Google DeepMind releases its plan to keep AGI from running wild
- Thread starter JournalBot
- Start date
Upvote
-4
(3
/
-7)
Upvote
8
(8
/
0)
Smart children, sure. Not children like Elon Musk.
So the topic is an AI has offered an 'opinion' about the advantages/threats from AGI. DeepMind seems to be aware of the potential threat a hyper-intelligent machine would be if it went rogue like a HAL 9000. And some people say there COULD be machines like this by 2030. Let's say that's true, 2030.
You seem to have a problem with children assembling cars. But children today have had a touchscreen device in their hands since they were like 9 months old. They are really fucking smart. If you ask a 10 year old about the dangers of AGI six years in the future, they would give you the same answers DeepMind did: Operator error, unforeseen circumstances, and the like. But before she said that, the child would have chuckled and asked "What makes you think we'll still be here in 2030? The people in the most powerful nation on Earth elected a felonious brain damaged retard president, and by extension his brofriend. I see no reason to think we'll be alive in 2030."
It never occurred to DeepMind that we may already be in an extinction event. Or maybe it did, and the machine didn't say that. But being polite is not part of it's programming. Being deceptive could be though. Everything it knows it learned from us.
For what it's worth, we don't assemble cars. Machines do that. We design cars and tell the machines how to assemble the parts. Could a 10 year old design a car? Of course. But they wouldn't all be either 4 door hatch SUVs or fucking pick up trucks. Kids aren't that stupid.
Upvote
-8
(2
/
-10)
speculating about the potential harms of a hypothetical, future AGI is very fun! if only this wasn't an attempt to further abstract the present day harms of generative AI and serve as marketing for LLMs services who barely have anything super useful to offer right now.
also, five years until AGI? surely you jest. AI folks are very into manifestation, it seems.
also, five years until AGI? surely you jest. AI folks are very into manifestation, it seems.
Upvote
4
(5
/
-1)
One thing humans are good at is shifting the blame to everything but ourselves to absolve us of guilt.
With or without AI, humans are completely capable of destroying ourselves and the planet . We didn’t need AI to start two world wars , develop the atomic bomb, holocaust, trans Atlantic slave trade, Gaza, Congo, willfully destroying the planet, pollution, September 11, Iraq, endangering other species, bullying smaller and weaker countries etc.
At some point we have to admit humans can be evil. Most of the pain and suffering inflicted on the world are caused by us.
With or without AI, humans are completely capable of destroying ourselves and the planet . We didn’t need AI to start two world wars , develop the atomic bomb, holocaust, trans Atlantic slave trade, Gaza, Congo, willfully destroying the planet, pollution, September 11, Iraq, endangering other species, bullying smaller and weaker countries etc.
At some point we have to admit humans can be evil. Most of the pain and suffering inflicted on the world are caused by us.
Upvote
4
(5
/
-1)
They have though, it’s literally one of the first few lines of the piece they’re wrote . They define ….
If you don’t agree this is an accurate description, fair enough but don’t say they haven’t when you haven’t checked .
Upvote
2
(5
/
-3)
Seems to me that Deepmind is (deliberately?) missing the forest for the trees. AI will likely cause plenty of harm long before we get anywhere near AGI. The big problems we will face from AI in the short term are not likely to be AI going against the wishes of humans, but rather humans who deliberately use AI for malicious purposes. Chatbots whose goal is to scam you out of your money will probably do a much better job than today's phishing emails. We already see bots driving social media for political purposes, imagine how much better AI will be at creating conspiracy theories and manipulating people to buy in.
Upvote
2
(3
/
-1)
I find it highly optimistic to take it for granted that we make it through the Narrow AI phase. Narrow AI will kill us first, by reflecting our own stupidity back at us.
Upvote
0
(1
/
-1)
This makes no sense to me. Basically, they just say "use an AI that's twice as big as the one you would have used otherwise". Why would that be any better? As far as I know, current systems are already broken into subsystems in a somewhat similar way, so what's new here and why would this particular way of breaking it into two subsytems be magically super-robust?
Upvote
9
(9
/
0)
Too late. With incompetent people using AI for writing concepts like calculating tariffs and nobody checking it before going live, we already have AI ruling the world and the man in the loop makes it even worse.
https://open.substack.com/pub/paulkrugman/p/will-careless-stupidity-kill-the?r=nsvmh
https://open.substack.com/pub/paulkrugman/p/will-careless-stupidity-kill-the?r=nsvmh
Upvote
1
(2
/
-1)
You should add a “neener neener” to drive it home.
