This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
Sooo, Big Yud appeared on Lex Fridman for 3 hours, a few scattered thoughts:
Jesus Christ his mannerisms are weird. His face scrunches up and he shows all his teeth whenever he seems to be thinking especially hard about anything, I didn't remember him being this way in the public talks he gave a decade ago, so this must either only be happening in conversations, or something changed. He wasn't like this on the bankless podcast he did a while ago. It also became clear to me that Eliezer cannot become the public face of AI safety, his entire image, from the fedora, to the cheap shirt, facial expressions and flabby small arms oozes "I'm a crank" energy, even if I mostly agree with his arguments.
Eliezer also appears to very sincerely believe that we're all completely screwed beyond any chance of repair and all of humanity will die within 5 or 10 years. GPT4 was a much bigger jump in performance from GPT3 than he expected, and in fact he thought that the GPT series would saturate to a level lower than GPT4's current performance, so he doesn't trust his own model of how Deep Learning capabilities will evolve. He sees GPT4 as the beginning of the final stretch: AGI and SAI are in sight and will be achieved soon... followed by everyone dying. (in an incredible twist of fate, him being right would make Kurzweil's 2029 prediction for AGI almost bang on)
He gets emotional about what to tell the children, about physicists wasting their lives working on string theory, and I can see real desperation in his voice when he talks about what he thinks is really needed to get out of this (global cooperation about banning all GPU farms and large LLM training runs indefinitely, on the level of even stricter nuclear treaties). Whatever you might say about him, he's either fully sincere about everything or has acting ability that stretches the imagination.
Lex is also a fucking moron throughout the whole conversation, he can barely even interact with Yud's thought experiments of imagining yourself being someone trapped in a box, trying to exert control over the world outside yourself, and he brings up essentially worthless viewpoints throughout the whole discussion. You can see Eliezer trying to diplomatically offer suggested discussion routes, but Lex just doesn't know enough about the topic to provide any intelligent pushback or guide the audience through the actual AI safety arguments.
Eliezer also makes an interesting observation/prediction about when we'll finally decide that AIs are real people worthy of moral considerations: that point is when we'll be able to pair midjourney-like photorealistic video generation of attractive young women with chatGPT-like outputs and voice synthesis. At that point he predicts that millions of men will insist that their waifus are actual real people. I'm inclined to believe him, and I think we're only about a year or at most two away from this actually being a reality. So: AGI in 12 months. Hang on to your chairs people, the rocket engines of humanity are starting up, and the destination is unknown.
I would enjoy engaging more with the AGI x-risk doomer viewpoint. I fully agree AI narrow risks are real, and AI sentience/morality issues are important. Where my skepticism lies is when presented with this argument:
Human intelligence is apparently bounded by our biology
Machine intelligence runs on machines, which is not bounded by biology!
Therefore, it may rapidly surpass our intelligence
Machine intelligence may even be able to improve its own intelligence, at an exponential rate, and develop Godlike power relative to us
This would be really bad if the MI was not aligned with humanity
We can't (yet) prove it's aligned with humanity!
Panicdoom
Where I have trouble is #2-4.
One variant of this Godlike argument I've seen (sorry if this comes across as a strawman, gaining traction on this debate is part of why I'm even asking) is that humans just becoming a little bit smarter than monkeys let us split atoms and land on the moon. Something much smarter than us might continue to discover fundamental laws about reality and they would similarly be Gods compared to us.
The reason I don't buy it is because we've been able to augment our intelligence with computers for some time now: by moving our thinking into computers we can hold more stuff in our head, evaluate enormous computations, have immediate recall, and go real fast. Sadly, the number of new game-changing fundamental laws of nature that have popped out of this have been approximately zero.
I believe we've discovered all of the fundamental laws of nature low-hanging fruit and the higher hanging fruit just isn't so computationally reducible: to learn more about reality we'll have to simulate it, and this is going to require the marshaling of an enormous degree of computation resources. I'm thinking less on the scale of entire data-centers in The Dalles full of GPUs and more like something the size of the moon made of FPGAs.
Stated another way, what I think holds humanity back from doing more amazing stuff isn't that we've failed to think hard and deep and uncover more fundamental truths and we could do that if we were smarter. What holds us back are coordination issues and simply the big hill to climb to boot up being able to harness more and bigger sources of energy and mine progressively stronger and rarer materials.
