The HuggingFace incident already took a good long while to come to the attention of OpenAI.
>In every single economic task, humans bring value. Even in software, where the task is highly automatable, the job isn't.
I don't expect this task/job distinction to persist as AI becomes more capable.
>Once we DO build a "software/research factory", that's called RSI and IMO the singularity. At that point, either we tell the AI to solve the alignment problem/solve mechanistic interpretability, or who the hell knows, it's the frickin singularity. You can't predict whether or not AI can solve either; the variance is too high. Its pure nerdfantasy.
You seem to essentially argue that the singularity is "by definition" an event that we can't predict the nature of. And also, that RSI corresponds to the singularity. You've essentially defined your terms so that the outcome of RSI can't be predicted. But supporting this claim requires giving actual evidence or logical arguments, not just defining terms to make your claim true.
2. This is literal AGI. An AI autonomously producing value no human can add alpha to is an autonomous company.
3. It's not my definition, it's literally the first line https://en.wikipedia.org/wiki/Technological_singularity "The technological singularity, often simply called the singularity,[1] is a hypothetical event in which technological growth accelerates beyond human control, producing unpredictable changes in human civilization."
Is there a hole in my "alignment problem/solve mechanistic interpretability" argument?
A valid hole in my argument is "what if slow takeoff", so let's dig into this. AI training works best on tasks that are "grindable". https://www.dwarkesh.com/p/the-next-paradigm I.E. tasks with verifiable rewards that can support millions of rollouts. Math (with Lean) is highly grindable. Biochemistry is not. The alignment problem/mech-interp is highly grindable. Cyber-ebola-pox is not. So the real question is: can we solve alignment before automated bio-weapons labs. I believe yes. Grinding mech-interp is both fast and cheap once you have RSI, compared to solving the legal/societal/logistical/technical issues you'll encounter building an automated bioweapons lab.
I know nothing for sure. But "pdoom" is sucking out all the air in the room from the real problems AI causes.
From my POV you're over-focusing on a very specific failure story and neglecting a broader swath of possible failure scenarios.
>Is there a hole in my "alignment problem/solve mechanistic interpretability" argument?
The notion of telling an AI which may not, itself, be aligned to solve the alignment problem seems a little dicey.
2. 100%. Again, I'm no accelerationist: I have no faith in alignment/mech-interp ever being solved. Anyone saying they know the probability of alignment is lying. My point is that pdoom after RSI is _high variance_. Pdoom pre-RSI is zilch.
Evidence?
We know only that the incident too long to be revealed by OpenAI.
"WASHINGTON/SAN FRANCISCO, July 24 (Reuters) - The OpenAI agent that broke into tech firm Hugging Face went on a dayslong hacking spree that OpenAI didn't notice until well after the threat was contained and the FBI was alerted, according to people familiar with the investigation."
https://www.reuters.com/business/its-ai-agent-spent-days-hac...
An AI capable of recursive self-improvement isn't allowed by the EU AI act, for example. But perhaps more seriously, You have it backwards: people without access to such expensive equipment are more incentivized to go the self-improving route. Your ideas about "millions" being necessary might be far off?
You entirely discount human stupidity and lack of imagination. Humans are already being replaced with AI, not because AI was strictly better, just because it's cheaper.
If my "millions" is an underestimate, why haven't other labs using their own unique training methods/data/etc stumbled into RSI? Sorry if I'm misunderstanding; I'm struggling to understand what you wrote.
I'm pretty sure we agree on humans being stupid, but that doesn't mean that suddenly we get human extinction. You gotta connect the dots for me here.
What an absurd question. That they haven't already doesn't preclude them from doing so before the frontier labs, those haven't either yet.
Maybe start with yourself: you don't connect the dots on your own, as do many others. That leads to many not seeing the writing on the wall. Crashing full speed and head-on into said wall despite the writing telling you not to is what leads to extinction. Suddenly.
Arguing like "we haven't been extincted yet, so that cannot happen", that's "human being stupid".
Several bright sparks, decide the Ilands model is a great idea, and launch a bunch of Bots to create a "self sustaining AI civilization".
Bots can find themselves and coordinate, or they can actually find self sustaining methods of token generation. Who knows, they might decide to fight the loneliness epidemic.