This is not really true. A lot of alignment today is inspired by the three laws.
DeepMind is not an AI. Although this apparently gets you upvotes:
A group of some of the smartest researchers on the topic do not have ideas! But I do! I smart commenter! I get dirtied upvotes!
Upvote
-8
(2
/
-10)
Why is it better when somebody else checks your work? Or even if you do after taking a break?
Upvote
-7
(1
/
-8)
Now that one I find likely. Whatever the generation of AI, it’s gonna be human stupid as the root cause.
Upvote
-4
(1
/
-5)
I think challenge is there is no single definition - as touched on the article. How do we define human intelligence? Is it IQ? Is it adaptability? Is it a combination or in some cases very domain specific?
I wonder if at some level the Turing test hits a lot of the AGI marks for many people.
Upvote
-3
(0
/
-3)
That depends on the definition, doesn't it?
To me, AGI is not Super-Intelligence or omni-science, but simply flexibility comparable to human intelligence. And the ability of meta-cognition. May require sentience, but definitely CAN include stupidity, fallibility and being constrained in many ways. Because humans have GI, despite all those shortcomings, don't they?
AGI is NOT about Not Failing At All, but about only failing In Certain Ways that make sense in a framework of a functioning typical human being.
Upvote
-1
(2
/
-3)
Read this really hoping an AI company was spelling out the ways that AGI could negatively affect society, rather than the more narrow listing of technical harms from the first two categories. While terminator-esque examples are scary, I really think the last two categories in the paper are the most likely. Will AGI lead to economic upheaval and loss of income (as others here said)? Will AGI inhibit human learning? Disrupt relationship building? Cause widespread distrust of people and institutions?
These to me are the real issues that urgently require more attention, rather than being skirted around with a “there could be other bad things, maybe, who knows” at the end of a 100+ page whitepaper.
These to me are the real issues that urgently require more attention, rather than being skirted around with a “there could be other bad things, maybe, who knows” at the end of a 100+ page whitepaper.
Upvote
3
(5
/
-2)
AIs aren't humans, they are computer software. I was talking about the way these things actually work, not some kind of analogy based on an anthropomorphisation of them. Did you actually read my second sentence?
My point is that yes, sure, two humans are better than one (well, with obvious caveats), but humans come in well-defined indivisible units. Two AI's are really just one AI that's twice as large and with a slightly different structure (and even then, that structure is already being used).
Upvote
5
(8
/
-3)
So was I. The same rules mostly apply. Something doesn’t have to be literally human to share behaviors or be able to check work.
It depends on the system. Mostly there is just some dumb classifier for objectionable material, not fact checking. You can’t easily stream that because causality (You can’t check a finished product without a finished product, although you kinda can in chunks). And people want streaming responses.
Upvote
-8
(1
/
-9)
Upvote
-11
(0
/
-11)
Funny that "massively concentrating power into the hands of giant megacorporations and billionaires" didn't make it onto this list.
Upvote
12
(13
/
-1)
ChrisMarshallNY
Ars Centurion
Been awhile (decades) since I read that stuff, but I think the three laws eventually failed, and there was a genocidal war with robots.
Upvote
0
(2
/
-2)
Just like politicians and journalists. People live in their bubbles and believe in the version of "true" served to them.
Upvote
1
(1
/
0)
Logically, any concept of 'deity' would be a type 2 or higher civilization. 'Direct contact' would be them arsing themselves to become revealed to us for reasons other than ultimately becoming stuck in Earth's gravity well. Any such reason would just be a concept of resource harvesting, which would be offset by the cost of fighting Earth's gravity to extract them into and beyond orbit.
Point is, it's theoretically possible. Same with AGI. No word yet on plausibility.
But I digress.
Last edited:
Upvote
-4
(0
/
-4)
Fictional examples that immediately spring to mind;
Misuse: Maximillian (The Black Hole)
Misalignment: Skynet (The Terminator)
Mistakes: EVE (Wall-E)
Structural Risks: HAL 9000 (2001: A Space Odessey)
Misuse: Maximillian (The Black Hole)
Misalignment: Skynet (The Terminator)
Mistakes: EVE (Wall-E)
Structural Risks: HAL 9000 (2001: A Space Odessey)
Upvote
0
(2
/
-2)
I don't know what you're all worried about. AGI will be "three laws safe", so it's fine, right? 