An AGI that wanted to do game-changing stuff to us would need to climb similar hills, which is a risk but that's not really a Godlike adversary -- we'd probably notice moon-sized FPGAs being built.
I recognize an AGI that was fast and coordinated and numerous could be a dangerous adversary too, but I'd like to only focus on why we think a massive increase in IQ (or rather, g) is the big x-risk.
Well what about 95% of the energy of the universe being unknown to us? We call it 'dark' as though that's some kind of explanation. Something is out there and it's far more important than everything we can see. Back in the late 19th century they thought they'd discovered all the laws of nature too. Newton got the job done, there were only a few weird puzzles about blackbody radiation and the orbit of Mercury being a bit odd. They got relativity, quantum physics, radio and so on. Our 'weird puzzle' is 95% of the universe being invisible! Either there's an immense amount of aliens or there's an extremely big secret we're missing.
Anyway, we haven't even left the tutorial stage of applying the physics we already know. No fusion, no 3D nanoscale engineering or nanorobotics. No quantum computing worth caring about. These are mostly engineering challenges that need optimization, AI can do that. It's already doing that. It's optimizing our chip layouts, it's used in controlling the plasma in fusion, it is necessary for understanding protein folding. AI is giving us the optimizing power to keep advancing in all these fields. These fields are immensely powerful! Mastering nanoscale robotics and fusion means you can start scaling your industrial base very quickly.
But the implementation is incredibly powerful. Look what the US did to Iraq - that's precision-guided weapons (computers), advanced fire-control (computers), night vision (computers), anti-radiation missiles (computers). Everything is using computer aided design, computers model how the armour holds up, computers let you command and control these powerful forces and bring firepower where it's needed, computers do the ballistics... The US didn't know some fundamental principle unknown to Iraq, it was only their implementation that was better.
If the AI has better implementation than us, we're fucked. It can use deceptive tactics to turn us against eachother, interfere with our command and control, snipe leaders with drones, weaponized mosquito-bots, smart mortars that pop out of vans. It can compromise people with blackmail, spy on our forces via satellite, bribe people. With fusion and nanotech, it can brute-force us directly, drowning us in robots.
Is your intuition that we're just totally missing a basic fundamental truth of the universe that fits on a cocktail napkin and if only we weren't such pathetic meat sacks we'd figure it out?
Because to me this screams "computationally irreducible" and we're not going to get traction without big simulations.
I believe I agree with you here? Is there a delta from your POV with my last paragraph? (Repeated below)
More options
Context Copy link
More options
Context Copy link
Your core misunderstanding is assuming that the AI will have to discover new fundamental laws of nature in order to reliably kill everyone. The AI doesn't need to reconcile quantum mechanics with general relativity to make super-smallpox. It just needs really good computational biochemistry skills.
I think the most common failure mode for people who don't get how AI could physically kill everyone is that they don't realize all the crazy shit that proteins can do.
Well, not just really good computational biochemistry skills? Wouldn't it also need a revolution in synbio to access an API where it input molecules and they were then produced? Where would that get sent? How do you convince people to inhale it?
Aside: I expect this synbio revolution would usher in an era of corresponding print-at-home immunity, reducing the threat vector from bespoke bioweapons. I don't expect all x-risk from weapons defense to be this symmetrical, shooting down an ICBM is much much harder than launching one for example. I would like to be as concrete as possible about the risks though.
Right here. You literally just type in the DNA sequence. Of course, that's assuming there are no biotech labs that have hooked up the AI directly to the protein synthesis machines to streamline development.
"Introducing Project Telos, a groundbreaking first-line clinical trial harnessing the power of a retrovirus to deliver telomere-repair enzymes into somatic cells. Join us in revolutionizing healthcare and longevity by combating age-related diseases, enhancing tissue repair, and promoting overall well-being. As a participant, you'll play a crucial role in shaping the future of regenerative medicine, while gaining exclusive access to potential health benefits. Be a part of history and help unlock the potential for a healthier, longer life."
And yes, GPT-4 wrote that.
Welcome to the future.
Let me see if I understand the threat model.
Unaligned AGI decides humans are a problem, engineers virus more infectious than measles with very long asymptomatic incubation period and 100% lethality.
Virus is submitted to idtdna.com with stolen payment info that the AGI hacked from elsewhere.
