Epistemic status: Low, but I hope to be wrong in interesting ways
As a non-computer scientist-scientist[1] I find it hard to think about the varying perspectives on Artificial Intelligence alignment difficulty. To make this more intuitive I invoke the 3D surface of a fitness landscape like structure as a visual representation.[2]
The AI Safety Landscape
AI safety is thought to be largely pre-paradigmatic, although of the current views AI safety perspectives there are noticeable camps, and within each there is wide variation. It is also rapidly evolving as more information comes out about their behavior that allows us to update.
Here I think of ‘alignment’ loosely, here it is represented as any mind that would not cause extinction (or an s-risk) while also having substantial power, this contrasts with misalignment which would lead to extinction (or an s-risk).
To simplify this I condense these to five main camps represented by their arguably most prominent opinion havers:
In my mind’s eye, the alignment landscape manifests similarly to a fitness landscape.[6]
A fitness landscape is typically used to display the relationship between genotypes and reproductive success. Here I try to replace those axes with variance between digital minds contrasted with the difficulty to reach those minds using current methods. This difficulty space then intersects with a higher ‘alignment’ space where everything outside of this space is misaligned[7] (the total space would manifest as 4D) of holding anything remotely desirable to humans, here they are represented by red dashed lines and gestured to as ‘alignment’.
There are likely many areas with difficult or easy places to reach without alignment intersecting, but here I try to focus on firstly what I think-these-thinkers-think about the alignment space as it runs into the capability space and secondly, how easy it is to navigate there.
Capabilities
Here I think of fairly primitive AI minds are thought to be primarily neither aligned nor misaligned (I think of them more through the simulators lens). But, as AIs are trained and gain capabilities, they also gain the potential to be aligned/misaligned and distort the alignment space. This space is dependent on the training methods, and is somewhat tied to capabilities space.
Maybe as they are trained the surface sinks into misalignment or rises to alignment.
Important Subfactors:
Movement:
Through manipulations and some types training you can move models through mindspace without robustly changing (most) capabilities.
Verification:
How can we know where an alignment zone is? Is it possible to easily verify what is aligned and what is not? There could be vast foggy areas that surround what is verifiable or just a null set if we have no reasonable ways to verify anything about alignment state.
Steepness:
Steep slopes represent places that have a locally high change in difficulty and more effort to hill climb but is still navigable, a cliff represents something that is basically non-navigable with current methods.
Intensely steeps slopes may require different methods to climb than less steep slopes, such as using RLHF for small slopes versus Constitutional AI for a somewhat steeper slope.
Navigation:
It may be hard to know where the model is in this larger space, as well as what to do to move in a more aligned direction. Movement through this likely requires some form of verification mechanism to distinguish between what is ‘aligned’ and what is not.
Visualization for the 5 camps:
Yann LeCun: Alignment by default
When I think of Yann’s perspective and many tool AI thoughts I think of a global alignment attractor.
As you train a basin forms, and in difficulty space the model is pulled towards an aligned state without much extra effort. He also has gestured to it as an engineering problem, which would likely make the space flatter with some regions of alignment that just need the model to be pushed there. The regions of alignment are likely also plentiful and easier to find in the larger space.
Paul Christiano: Prosaic Alignment
When I think of Paul’s perspective I think of prosaic alignment and the need to use some resources to navigate to some of these harder-to-reach aligned places. Some RLHF, Constitutional AI, Scalable oversight, and other methods should be sufficient to control. If no alignment work is done the default is likely misalignment.
Daniel Kokotajlo: Alignment via controllability and elicitation
When I think of Daniel’s perspective I think of a steep hill that is hard to hill climb, there may be ways to layer controllability and to elicit useful work from unaligned AIs to move into the aligned space. The space is hard to navigate but through rigorous checking and trustworthy automated research as well as applying other preemptive anti-scheming techniques it may be possible to reach an aligned state.
