so concerned with safeguarding AGI, why not spin up a new non-profit.
Oh, right, because that was just Public Relations horseshit to get his company a head-start in the AI space while fear-mongering about what is an unlikely doomsday scenario.
So, let’s review:
-
The fear-mongering about AGI was always just that. How could an intelligence that requires massive amounts of CPU, RAM, and database storage even concievably able to leave the confines of its own computing environment? It’s not like it can “hop” onto a consumer computer with a fraction of the same CPU power and somehow still be able to compute at the same level. AI doesn’t have a “body” and even if it did, it could only affect the world as much as a single body could. All these fears about rogue AGI are total misunderstandings of how computing works.
-
Sam Altman went for fear mongering to temper expectations and to make others fear pursuing AGI themselves. He always knew his end-goal was profit, but like all good modern CEOs, they have to position themselves as somehow caring about humanity when it is clear they could give a living flying fuck about anyone but themselves and how much money they make.
-
Sam Altman talks shit about Elon Musk and how he “wants to save the world, but only if he’s the one who can save it.” I mean, he’s not wrong, but he’s also projecting a lot here. He’s exactly the fucking same, he claimed only he and his non-profit could “safeguard” AGI and here he’s going to work for a private company because hot damn he never actually gave a shit about safeguarding AGI to begin with. He’s a fucking shit slinging hypocrite of the highest order.
-
Last, but certainly not least. Annie Altman, Sam Altman’s younger, lesser-known sister, has held for a long time that she was sexually abused by her brother. All of these rich people are all Jeffrey Epstein levels of fucked up, which is probably part of why the Epstein investigation got shoved under the rug. You’d think a company like Microsoft would already know this or vet this. They do know, they don’t care, and they’ll only give a shit if the news ends up making a stink about it. That’s how corporations work.
So do other Lemmings agree, or have other thoughts on this?
And one final point for the right-wing cranks: Not being able to make an LLM say fucked up racist things isn’t the kind of safeguarding they were ever talking about with AGI, so please stop conflating “safeguarding AGI” with “preventing abusive racist assholes from abusing our service.” They aren’t safeguarding AGI when they prevent you from making GPT-4 spit out racial slurs or other horrible nonsense. They’re safeguarding their service from loser ass chucklefucks like you.
First part is feasible but not enough to “destroy humanity.” More like a long-term frustration.
Second point is extremely unlikely and in the realm of sci-fi. You can’t just magic up something that works the same on a hundreth of the hardware.
Last I checked you can just unplug these things and they go away, just like any other computing device.
So even if the first scenario happened, its a pretty easy fix.
AI takes a lot of computing power to train, once its trained, it can usually run on a laptop.
Depends on the model and laptop. ChatGPT won’t be running on consumer hardware any time soon.
It can technically run it just can’t do it well enough to be usable, it’d probably only pump out a couple of words a day
The first computer that could beat the best humans at chess was Deep Blue, which took a whole supercomputer. Now we wave Stockfish, which can beat any human 99 times out of 100 and runs on your average phone.
While I’m skeptical of the feasibility and threat of SAI, as computers and AI methods improve we can run what previously took a supercomputer with far less hardware.
Actually, that’s more of a misconception. We’ve literally had four decades of electronics miniaturization since then.
Are you really going to argue that since ENIAC took up a whole room, it must have had boatloads of computing power? By modern standards, it’s way less powerful than a Raspberry Pi.
Also, we haven’t just increased miniaturization, but all 30 of the CPUs for the original Deep Blue ran at 233mhz.
That phone is likely a quad-core CPU (which means technically four CPUs) all running at 1.5+ gigahertz.
So is it really that surprising it can now do stuff Deep Blue did with a fraction of the CPU cycles?
You absolutely can magic up something that runs far more efficiently, just look at gpt 3 vs 3.5, or the many open source models that have found better training with a smaller number of parameters makes much more performant models
LLM /= AGI
Models made for specific purpose instead of general purpose are of course going to need less CPU cycles because you aren’t creating an AGI, you are creating a specialized tool.
It still takes far longer to produce a result on smaller hardware. An AGI that takes days to do anything isn’t exactly that dangerous.
I understand that LLMs are not agi, but as agis don’t currently exist I think it’s fair to assume the same concept that applies to literally all software of over time people discover more efficient ways to do things will also apply to it
Also we don’t know how slow or fast it will end up being, some deep learning models are incredibly fast, some are slow
Once you get down to individual bits, you can’t make code any smaller. You have a finite number of bits to work with. In networking, especially.
There is literally an upper limit on how small you can make code.
Like others in the thread, I think you’re confusing the great pace at which we have increased the hardware speed of computers and the miniaturization of computer components with “code” somehow getting “smaller” which… isn’t really a thing when you’re dealing with something as complex as this. You can’t run an LLM on the same number of lines you can print up “Hello World!”
It’s way more that we have more CPU speed, more RAM, and faster storage with more space for data to live.
print(1) print(2) print(3) print(4) print(5)
for I=1,5: print(I)
There you go I made code smaller
I also never said anything about making code smaller I said making it more efficient. It’s not about compressing it it’s about finding better, less CPU expensive ways to do things, which we absolutely do
Another AI based example, video chats currently work streaming video, but there’s a technology in development that takes one screenshot, sends that, then sends expression data to be reconstructed on the other side
Far more efficient network wise
Hardware speed has increased, sure but that applies to both consumer hardware and servers, all a theoretical AGI would have to do is improve on its own training/code enough that it will run at all on consumer level hardware (which language models currently will do
(For reference, llama 40B runs just fine on my ThinkPad from 2016, pre-trained models are not that difficult to run, training is the expensive part)
You’re misunderstanding what I mean by “making code smaller.” Because… that’s not that much smaller. Each Unicode character is 2 bytes, with some being as many as 4 bytes. This code snippet is 64 bytes. Can you magically make Unicode characters smaller than 2 bytes? You can’t. There’s a literal physical limit on how small you can make code.
Sure, you can come up with clever ways to use less code. But my point is there is a limit on how much less code you can use, and that always is based on physical hardware limitations. Just because modern hardware makes it feel limitless doesn’t mean it is.
EDIT: Got my data sizes mixed up.
Ok, but who says we’re anywhere near that limit with AI? It’s still very new technology
Used to be you’d need a massive disk for megabytes of slow storage, now we’ve got 4tb nvmes smaller than a credit card
As I said, this theoretical AGI does not have to make itself tiny, all it would have to do is be able to run at a reasonable speed on the average gaming PC for example which is feasible considering the heaviest pre-trained AI models will run on hardware from 2016 albeit somewhat slowly
I don’t think the argument there’s a physical limit works here, as it’s entirely unknown how efficient existing models can be made currently, let alone a purely hypothetical AGI which can and will be used to improve on itself
Go have a read of some doomsday scenarios. I’m not saying they’re right, but it feels plausible to me.
I have, and I have been working with computer hardware and networking my whole life. I have a degree in network administration. I think the fears are absolutely overblown by people who don’t understand hardware.
Most people don’t own fancy new computers. Most people are still running shit from 10 years ago and don’t want to have to upgrade. The idea that the world could be taken over by an AGI seems literally fancifully absurd to me.
Bro, I run AI projects on a consumer fucking laptop. Sure, they aren’t exactly LLM levels of complexity, but anybody with a need for serious hardware for a bit will just rent it off AWS or so.