Idtdna.com processes the request, builds the supervirus, and ships it somewhere.
????
Everyone dies.
I assume you'll have a clever solution for 4.
Why do you assume the lab would synthesize any arbitrary protein? Surely they would want some confidence they're not building a dangerous supervirus?
Or are we assuming the evil AGI can submit a convincing doc with it that says it's totally benign?
It might be difficult to find an individual lab to synthesize the whole virus from scratch. You could do it by coming up with a 2-component capsid which spontaneously assembles in aqueous solution at a particular pH, getting two individual shipments of the two capsid proteins, getting a third shipment containing the viral genome, then mixing them together and adjusting the pH.* You could bribe, brainwash, or trick an undergrad to do this part.
Sequences are blacklisted, not whitelisted. The whole point of research is that you don't know exactly what the proteins are going to do yet. It would not be hard for the AI to come up with sequences that look benign at first glance. If need-be, yes, a convincing doc could be submitted. GPT-4 could probably do that part today.
I admit I'm a bit confused here. Do you need an explanation for how we get from "supervirus in a tube" to "everyone dies"? Imagine Covid but with a 3-month incubation period and a 100% fatality rate. Convince an undergrad to pour it over his face or something. Tell him its the Fountain of Youth.
More options
Context Copy link
More options
Context Copy link
"Super-smallpox" is a metaphor. There's no reason an engineered pathogen needs to bear an apparent similarity to any known pathogen.
An AI that's solved protein-folding can make its own custom restriction enzymes and ligases with different nucleotide substrates than are currently known (good luck predicting the substrate from the amino acid sequence without the AI's help). The final DNA sequence need not be obvious from the fragments sent to the lab for first synthesis.
Maybe the virus looks like it isn't replication-competent? Maybe they even run tests on an immortal cell-line in a petri dish. Perfectly safe right? Uh-oh, turns out it was specifically designed to not replicate in the exact cell-line used for testing. Whoops.
This is what Eliezer means by "security mindset". The above safeguards are the equivalent of computer security via increasingly large passwords. It is predictable that they will fail, even if you can't see the exact failure scenario.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Are prions a legitimate thing to worry about given that Brits ate prion tainted beef for like a decade and nothing happened?
Well, a hundred people died or so, all of whom shared a certain rare mutation.
I guess there's also Kuru, but are we certain that tribe wasn't inbred, with weird genetic susceptibility to it ?
Of course, that doesn't rule out some funny AI reading through themotte.org for inspiration and then figuring out especially good prions against whom we aren't mostly immune, I guess. Well, hello to you if you're reading this.
More options
Context Copy link
More options
Context Copy link
Okay, but - take a 110 IQ person and give them a computer that runs at a petahertz. They're not gonna become Einstein. Does this imply it's hard or impossible to get to Einstein-level intelligence? Yet Einstein existed, with 99.9% of the same genetic material, and the same genetic mechanisms, as the 110 IQ person.
Yeah, we've plumbed the 'laws of nature' pretty deeply, hence fundamental physicists having trouble finding anything new. But that hasn't stopped the progress of technology, and it hasn't stopped smarter people from using technology much better than dumber people.
Yeah! But given we will, in the next few hundred years, give AGI a core role in our datacenters and technological development?
Not sure I follow?
But while we're here, I present to you Terence Tao. He has a 230 IQ, which is an unbelievably off-the-charts score and pushes the whole notion of IQ testing to absurdity, and he's clearly not even slightly Godlike?
Surely we'll have made a lot more progress on the interpretability and alignment problem by then (Context: the x-riskers, like The Yud in that Lex podcast, are arguing we need to pause AI capabilities research to spend more time on interpretability, since capabilities are drastically outpacing for their comfort)
A God creates, and is there any purer form of creation than bending Logos like Legos till you have a construct far greater than the sum of its parts? That's Mathematics in a nutshell.
On a less poetic note, Tao is still a human, his insane cognitive abilities are still the product of about 1.4 kilos of meat sipping 10 watts of power within a human cranium.
Beyond showing that the average human carries around a grossly underoptimized brain in comparison, you only need to reflect on all the relevant ways an AGI differs from Tao. For one, they scale, as the existence of GPT-4 shows us. Tao can't recursively self-improve, self replicate on short timespans, has plenty of biological drives distracting him, and is a squishy evolutionary kludge in comparison to an AGI.