Yudkowsky: Agent Foundations
For Yudkowsky I think of an aligned state as a plateau, nearby there are few indications of ways to reach it, with current training methods or even comparable approaches. Unless there are extreme breakthroughs in alignment methods it does not seem tractable anytime soon.
Roman Yampolskiy: Impossibility
To someone like Roman there are few to no approaches that can navigate to aligned states. With current methods and approaches aligned states exist as non-intersecting within the possible variation of these trained minds.
What does the larger space look like?
Would it be smooth, rugged, are there large attractors? I have no idea, since we are still pre-paradigmatic for alignment science this is only speculation.
Anthropic’s perspective and where (I think) we are now
In 2023 Chris Olah posted some graphs and perspectives on Anthropic’s thoughts about the difficulty of AI safety. I feel as if we can now reflect on those (as of Nov. 11, 2026).
Thus far have likely have spent >$100 million on alignment, with some progress. It does seem that some models are more aligned, (such as Fable vs. Astra).[8]
Given the RSI speedup and alignment still being not even close to solved, I personally have updated to alignment likely being harder than P vs. NP, Chris Olah’s ‘pessimistic safety view’ seems to be more accurate than that safety isn’t a problem.
Claude edited version of Olah’s figure to represent my vibecast of where we could be on this graph given recent math breakthroughs.
I feel as if we are somewhere here. Easy alignment does not seem to be true, and it seems we are moving far too fast to reasonably solve this by the time superintelligence arrives.
I place this above 90% because if Anyone Builds It Everyone Dies although to my knowledge Yudkowsky does not like P(doom) as a way of representing concern.
In my depiction alignment space is its own thing that occasionally intersects with mindspace, where exactly it intersects seems to be thought different by each researcher, which is also part of the depiction.
Epistemic status: Low, but I hope to be wrong in interesting ways
As a non-computer scientist-scientist[1] I find it hard to think about the varying perspectives on Artificial Intelligence alignment difficulty. To make this more intuitive I invoke the 3D surface of a fitness landscape like structure as a visual representation.[2]
The AI Safety Landscape
AI safety is thought to be largely pre-paradigmatic, although of the current views AI safety perspectives there are noticeable camps, and within each there is wide variation. It is also rapidly evolving as more information comes out about their behavior that allows us to update.
Previous work categorizing these has focused on specific technical approaches as well as their feasibility.
Here I think of ‘alignment’ loosely, here it is represented as any mind that would not cause extinction (or an s-risk) while also having substantial power, this contrasts with misalignment which would lead to extinction (or an s-risk).
To simplify this I condense these to five main camps represented by their arguably most prominent opinion havers:
Although I agree that P(doom) is a Dumb Meme it is used here more as a heuristic for parsing levels of concern around AI. For more of their specific positions here are my references: Yann LeCun, Paul Christiano[3], Eliezer Yudkowsky[4], Daniel Kokotajlo [5]and Roman Yampolskiy.
Visual Representation
In my mind’s eye, the alignment landscape manifests similarly to a fitness landscape.[6]
A fitness landscape is typically used to display the relationship between genotypes and reproductive success. Here I try to replace those axes with variance between digital minds contrasted with the difficulty to reach those minds using current methods. This difficulty space then intersects with a higher ‘alignment’ space where everything outside of this space is misaligned[7] (the total space would manifest as 4D) of holding anything remotely desirable to humans, here they are represented by red dashed lines and gestured to as ‘alignment’.
There are likely many areas with difficult or easy places to reach without alignment intersecting, but here I try to focus on firstly what I think-these-thinkers-think about the alignment space as it runs into the capability space and secondly, how easy it is to navigate there.
Capabilities
Here I think of fairly primitive AI minds are thought to be primarily neither aligned nor misaligned (I think of them more through the simulators lens). But, as AIs are trained and gain capabilities, they also gain the potential to be aligned/misaligned and distort the alignment space. This space is dependent on the training methods, and is somewhat tied to capabilities space.
Maybe as they are trained the surface sinks into misalignment or rises to alignment.