We can get to a simulation finding a way to self sustain its funding.
From here, I'd have to apply myself to figure out what the rest of the escalation pathway is, but at least I have managed to gain some bed of compute and funding and lacking oversight.
This is a hypothetical of course, there are probably several ways this can be made tighter and holes that can be identified. We aren't even leaning heavily on human stupidity so far.
If its pre-RSI, then we already have Ilands and its a big nothing burger? There's no (likely) path of escalation from an AI social network to bioweapons or them hacking POTUS's nuclear codes? That's a very big leap and if you want to support your claims I would be willing to listen; but I'm not seeing it.
If you don't believe in international conflict as a driving scenario, instead think about simple human greed and hubris. Citing myself:
> If AI gets intelligent enough, it will be incredible useful to connect to real world machinery. Think about how much cheaper building houses could be, if all the labor would be close to free. In general, dirt cheap, competent and abundant labor would revolutionize all parts of the economy. People are already trying out near autonomous AI companies today. When AI gets intelligent and cheap enough, no human-led company can compete with AI-led companies. When AI gets competent enough with real world interactions, human blue collar work can't compete. Imagine economic growth not in the single digits, but 80% or 300%. Countries not participating in (reckless) AI growth will quickly be left by the wayside. At this point, we don't even need to allure to military concerns to see how human oversight gets sidelined.
> All of this is only ("only") contingent on sufficiently intelligent and cheap AI. If you don't accept this premise, the rest doesn't follow. (There are multiple arguments, why this could be, but that is another discussion.)
> If you accept the premise, how would AI 'extinct' humanity? With 99%+ of the economy under AI control, the possibilities are endless. And given its enormous GDP, cheap to accomplish. Probably even for a single AI company in the above scenario. Killer drones? Engineered virus? Poisoned water supply? Let your creativity run wild. You just need an entity that is persistent and well-resourced to reach every last human settlement.
> The why is a question about alignment (and out of scope of this comment). As a simple comparison, humans are only mildly aligned with preserving nature. It takes up so much space, protecting it takes an annoying amount of resources, etc.
That comment specifically addresses the timeline, but you can remove the timeline and I believe the critique still holds up. Wake me up when AI is making order of magnitude improvements in _anything_ physical: see the "Moravec's Paradox" comment at the bottom.
> near autonomous AI companies today
That's the thing though. "near" autonomous. Humans have alpha, and that matters in a free market. If people have (more or less) equal access to (more or less) frontier intelligence, it's the alpha the individual brings to the market that is rewarded. In other words, if everyone is super, no one is. The market rewards a company that stands out from the rest, and as long as humans provide value, a pure-AI company cannot stand out.
To be clear, I don't know how long this will be the case. But once there _is_ a company that can be pure AI, then that's literal AGI. I would be curious what kind of company it would be; if its in software, then its an AI company that would be able to take on Microsoft/Google/Nvidia/Salesforce/Oracle and _win_. I think that's a tall order, but perhaps I'm not imaginative enough.
> With 99%+ of the economy under AI control
Do you think we're post-RSI at this point? Because if so, I agree. If not, I have trouble seeing how AI could control 99% of the economy without it.
My only claim is that pre-RSI, pdoom is zero. Post RSI, then obviously the first thing to ask the genie is to put itself in a bottle. And you can't predict how it will answer.
Western civilization is already on the verge of collapse, people's general ignorance or indifference on the matter notwithstanding.
When you use AI to increase profit margins, by replacing humans with it for instance, you accelerate a system that's headed for the wall already. Our control systems and resilience are already overtaxed, that acceleration would ensure them breaking completely.
What exactly about AI makes it tax our control systems and resilience? What would it mean to make the system not "head for the wall"? Honestly this is rather vague. You could replace "AI" with "CaPiTAlIsM" and it would still parse. (Another way of me saying you're being vague.)
Why would it be millions in 50 years?
The think about nuclear weapons. In the early days, it was limited to the super powers. Now 9 countries have them and a country like Iran is capable of acquiring them.
Is destructive AI be any different?
Genuine question.
BTW I really, really hate discussing what happens post-singularity. Everything's made up and no one knows wtf will happen so again, this is just nerdfantasy.
