I’ve seen a lot of discussion about the METR acceleration report. I think many of the models used to discuss it predict acceleration that is too fast. Here is a simple-ish model that is tied as closely as I can make it to relevant empirical data to predict how things will unfold given what we know of AI acceleration.
Suppose that Anthropic uses some amount of compute to train models using their training pipeline. The amount of compute and the efficiency of the pipeline are both growing rapidly over time. The difficulty of improving models also increases rapidly as a function of model capacity, leading to the roughly linear growth in model capability we have historically observed. The rate of improvement in the pipeline is a function of average researcher productivity. Historical estimates suggest that training compute used to train models at the frontier has been growing 5x per annum and training efficiency has been growing at 3x. Given the historical growth rate of model capabilities (measured via ECI or AECI) is about 15 points a year, it suggests that multiplying the effective training effort by 15X results in a new model with ECI 15 points higher than the old one, or about 20% more effective training input needed for a model with 1 more ECI.
Then in october 2025 we have a shift with 2 elements. Researcher productivity begins to be uplifted by AI assistance, and other unspecified factors also increase the rate of research efficiency increase. The net effect is that over the following year, capabilities improve almost twice as fast.
Modelling training efficiency and capability
We can express our first assumption with a simple equation:
Here, C is training compute, E is training efficiency, I is model capability measured in AECI, and K is a constant. The parameter κ represents the increasing difficulty of training more capable models.
Rearranging:
This gives us an important result. If training compute and training efficiency grow exponentially, model capability grows linearly. In particular, 5× annual compute growth and 3× annual training-efficiency growth give us 15× annual growth in effective training inputs.
We can therefore estimate κ from the historical rate of capability improvement:
Now we can ask what must have changed to produce the acceleration observed by METR.
If we assume training compute continues growing at the same rate, the additional capability improvements must come from faster improvements in training efficiency.
With 1.9× acceleration, the annualized rate of capability improvement rises from approximately 15 to 28.5 AECI points. Our equation implies that achieving this requires training efficiency to improve at an annualized rate equivalent to approximately 34×, rather than the historical 3×.
But not all of that improvement is attributable to AI. Taking METR’s tentative 80% estimate as a share of the additional acceleration, approximately 10.8 of the additional 13.5 AECI points per year are attributable to AI, while 2.7 are attributable to other factors.
Without AI assistance, the current rate of progress would therefore be approximately 17.7 AECI points per year. This corresponds to an annualized training-efficiency improvement of approximately 4.9×.
This gives us the quantities we need to model the contribution of researcher productivity.
Modelling the productivity of research
Let P represent the productivity of Anthropic’s researchers relative to working without AI assistance. We want to know how improvements in researcher productivity P translate into improvements in training efficiency E.
There is no reason to expect these to be proportional. Doubling the amount of useful work researchers can accomplish need not double the rate at which the training pipeline improves. Additional experiments may produce diminishing returns, and progress may be constrained by experimental compute, the availability of useful research ideas, or other bottlenecks.
We can initially represent this relationship with a simple function:
Here, g is the current rate of improvement in training efficiency without AI assistance, incorporating the contributions of existing research resources and the other factors identified by METR. Q(P) represents how improvements in researcher productivity translate into additional research output, measured by improvements in training efficiency.
We normalize the function so that Q(1)=1: without AI assistance, research proceeds at the baseline rate of about 17.7 ECI points per year after the 2025 transition.
Anthropic’s researchers have reported productivity improvements reaching approximately 12× with the latest models, which we will adopt while recognizing that the uncertainty is substantial.
Our earlier calculation suggests that AI assistance must currently be increasing the rate of improvement in log training efficiency by approximately 2.23× relative to the no-AI counterfactual. Thus, using these observations as an approximate calibration, we have:
This is an important distinction. A 12× improvement in individual researcher productivity produces a much smaller improvement in the rate at which the training pipeline improves.
The simplest way to model this is to assume diminishing returns to additional researcher productivity:
Using our two calibration points:
The resulting equation is:
That is an estimate of the diminishing returns of increased researcher productivity on research efficiency over the year for which we have data.
