Well, we have surveys like this one indicating that they don’t take into account the likelihood of an existential catastrophe.
It seems to me that many forecasters are thinking about a trajectory that could lead to the creation of ASI with a certain probability in a certain year. But these models can be disrupted due to other factors such as wars, sanctions, restrictions on research, social upheaval, and so on.
teradimich
Karma: 266
I share the high concern about the potential for an existential threat from ASI, but I find the confidence levels of ≥50% X-risk in the near term to be epistemically overconfident.
To me, the probability of human extinction due to unaligned ASI must be decomposed into sequential factors.
P(ASI by 2035 while maintaining the current trajectory)≤0.75
P(no major disruption to development before 2035)≤0.8
P(alignment/control problem is NOT solved before 2035)≤0.9
P(unaligned ASI will not leave us alive for some reason)≤0.9
So 0.75×0.8×0.9×0.9≤0.486.
I am highly uncertain about these specific values and would personally prefer estimates closer to 0.5 for each component. But it seems to me that such variables should be taken into account when thinking about how doomed the world we know is.
But is it appropriate to be ~98% sure that the ASI level will be achieved in the coming years?
If not, then it seems reasonable to allow more uncertainty.
To prove that the forecasts are well calibrated, it would be worthwhile to make more verifiable statements. I have often seen claims that Yudkowsky has perfectly calibrated probabilities, but according to his other public forecasts or his page in Manifold, it does not seem so.
What do you think about GPT-5? Is this a GPT-4.5 scale model, but with a lot of RLVR training?
keeps the future of humanity in a good shape (as well as making it harmless)
Is this the result you expect by default? Or is this just one of many unlikely scenarios (like Hanson’s ‘The Age of Em’) that are worth considering?
I am sitting here crying as the last remaining bits of diplomatic goodwill and hope for internationally coordinated treaties on coordinating the AI takeoff evaporates.
We can still hope that we won’t get AGI in the next couple of years. Society’s attitude towards AI is already negative, and we’re even seeing some congressmen openly discuss the existential risks. This growing awareness might just lead to meaningful policy changes in the future.
plausibly about 3e26 FLOPs
Or 6e26 (in FP8 FLOPs).
And already on February 17th, Colossus had 150k+ GPU. It seems that in the April message they were talking about 200k GPUs. Judging by Musk’s interview, this could mean 150,000 H100 and 50,000 H200. Perhaps the time and GPU were enough to train a GPT-5 scale model?
I sympathize with this line of thinking, but I’ve never understood something like P(doom)>0.8.
The analogies with cancer or poison seem a bit odd, because we’re trying to estimate the probability of an event that has never happened before. Without relying on anything like physical laws, without anything close to consensus. Even among the people who proposed the key ideas of the AI Risk discussions, not all were confident pessimists.
We have too many unknowns. We don’t know when superintelligence will appear. We can’t predict how governments and corporations will treat AI in the coming years. We don’t know what will happen if someone tries to use a sufficiently advanced AI for automated safety research. Or narrow AI might change the situation in the world before superintelligence appears. Our civilization could collapse for any number of reasons.
And I don’t think we can say for sure what superintelligence will do to humans.
Earlier, you wrote about a change to your AGI timelines.
What about p(doom)? It seems that in recent months there have been reasons for both optimism and pessimism.
It seems a little surprising to me how rarely confident pessimists (p(doom)>0.9) they argue with moderate optimists (p(doom)≤0.5).
I’m not specifically talking about this post. But it would be interesting if people revealed their disagreement more often.
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.
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.
Have you seen this market?
When will an AI model using neuralese recurrence be first released to the public? (currently 11 Oct 2027)