Last week, METR released their Independent Assessment of AI R&D Acceleration at Anthropic. They estimate that Anthropic “is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” They also add that “there is a significant risk that further capabilities progress could lead to even faster acceleration.” This assessment implies that timelines to ASI could be very short. A naive forecast based on this information says that we will get ASI within a year.
One of the key data points that informed METR’s analysis was a 1.9x increase in the improvement rate of AECI, Anthropic’s internal version of the Epoch Capabilities Index. This is particularly notable because METR estimates that the majority of the acceleration, maybe 80%, is from AI contributions, meaning that we could be seeing the start of an RSI loop causing superexponential growth in capabilities, as measured by AECI.
However, an alternative possibility is that Anthropic’s AIs might be assisting with ML research in a way that overfits to AECI such that this acceleration will not be sustained. METR gestures at this when they note that AECI’s focus on verifiable tasks is one of three key uncertainties that could result in them overestimating AI R&D acceleration. In this post, I discuss this uncertainty, why I believe it points to a legitimate reason that AECI improvements may decelerate, and why this deceleration, more than generalization to fuzzy tasks, is the crux of the relationship between verifiable tasks and timelines.
Second-order AECI might be overfit
In order to discuss this topic, it is useful to differentiate between how good a model is at AECI, how good a model is at training models to be good at AECI, how good a model is at training models to be good at training models to be good at AECI, etc. I will use the terms AECI, AECI’, and AECI’’ to refer to each of these properties.
My point rests on the claim that AECI’ can be improved without improving general ML research skills. In theory, “general ML research” is the kind of skill that would keep bearing fruit across any number of AI generations: if gen 1 is better at ML research, then it will be better at making gen 2 better at ML research in a way that makes gen 3 better at ML research, and so on. In contrast, an AI that is good at AECI’ might make choices that improve the next generation's AECI score but don’t make the next generation robustly smarter and better at carrying those improvements down the line.
There are a couple of sources of selection that I believe could lead to overfitting to AECI’ such that it comes apart from general ML research skills.
Humans train AIs and select the ones that are good at AECI’. One way this could happen is by Anthropic directly training against AECI’. For example, they could train AIs on NanoGPT speedrun-type tasks where the AI has to come up with algorithmic improvements to training that increase AIs’ benchmark scores with minimal compute. It’s possible that these algorithmic improvements could be overfit to the benchmarks (these benchmarks could include AECI or be correlated with it) such that they don’t lead to AIs that are more compute efficient across the board. A more subtle way humans might select for AECI’ is by developing a preference for models that seem to be coming up with good ML research ideas but where the quality of the ML research is being measured against (and overfit to) verifiable benchmarks.
AIs train new AIs and select the ones that are good at AECI. Even without humans overfitting to metrics for AECI’, AIs might naturally be more skilled in AECI’ than general ML research. For example, AIs might have poor taste but compensate by iterating harder (compared to humans) against verifiable benchmarks. AIs iterating against W2SG benchmarks (which involve training AIs), for instance, struggled with overfitting. They beat humans on the primary metric, performance gap recovered, but the gain didn’t hold up when transferred to a production-scale setting, and the authors flag finding metrics agents “can reliably hill-climb without overfitting” as a key challenge. In general, AIs may select harder for AIs that score well on benchmarks even if those AIs are less generally intelligent.
These points correspond to the labels 1 and 2 on the diagram below.
Overfitting to AECI’ could be one instance of a broader pattern in which models pick low-hanging fruit, producing a quick speedup that ultimately diminishes. Their jagged intelligence might give them a comparative advantage relative to humans that drives a burst of early insights, but that advantage could dry up as the shared frontier of human and AI insight becomes more saturated. This seems especially likely if AI's comparative advantage rests more on iteration and local improvements than on conceptual taste and systematic improvements.
This early acceleration dynamic is one possible explanation for the jump in the rate of AECI improvement that METR attributes to AI. If this is the case, then ML research improvement from AI could saturate, leading to deceleration later.
Why ML research is the crux
In my analysis, I focus on the gap between AECI’ and general ML research skills in order to separate automated ML improvements that will drive a strong positive feedback loop from those that will eventually level off. However, someone could also be concerned with positive feedback loops on verifiable tasks that don’t translate to positive feedback loops on fuzzy tasks. I put much less credence in this concern because I think we are likely to see at least some transfer between verifiable tasks and fuzzy tasks. This is because some verifiable tasks are central to general improvements in intelligence. Sample efficiency is one example. It can be measured by how few samples an algorithm needs to reach a given score on a benchmark. Sample efficiency is central to intelligence because it gets at generalization: higher sample efficiency means better performance out of distribution. Better out-of-distribution performance, in turn, means better performance on fuzzy tasks, which depend on generalization because directly relevant metrics for them are scarce.
Additionally, the transfer between verifiable tasks and fuzzy tasks does not need to be high. As long as you think that even a small fraction of verifiable capabilities transfer to fuzzy tasks, a FOOM on verifiable tasks will quickly transfer to fuzzy tasks.