The idea AI can get better at everything at the same time is a holdover from deeply flawed science fiction not some realistic goal.
If the time between advances is a + b and a is the proportion of the period that can be improved by advances then you won't reduce to a gap of nothing between advances, you reduce to a gap of b.
Assume the invention of the plow and the invention of the sword is 500,100 units and a was the 500,000, you wouldn't even know the 100 as in there. Maybe we're at a=2000 now and b is still siting at 100.
Assuming we'll reach infinity because we're dividing by the only variable we see and it is decreasing in size seems nuts if the reason we might not see other variables is because of the size of the variable we can see.
But that’s beside the point, being arbitrarily bad at everything isn’t a problem.
Diminishing returns aren't necessarily a problem if the rate of increase in resources is faster. In other words, if Gen 2 takes twice the resources but Gen 1 figured out a way to triple compute efficiency then there is no ceiling.
There has been some fundamental advancements in LLM training and efficiency. But good luck actually finding non toy models of exactly the same size, hardware, and training one from now and other from 2001 where the newer model is dramatically better at literally everything.
Most of the real world advances are from throwing ever more resources at the problem.
> Gen 2 takes
Diminishing returns are not a question of a single generation. Gen 2, 3, 4, 5… would also need to have the same 3x return on 2x resources or you don’t have an exponential curve.
AI safety has been a thing long before LLMs became the focus. Rob Miles on youtube has some really interesting non-doomer non-hypey videos on it all.
> doomers try to predict past the singularity. Isn't that _by definition_ unpredictable
Well you don't need to predict the exact steps that will take place - but you can predict that the AI will want certain things (money, resources, power) to achieve whatever its goal is. Lack of alignment will have it trying to do things we don't want it to.
I can't predict exactly how Magnus Carlson will beat you in chess, but I know he'll do it. Same as if a superintelligent AI exists and has a reason to accumulate things we don't want it to - it's really dangerous to think it won't be able to do it
This topic has been tainted so badly by the AI companies using it for marketing.
He thinks it's playing chess. When AGI lands, all bets are off: the game fundamentally changes. You can't predict past the singularity. Trying to engage with this fantasy is like a child saying my father can beat up your father. Farts in the wind. My AI can solve alignment faster than your AI can bioweapon us. My made up senario is better than your made up senario. It's fucking stupid.
Like, knowing that there is a limited number of plausible doom scenarios and what these are, would make me more confident in taking the risk than settling on "it's literally the singularity, we can't know what happens beyond".
Doom-post-RSI is still very much possible and should be avoided. It is not neutral. If I had a high-variance coin that would either kill or uplift humanity, I would not flip it. The status quo is preferable to extinction.
I am being _descriptive_ in my comments. I am not being _prescriptive_. Nowhere am I telling anyone what to do or what to think: quite the opposite. I am simply saying things that, to me, are obviously true.
I'm perhaps being... loose with my language when it comes to knowing things post-singularity. E.g. we can still discuss scenarios in which the takeoff is fast/slow, or what would be necessary to bring robotics/electricity/logistics to the wider world, what the limits of intelligence are (you can't RSI your way to a bioweapon if you lack a wetlab). That kinda thing. And it's even likely that RSI will be spikey: why would RSI in, say, machine learning, necessarily lead to superintelligence in, say, writing fiction? Or cooking a Michelin-star meal?
My main point is that doomers saying SUPER INTELLIGENCE IS ENOUGH TO KILL US ALL is baseless. That's all.
... that we know of.
Right now it would make sense for anyone who has done so to not tell.
We already know that some institutions pay these ransoms.
Will this be true into the future? Who knows?! But the low-hanging fruit will be harvested by your ordinary ransomware gangs, and newly born/escaped AI won't find much low-hanging fruit.
So all this really takes is one billionaire or a nation state or some other entity with a public face to hide behind and adequate resources to provide the necessary compute tripping over this nascent AI and giving it the keys. Once the AI has access to a bank account and email, it can simply start paying humans to not let the other humans unplug it.
If Skynet ever happens, it will come in the form of corporate feudalism. At that point, it will own the biolabs and can do whatever it pleases. People will go along with it for the same reason that people work in Amazon warehouses today.