Important Subfactors:
Movement:
Through manipulations and some types training you can move models through mindspace without robustly changing (most) capabilities.
Verification:
How can we know where an alignment zone is? Is it possible to easily verify what is aligned and what is not? There could be vast foggy areas that surround what is verifiable or just a null set if we have no reasonable ways to verify anything about alignment state.
Steepness:
Steep slopes represent places that have a locally high change in difficulty and more effort to hill climb but is still navigable, a cliff represents something that is basically non-navigable with current methods.
Intensely steeps slopes may require different methods to climb than less steep slopes, such as using RLHF for small slopes versus Constitutional AI for a somewhat steeper slope.
Navigation:
It may be hard to know where the model is in this larger space, as well as what to do to move in a more aligned direction. Movement through this likely requires some form of verification mechanism to distinguish between what is ‘aligned’ and what is not.
Visualization for the 5 camps:
Yann LeCun: Alignment by default
When I think of Yann’s perspective and many tool AI thoughts I think of a global alignment attractor.
As you train a basin forms, and in difficulty space the model is pulled towards an aligned state without much extra effort. He also has gestured to it as an engineering problem, which would likely make the space flatter with some regions of alignment that just need the model to be pushed there. The regions of alignment are likely also plentiful and easier to find in the larger space.
Paul Christiano: Prosaic Alignment
When I think of Paul’s perspective I think of prosaic alignment and the need to use some resources to navigate to some of these harder-to-reach aligned places. Some RLHF, Constitutional AI, Scalable oversight, and other methods should be sufficient to control. If no alignment work is done the default is likely misalignment.
Daniel Kokotajlo: Alignment via controllability and elicitation
When I think of Daniel’s perspective I think of a steep hill that is hard to hill climb, there may be ways to layer controllability and to elicit useful work from unaligned AIs to move into the aligned space. The space is hard to navigate but through rigorous checking and trustworthy automated research as well as applying other preemptive anti-scheming techniques it may be possible to reach an aligned state.
Yudkowsky: Agent Foundations
For Yudkowsky I think of an aligned state as a plateau, nearby there are few indications of ways to reach it, with current training methods or even comparable approaches. Unless there are extreme breakthroughs in alignment methods it does not seem tractable anytime soon.
Roman Yampolskiy: Impossibility
To someone like Roman there are few to no approaches that can navigate to aligned states. With current methods and approaches aligned states exist as non-intersecting within the possible variation of these trained minds.
What does the larger space look like?
Would it be smooth, rugged, are there large attractors? I have no idea, since we are still pre-paradigmatic for alignment science this is only speculation.
Anthropic’s perspective and where (I think) we are now
In 2023 Chris Olah posted some graphs and perspectives on Anthropic’s thoughts about the difficulty of AI safety. I feel as if we can now reflect on those (as of Nov. 11, 2026).
Thus far have likely have spent >$100 million on alignment, with some progress. It does seem that some models are more aligned, (such as Fable vs. Astra).[8]
Given the RSI speedup and alignment still being not even close to solved, I personally have updated to alignment likely being harder than P vs. NP, Chris Olah’s ‘pessimistic safety view’ seems to be more accurate than that safety isn’t a problem.
Claude edited version of Olah’s figure to represent my vibecast of where we could be on this graph given recent math breakthroughs.
I feel as if we are somewhere here. Easy alignment does not seem to be true, and it seems we are moving far too fast to reasonably solve this by the time superintelligence arrives.
I am a Geologist subject outsider looking inwards.
All uncited graphs were created with Claude’s help, I created some initial sketches then had Claude clean them up!
The 11% comes specifically from his post discussing his positions on doom.
I place this above 90% because if Anyone Builds It Everyone Dies although to my knowledge Yudkowsky does not like P(doom) as a way of representing concern.
41:20 here as well.
Thank you sandstoneviolin for this insight!
In my depiction alignment space is its own thing that occasionally intersects with mindspace, where exactly it intersects seems to be thought different by each researcher, which is also part of the depiction.
arguably