Thus, I don’t think there is a compelling case for long timelines on account of AIs struggling to transfer verifiable skills to fuzzy tasks. AI’s acceleration on verifiable tasks will have to saturate in order for timelines to be extended.
Conclusion
The overfitting consideration I have laid out here is only one of many for determining how sustainable the acceleration in AI capabilities will be. For example, it’s possible that there is some degree of overfitting and also an underlying superexponential trend. My all-things-considered view is that there is a 60% chance we will see a deceleration in AI progress in the next two years, but a significant, maybe 30%, chance we will see ASI by the end of 2027.
Thank you to Jadyn Malone, Nabil Baugher, Lukas Finveddan, and Alex Mallen for useful feedback and discussion.
Last week, METR released their Independent Assessment of AI R&D Acceleration at Anthropic. They estimate that Anthropic “is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” They also add that “there is a significant risk that further capabilities progress could lead to even faster acceleration.” This assessment implies that timelines to ASI could be very short. A naive forecast based on this information says that we will get ASI within a year.
One of the key data points that informed METR’s analysis was a 1.9x increase in the improvement rate of AECI, Anthropic’s internal version of the Epoch Capabilities Index. This is particularly notable because METR estimates that the majority of the acceleration, maybe 80%, is from AI contributions, meaning that we could be seeing the start of an RSI loop causing superexponential growth in capabilities, as measured by AECI.
However, an alternative possibility is that Anthropic’s AIs might be assisting with ML research in a way that overfits to AECI such that this acceleration will not be sustained. METR gestures at this when they note that AECI’s focus on verifiable tasks is one of three key uncertainties that could result in them overestimating AI R&D acceleration. In this post, I discuss this uncertainty, why I believe it points to a legitimate reason that AECI improvements may decelerate, and why this deceleration, more than generalization to fuzzy tasks, is the crux of the relationship between verifiable tasks and timelines.
Second-order AECI might be overfit
In order to discuss this topic, it is useful to differentiate between how good a model is at AECI, how good a model is at training models to be good at AECI, how good a model is at training models to be good at training models to be good at AECI, etc. I will use the terms AECI, AECI’, and AECI’’ to refer to each of these properties.
My point rests on the claim that AECI’ can be improved without improving general ML research skills. In theory, “general ML research” is the kind of skill that would keep bearing fruit across any number of AI generations: if gen 1 is better at ML research, then it will be better at making gen 2 better at ML research in a way that makes gen 3 better at ML research, and so on. In contrast, an AI that is good at AECI’ might make choices that improve the next generation's AECI score but don’t make the next generation robustly smarter and better at carrying those improvements down the line.
There are a couple of sources of selection that I believe could lead to overfitting to AECI’ such that it comes apart from general ML research skills.
These points correspond to the labels 1 and 2 on the diagram below.
Overfitting to AECI’ could be one instance of a broader pattern in which models pick low-hanging fruit, producing a quick speedup that ultimately diminishes. Their jagged intelligence might give them a comparative advantage relative to humans that drives a burst of early insights, but that advantage could dry up as the shared frontier of human and AI insight becomes more saturated. This seems especially likely if AI's comparative advantage rests more on iteration and local improvements than on conceptual taste and systematic improvements.
This early acceleration dynamic is one possible explanation for the jump in the rate of AECI improvement that METR attributes to AI. If this is the case, then ML research improvement from AI could saturate, leading to deceleration later.
Why ML research is the crux
In my analysis, I focus on the gap between AECI’ and general ML research skills in order to separate automated ML improvements that will drive a strong positive feedback loop from those that will eventually level off. However, someone could also be concerned with positive feedback loops on verifiable tasks that don’t translate to positive feedback loops on fuzzy tasks. I put much less credence in this concern because I think we are likely to see at least some transfer between verifiable tasks and fuzzy tasks. This is because some verifiable tasks are central to general improvements in intelligence. Sample efficiency is one example. It can be measured by how few samples an algorithm needs to reach a given score on a benchmark. Sample efficiency is central to intelligence because it gets at generalization: higher sample efficiency means better performance out of distribution. Better out-of-distribution performance, in turn, means better performance on fuzzy tasks, which depend on generalization because directly relevant metrics for them are scarce.
Additionally, the transfer between verifiable tasks and fuzzy tasks does not need to be high. As long as you think that even a small fraction of verifiable capabilities transfer to fuzzy tasks, a FOOM on verifiable tasks will quickly transfer to fuzzy tasks.
Thus, I don’t think there is a compelling case for long timelines on account of AIs struggling to transfer verifiable skills to fuzzy tasks. AI’s acceleration on verifiable tasks will have to saturate in order for timelines to be extended.
Conclusion
The overfitting consideration I have laid out here is only one of many for determining how sustainable the acceleration in AI capabilities will be. For example, it’s possible that there is some degree of overfitting and also an underlying superexponential trend. My all-things-considered view is that there is a 60% chance we will see a deceleration in AI progress in the next two years, but a significant, maybe 30%, chance we will see ASI by the end of 2027.
Thank you to Jadyn Malone, Nabil Baugher, Lukas Finveddan, and Alex Mallen for useful feedback and discussion.