Posts
Comments
Thanks for the reply. I remembered a recent article by Evans and thought that reasoning models might show a different behavior. Sorry if this sounds silly
Are you planning to test this on reasoning models?
I agree. But now people write so often about short timelines that it seems appropriate to recall the possible reason for the uncertainty.
Doesn't that seem like a reason to be optimistic about reasoning models?
There doesn't seem to be a consensus that ASI will be created in the next 5-10 years. This means that current technology leaders and their promises may be forgotten.
Does anyone else remember Ben Goertzel and Novamente? Or Hugo de Garis?
Yudkowsky may think that the plan 'Avert all creation of superintelligence in the near and medium term — augment human intelligence' has <5% chance of success, but your plan has <<1% chance. Obviously, you and he disagree not only on conclusions, but also on models.
EY is known for considering humanity almost doomed.
He may think that the idea of human intelligence augmentation is likely to fail. But it's the only hope. Of course, many will disagree with this.
It seems that we are already at the GPT 4.5 level? Except that reasoning models have confused everything, and increasing OOM on output can have the same effect as ~OOM on training, as I understand it.
By the way, you've analyzed the scaling of pretraining a lot. But what about inference scaling? It seems that o3 has already used thousands of GPUs to solve tasks in ARC-AGI.
Thank you. In conditions of extreme uncertainty about the timing and impact of AGI, it's nice to know at least something definite.
Can we assume that Gemini 2.0, GPT-4o, Claude 3.5 and other models with similar performance have a similar compute?
If we don't build fast enough, then the authoritarian countries could win.
Ideally it would be something like the UN, but given the geopolitical complexities, that doesn't seem very possible.
This sounds like a rejection of international coordination.
But there was coordination between the United States and the USSR on nuclear weapons issues, despite geopolitical tensions, for example. You can interact with countries you don't like without trying to destroy the world faster than them!
2 years ago, you seemed quite optimistic about AGI Safety/Alignment and had a long timeline.
Have your views changed since then?
I understand that hiring will be necessary in any case.
Keeping people as a commodity for acasual trade or pets seems like a more likely option.
If only one innovation separates us from AGI, we're fucked.
It seems that if OpenAI or Anthropic had agreed with you, they should have had even shorter timelines.
A short reading list which should be required before one has permission to opine. You can disagree, but step 1 is to at least make an effort to understand why some of the smartest people in the world (and 100% of the top 5 ai researchers — the group historically most skeptical about ai risk) think that we’re dancing on a volcano . [Flo suggests: There’s No Fire Alarm for Artificial General Intelligence, AGI Ruin: A List of Lethalities, Superintelligence by Nick Bostrom, and Superintelligence FAQ by Scott Alexander]
But Bostrom estimated the probability of extinction within a century as <20%. Scott Alexander estimated the risk from AI as 33%.
They could have changed their forecasts. But it seems strange to refer to them as a justification for confident doom.
I would expect that the absence of a global catastrophe for ~2 years after the creation of AGI would increase the chances of most people's survival. Especially in a scenario where alignment was easy.
After all, then there will be time for political and popular action. We can expect something strange when politicians and their voters finally understand the existential horror of the situation!
I don't know. Attempts to ban all AI? The Butlerian jihad? Nationalization of AI companies? Revolutions and military coups? Everything seems possible.
If AI respects the right to property, why shouldn't it respect the right to UBI if such a law is passed? The rapid growth of the economy will make it possible to feed many.
In fact, a world in which someone shrugs their shoulders and allows 99% of the population to die seems obviously unsafe for the remaining 1%.
It's possible that we won't get something that deserves the name ASI or TAI until, for example, 2030.
And a lot can change in more than 5 years!
The current panic seems excessive. We do not live in a world where all reasonable people expect the emergence of artificial superintelligence in the next few years and the extinction of humanity soon after that.
The situation is very worrying, and this is the most likely cause of death for all of us in the coming years, yes. But I don't understand how anyone can be so sure of a bad outcome as to consider people's survival a miracle.
Then what is the probability of extinction caused by AI?