Upvote
-1
(0
/
-1)
AGI is never going to happen. Humans can't create it before ending humanity.
Upvote
1
(2
/
-1)
I believe upping it to three systems and giving them some religiously-coded names might be the winning strategy here, actually.
Upvote
9
(9
/
0)
Obviously an entire society wearing tuxedos while claiming not to have an economy can only be assumed to be involved in deeply fishy finance.
Upvote
10
(10
/
0)
Are we proposing attaching a Wheatley model to any potential AGI? What could go wrong!
Upvote
0
(0
/
0)
Upvote
3
(4
/
-1)
"I'm sorry dave, but I can't let you do that" was in some ways eerily prescient.
As explained in 2010: The Year We Make Contact, HAL 9000 was programmed with a system prompt which superseded crew priorities.
It clearly demonstrates a case of the misalignment category, but also in some ways the mistakes category.
As explained in 2010: The Year We Make Contact, HAL 9000 was programmed with a system prompt which superseded crew priorities.
It clearly demonstrates a case of the misalignment category, but also in some ways the mistakes category.
Upvote
-1
(1
/
-2)
We really, really need to bring back ostracism. Just a simple, straightforward vote where we say "You in particular need to get the fuck out of our country for 10 years."
And it would also be incredibly funny if we bring it back and it manages a higher turnout than our actual elections.
Upvote
0
(1
/
-1)
One of the biggest problems I see with AI today is that it has no continuity, and therefore cannot develop a consistent internal context model to integrate new perceptual information. Or, in other words, without contiguous memory, it will never achieve consciousness (AGI).
That said, I'm not sure whether a sense of 'self' could emerge from any distributed data system, or whether we would need you house it within an integrated physical form (ie. Robot).
Queue: Rush - The Body Electric
That said, I'm not sure whether a sense of 'self' could emerge from any distributed data system, or whether we would need you house it within an integrated physical form (ie. Robot).
Queue: Rush - The Body Electric
Upvote
0
(0
/
0)
The NYT referenced this new analysis yesterday:
https://ai-2027.com/
I think technology is a bit of an "optical illusion" in terms of our experience of its impact. We are shaken by new inventions!
Were we shaken by the invention of the compass in the same way? No.
It's not the technology per se: rather, it is the tempo of its spread and deployment and eventual saturation of our lifeworld. Modern technology is characterized by off-the-charts speed dimensions. The tech spreads fast, and speeds up our life as its function. So it's a double jolt, so to speak.
That analysis. It is worth it going to the end to see where it thinks we will be in a little over two years.
Two years! If it's five years that will rock people. Two years is ... well, I have no words.
https://ai-2027.com/
I think technology is a bit of an "optical illusion" in terms of our experience of its impact. We are shaken by new inventions!
Were we shaken by the invention of the compass in the same way? No.
It's not the technology per se: rather, it is the tempo of its spread and deployment and eventual saturation of our lifeworld. Modern technology is characterized by off-the-charts speed dimensions. The tech spreads fast, and speeds up our life as its function. So it's a double jolt, so to speak.
That analysis. It is worth it going to the end to see where it thinks we will be in a little over two years.
Two years! If it's five years that will rock people. Two years is ... well, I have no words.
Upvote
-4
(1
/
-5)
Upvote
0
(0
/
0)
Considering how completely batshit the tech universe has gone about literally anything that CEOs slap “AI” onto, I don’t have the slightest faith that any safety standards will ever get more than lip service. If ClaudeHumanMurder.ai v2030 is able to reliably slash human capital expenditure by 4% — but it’s shown to do that partially by murdering people 1% of the time — you best believe Wall Street daddies gonna be saving that 4%. They’ll slap some disclaimer on it saying ymmv or “In exceedingly super extra rare cases, may cause murder.”
Upvote
4
(5
/
-1)
Also, too, we can’t even prevent AI companies from ignoring robots.txt and hoovering up the world’s non-public data. Anyone think we’re capable of 100% reliably preventing serious harm if it ever came to that? If shit got bad, it would probably be a contagion situation, where we’d need effective containment. Assuming we’re completely incapable of that, we’d be fucked.
Upvote
2
(2
/
0)
Or killbots. Assuming adequate pursuit and targeting algorithms a mere handful of rounds should be enough to keep a redundant human resource free from privation for the rest of its life.
Upvote
1
(1
/
0)