If Tao somehow had his mind uploaded, then you ought to worry if he ever turned against the rest of humanity. No sign of that happening before the Singularity shows up by other means though.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Also, humans have been a little bit smarter than monkeys for a couple hundred thousand years at least, and yet we didn't go to the moon until my dad was twenty years old. It's clear that just being a little smarter than monkeys doesn't mean you're going to the moon next Tuesday, there's something more to it than that. Likewise, being a little bit smarter than humans doesn't necessarily mean you're going to be disassembling the solar system for atoms tomorrow.
That "little bit" was enough to create the exponential technology curves we take for granted today, and each generation built off the last.
Ergo, a Superhuman AGI would have access to all of our current technology, and thus be in a position to accelerate even ludicrously fast modern growth curves to its benefit.
Do you think it matters if it takes the malign AGI a day or a century to dismantle the Solar System? You won't be living much past a few days in all likelihood if it has it out for you.
Not really, at least until very recently. The life experience of the average human being was just about static for thousands of years before the industrial revolution. Humans may have become a little more intelligent over the centuries but they didn't suddenly become several hundred times more intelligent c. 1750.
Yes I think it matters if I die tomorrow or in 100 years.
I meant in the sense that each generation had access to all the knowledge and technology of the last, not that they got significantly smarter (the how declining Flynn effect on IQ not withstanding).
That's not what I mean, I expect that once a hostile AGI makes a move, it'll eliminate humans with speed and judiciousness in both scenarios, regardless of whether it takes it a day or a century to fulfill its resource demands afterwards. Both scenarios are considerably easier once you have pesky future competitors such as humans out of the way.
So from your perspective, you die at the same time, in both cases soon rather than later if you consider typical AI timeliness.
Right, but for most of human history technology and knowledge didn't really accumulate much over the generations. We probably learned and built more in the last 200 years than in the 20,000 before that, but the sudden explosion c. 1750 doesn't seem to be down to a similar explosion in "raw" human intelligence.
I'm skeptical that a human-level artificial intelligence will be in a position to wipe out mankind immediately (i.e a few days/weeks/years) after its 'awakening.'
Hmm? I don't see why you would say that tech and knowledge didn't accumulate, it did, but orally, and then picked up in volume as literature was invented.
There's definitely a pickup in progress in the 18th century onwards, but to me, that illustrates that even our barely changed intelligence as anatomically modern humans suffices for exponential growth. So I'm even more concerned by minor but significant advances in the same, a little intelligence goes a long way, and it won't be starting from scratch.
I consider human level AGI to be far less of a threat than superhuman AGI, but the latter seems to be a tiny roadbump on the road to the latter. GPT 3.5 went from being like an overly eager med student to 4 being a better doctor than I am!
That being said, even a human level AGI can exponentially self-replicate, cease control of industrial equipment, and create a super-pathogen with near 100% lethality, because humans could quite trivially do the latter if we were insane enough (GOF research suggests we are..).
But even then, I think it would lie in wait to become stronger, and do so in stealth, so once again, from our perspective, it appears out of nowhere and kills us post haste, without any warning. How long it takes to turn the solar system into spare parts is an academic exercise afterwards, we wouldn't be there to witness it!
Very very little. For most of human history, each generation improved only very slightly, if it improved at all, upon the knowledge base of the foregoing generation. Then in the 17th or 18th century or so everything changed. It was once possible for an educated man to be, more or less, an 'expert' in all fields because the pool of general knowledge was not very deep. The explosion in understanding and know-how of the scientific and industrial revolutions has rendered that impossible of course.
But this explosion happened without any similarly sudden explosion in the "raw" cognitive power of human beings (some argue that there was an increase in raw intelligence around this time, but even if so it clearly wasn't a several-hundred fold increase). The slight step up from monke to anatomically modern human wasn't enough on its own to take us to the moon or even create steam power, because we had to wait millennia for the proper conditions (whatever those were) in which such inventions could be realized.
Which is why I don't think "AI becomes a little smarter than us" is immediately followed by "AI becomes 50,000+ times smarter than us and then begins turning the universe into grey goo." Humans becoming a little smarter than monkeys wasn't followed by spaceflight, or even the agricultural revolution, for a long, long time.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link