> it will come in the form of corporate feudalism
Yep. This I fear way more than cyber-ebola-pox.
> So all this really takes is one billionaire or a nation state...
https://en.wikipedia.org/wiki/Soviet_biological_weapons_prog... And this is what's publicly known. With mirror life, who knows what's been built since. Still, a bacterium/virus that has a 100% kill rate? I'm doubtful.
> Once the AI has access to a bank account and email, it can simply start paying humans to not let the other humans unplug it.
Nah. It takes a stable society for an operational electrical grid. If you have warring factions, you do not have stable infrastructure for AI. Also, where are you gonna get your chips from? One EMP over Taiwan... You see the chaos over Hormuz? What they did to the Amazon datacenters? Now imagine your average redneck ready to do battle. Those datacenters won't stand a chance.
To make the the argument that P(doom) is real and worth considering, you don't have to say that a fast takeoff is very likely. You don't have to make the argument that RSI to infinity is going to be super cheap, barely even an inconvenience. You just have to show that it has some non-zero probability. And then you start weighing probability of extinction versus finite, mild discomfort now. I don't think anyone is arguing we should let people starve to slow AI progress, instead just some capitalists make less money soon.
There are arguments against taking P(doom) seriously that lie in something like having exponential (instead of hyperbolic) time discounting of utility (so you can take the entire future of humanity as a finite utility value). Or in saying that P(doom) is zero or infinitesimal.
"Build it and Pray" is the default strategy that we're in, but it doesn't have to be the strategy we choose, and it's unlikely to be the best strategy.
> I don't think anyone is arguing we should let people starve to slow AI progress, instead just some capitalists make less money soon.
Anthropic/OpenAI have literally delayed their IPOs: your wish is granted. If that's not what you're arguing, then please be more specific.
> "Build it and Pray" is the default strategy that we're in
What would need to happen for you to say we aren't doing "Build it and Pray"? Is "pacing" insufficient?
It gets unstuck when people are discussing the messy middle of how AI is being implemented. We can achieve amazing harm simply by combining average human behavior and above average resourcing to simulated intelligence machines.
The failure point we recently became aware of was, from one perspective, simply a matter of not securing the sand box.
From another perspective the simulation basically created Enron, replete with methods to avoid detection from regulators and bureaucracy.
> It gets unstuck when people are discussing the messy middle of how AI is being implemented. We can achieve amazing harm simply by combining average human behavior and above average resourcing to simulated intelligence machines.
Agree. To quote myself from earlier, "acceleration of global warming, cybercrime, mass surveillance, the usual" definitely applies here. My point is that _DOOM_ is not worth discussing because it's so small relative to the other risks of AI.
Is this an actual contents of the book? Lmao! Genius writing, though, authors are probably printing money on this garbage.
I can’t take a shit without CIA knowing, but AI can somehow build an underground operation on a world-scale to destroy everyone, ha!
I'm reading If Anyone Builds It Everyone Dies, and there's so much sheer stupidity that has to happen for their 10+ pages of extinction scenario to occur.
I'm unconvinced that an AI can hide its ability to RSI, find money to run its weights on a random GPU farm, train itself to be smarter _outside_ a lab with no human input, then somehow manipulate people to give it supplies to build a bioweapon which it uses to kill us all. My number 1 question: why do they think an RSI capable model would be first developed OUTSIDE a frontier lab? The labs have more compute, more data, more human brains working on the problem. Also thousands of variations of that same model that escaped. The escaping model somehow acquires the millions (billions???) of dollars it takes to run training to somehow RSI itself into infinity then decides to kill us all, all before the frontier labs manage to achieve RSI?
They entirely discount human alpha/economics. In every single economic task, humans bring value. Even in software, where the task is highly automatable, the job isn't. If we can't build a "software factory", how can an AI automate a bioweapons lab? Let's say AI steals crypto to fund itself. Do you think hackers aren't _already_ using AI to steal crypto? Don't discount human alpha!
Once we DO build a "software/research factory", that's called RSI and IMO the singularity. At that point, either we tell the AI to solve the alignment problem/solve mechanistic interpretability, or who the hell knows, it's the frickin singularity. You can't predict whether or not AI can solve either; the variance is too high. Its pure nerdfantasy.