Of course, capital is useful in order to exert influence now. Although I would suggest that for a noticeable impact on events, capital or power is needed, which are inaccessible to the vast majority of the population.
But can we end up in a world where the richest 1% or 0.1% will survive, and the rest will die? Unlikely. Even if property rights were respected, such a world would have to turn into a mad hell.
Even a world in which only people like Sam Altman and their entourage will survive the singularity seems more likely.
But the most likely options should be the extinction of all or the survival of almost all without a strong correlation with current well-being. Am I mistaken?
Most experts do not believe that we are certainly (>80%) doomed. It would be an overreaction to give up after the news that politicians and CEO are behaving like politicians and CEO.
But your P(doom) still only 0.6? Or are you considering disempowerment from AI separately?
It still surprises me that so many people agree on most issues, but have very different P(doom). And even long-term patient discussions do not bring people's views closer. It will probably be even more difficult to convince a politician or the CEO.
So what's your P(doom)?
I have already tried to collect the most complete collection of quotes here. But it is already very outdated.
It seems that in 2014 he believed that p(doom) was less than 20%
I do expect some of the potential readers of this post to live in a very unsafe environment - e.g. parts of current-day Ukraine, or if they live together with someone abusive - where they are actually in constant danger.
I live ~14 kilometers from the front line, in Donetsk. Yeah, it's pretty... stressful.
But I think I'm much more likely to be killed by an unaligned superintelligence than an artillery barrage.
Most people survive urban battles, so I have a good chance.
And in fact, many people worry even less than I do! People get tired of feeling in danger all the time.
'“Then why are you doing the research?” Bostrom asked.
“I could give you the usual arguments,” Hinton said. “But the truth is that the prospect of discovery is too sweet.” He smiled awkwardly, the word hanging in the air—an echo of Oppenheimer, who famously said of the bomb, “When you see something that is technically sweet, you go ahead and do it, and you argue about what to do about it only after you have had your technical success.”'
'I asked Hinton if he believed an A.I. could be controlled. “That is like asking if a child can control his parents,” he said. “It can happen with a baby and a mother—there is biological hardwiring—but there is not a good track record of less intelligent things controlling things of greater intelligence.” He looked as if he might elaborate. Then a scientist called out, “Let’s all get drinks!”'
Hinton seems to be more responsible now!
The level of concern and seriousness I see from ML researchers discussing AGI on any social media platform or in any mainstream venue seems wildly out of step with "half of us think there's a 10+% chance of our work resulting in an existential catastrophe".
In fairness, this is not quite half the researchers. This is half the agreed survey.
I expect that worried researchers are more likely to agree to participate in the survey.
Thanks for your answer, this is important to me.
I am not an American (so excuse me for my bad English!), so my opinion about the admissibility of attack on the US data centers is not so important. This is not my country.
But reading about the bombing of Russian data centers as an example was unpleasant. It sounds like a Western bias for me. And not only for me.
If the text is aimed at readers not only from the First World countries, well, perhaps the authors should do such a clarification as you did! Then it will not look like political hypocrisy. Or not write about air strikes at all, because people are distracted for discussing this.
I'm not an American, so my consent doesn't mean much :)
Suppose China and Russia accepted the Yudkowsky's initiative. But the USA is not. Would you support to bomb a American data center?
I can provide several links. And you choose those that are suitable. If suitable. The problem is that I retained not the most complete justifications, but the most ... certain and brief. I will try not to repeat those that are already in the answers here.
Jaron Lanier and Neil Gershenfeld
Magnus Vinding and his list
Maybe Abram Demski? But he changed his mind, probably.
Well, Stuart Russell. But this is a book. I can quote.
I do think that I’m an optimist. I think there’s a long way to go. We are just scratching the surface of this control problem, but the first scratching seems to be productive, and so I’m reasonably optimistic that there is a path of AI development that leads us to what we might describe as “provably beneficial AI systems.”
There are also a large number of reasonable people who directly called themselves optimists or pointed out a relatively small probability of death from AI. But usually they did not justify this in ~ 500 words…
I also recommend this book.
My fault. I should just copy separate quotes and links here.
I have collected many quotes with links about the prospects of AGI. Most people were optimistic.
Glad you understood me. Sorry for my english!
Of course, the following examples themselves do not prove the opportunity to solve the entire problem of AGI alignment! But it seems to me that this direction is interesting and strongly underestimated. Well, someone smarter than me can look at this idea and say that it is bullshit, at least.
Partly this is a source of intuition for me, that the creation of aligned superintellect is possible. And maybe not even as hard as it seems.
We have many examples of creatures that follow the goals of someone more stupid. And the mechanism that is responsible for this should not be very complex.
Such a stupid process, as a natural selection, was able to create mentioned capabilities. It must be achievable for us.
It seems to me that the brains of many animals can be aligned with the goals of someone much more stupid themselves.
People and pets. Parasites and animals. Even ants and fungus.
Perhaps the connection that we would like to have with superintellence, is observed on a much smaller scale.
I apologize for the stupid question. But…
Do we have more chances to survive in the world, which is closer to Orwell's '1984'?
It seems to me that we are moving towards more global surveillance and control. China's regime in 2021 may seem extremely liberal for an observer in 2040.
I guess I missed the term gray goo. I apologize for this and for my bad English.
Is it possible to replace it on the 'using nanotechnologies to attain a decisive strategic advantage'?
I mean the discussion of the prospects for nanotechnologies on SL4 20+ years ago. This is especially:
My current estimate, as of right now, is that humanity has no more than a 30% chance of making it, probably less. The most realistic estimate for a seed AI transcendence is 2020; nanowar, before 2015.
I understand that since then the views of EY have changed in many ways. But I am interested in the views of experts on the possibility of using nanotechnology for those scenarios that he implies now. That little thing I found.
Nanosystems are definitely possible, if you doubt that read Drexler’s Nanosystems and perhaps Engines of Creation and think about physics.
Is there something like the result of a survey of experts about the feasibility of drexlerian nanotechnology? Are there any consensus among specialists about the possibility of a gray goo scenario?
Drexler and Yudkowsky both extremely overestimated the impact of molecular nanotechnology in the past.
I do not know the opinions of experts on this issue. And I lack competence for such conclusions, sorry.
AlexNet was the first publication that leveraged graphical processing units (GPUs) for the training run
Do you mean the first of the data points on the chart? The GPU was used for DL long before AlexNet. References: [1], [2], [3], [4], [5].
But this is 1999, yes.
Probably that:
When we didn’t have enough information to directly count FLOPs, we looked GPU training time and total number of GPUs used and assumed a utilization efficiency (usually 0.33)
This can be useful:
We trained the league using three main agents (one for each StarCraft race), three main exploiter agents (one for each race), and six league exploiter agents (two for each race). Each agent was trained using 32 third-generation tensor processing units (TPUs) over 44 days
Perhaps my large collection of quotes about the impact of AI on the future of humanity here will be helpful.
Then it is worth considering the majority of experts from the FHI to be extreme optimists, the same 20%? I really tried to find all the publicly available forecasts of experts and those who were confident that AI would lead to the extinction of humanity, there were very few among them. But I have no reason not to believe you or Luke Muehlhauser who introduced AI safety experts as even more confident pessimists: ’Many of them are, roughly speaking, 65%-85% confident that machine superintelligence will lead to human extinction’ . The reason may be that not everyone agrees, whose opinion is worth considering.
What about this and this? Here, some researchers at the FHI give other probabilities.
I meant the results of such polls: https://www.thatsmags.com/china/post/15129/happy-planet-index-china-is-72nd-happiest-country-in-the-world. Well, it doesn’t matter.
I think that I could sleep better if everyone recognized the reduction of existential risks in a less free world.
I’m not sure that I can trust news sources that are interested in outlining China.
In any case, this does not seem to stop the Chinese people from feeling happier than the US people.
I cited this date just to contrast with your forecast. My intuition is more likely to point to AI in the 2050-2060 years.
And yes, I expect that in 2050 it will be possible to monitor the behavior of each person in countries 24/7. I can’t say that it makes me happy, but I think that the vast majority will put up with this. I don't believe in a liberal democratic utopia, but the end of the world seems unlikely to me.