Projecting Ahead
Since we have AECI values for the research models used by Anthropic over the course of the year, we can estimate the relationship between AECI and researcher uplift. This allows us to create a full recursive model of AI development that incorporates the feedback of model improvement on research efficiency. But there are two remaining uncertainties that need to be addressed. One is the question of research recalcitrance: how hard is it to increase research efficiency? The other is how the impact of ECI on research efficiency might change over time: we can’t assume that just because increasing ECI from 148 to 178 uplifted researcher productivity by 12x, that when ECI is 208 productivity will be 144x.
Let’s consider research recalcitrance first. We can make a reasonable assumption from the historical pattern of progress. Research efficiency has been improving geometrically, at approximately 3× annually. Meanwhile, the resources devoted to AI research have almost certainly also been increasing geometrically. Since the rate of improvement in log training efficiency has remained approximately constant despite increasing research inputs, this suggests that the difficulty of achieving further improvements has also been increasing geometrically.
We can represent this quite simply. Let X represent the logarithm of training efficiency:
Suppose effective research inputs, including researchers and the other resources available to them, have historically been growing exponentially:
Here, n is the growth rate of effective research inputs. If research output is proportional to effective inputs at a given level of research difficulty, maintaining a constant rate of improvement in X requires research recalcitrance to grow at the same rate.
We can therefore represent research recalcitrance as:
Here, X₀ is the current level of log training efficiency, and ρ measures research recalcitrance. The further training efficiency advances, the more difficult it becomes to achieve another improvement of the same proportional size.
During the historical period, X increased at approximately ln(3) per year, giving us:
This is useful, although it doesn’t give us an independent numerical estimate of recalcitrance. We would need to know how quickly effective research inputs were growing to calculate ρ. Notice that these are research inputs, not the compute used to train frontier models; we shouldn’t assume that the two have been growing at the same rate.
We can now incorporate research recalcitrance into our previous equation for research output:
Here, g₀ is the current rate of improvement in log training efficiency without AI assistance, which we estimated at approximately ln(4.9). This includes the additional non-AI contributions to research progress identified by METR. We treat these as a fixed improvement in the effectiveness of research, rather than trying to estimate how their contribution changed over time.
The equation says that the rate of improvement in training efficiency depends on the growth in research inputs, the productivity of researchers, and increasing research recalcitrance.
We also already know how training efficiency and training compute produce model capability:
Here, t is measured in years from October 2026, and κ is our previously estimated training recalcitrance of approximately 0.181.
These two equations give us most of what we need. We only have to specify how researcher productivity responds to increasing model capability.
How does researcher productivity respond to increasing ECI?
Anthropic’s published productivity surveys, together with the more recent estimates reported by METR, provide some evidence for this relationship.
As we saw earlier, a rough fit to the available observations suggests that reported researcher productivity uplift has been increasing approximately exponentially with ECI, roughly doubling every 8–11 additional points. The observations are sparse and the surveys differ substantially in methodology, so this is very tentative. But it gives us a starting point.
The simplest assumption would be that researcher productivity continues increasing exponentially with ECI:
Here, I₀ is current model capability, the initial uplift is 12×, and β measures the responsiveness of researcher productivity to further capability improvements. Our provisional estimate is β ≈ 0.063.
This produces a recursive feedback loop. More capable models increase researcher productivity, more productive researchers improve the training pipeline, and improvements in the training pipeline produce more capable models.
But extrapolating the observed exponential relationship indefinitely is probably unrealistic.
Consider what happens as models become capable of performing particular research tasks. Initially, they may contribute almost nothing. As their capabilities improve, they become increasingly useful, eventually approaching or exceeding human performance. But once they can perform a task extremely well, further improvements may provide progressively smaller gains in productivity.
For example, an AI that can implement research experiments reliably may provide an enormous productivity improvement over one that cannot. But once most experimental code can be produced quickly and reliably, further improvements in coding capabilities will not necessarily make research much faster. The bottlenecks will increasingly lie elsewhere. Some comments by researchers quoted in the METR report suggest that this may be accurate.
If so, this suggests that the relationship between model capability and research productivity may be sigmoidal: initially slow, then rapidly increasing, and eventually approaching a ceiling.
We can represent this with a logistic function:
Here, is the maximum productivity multiplier, is the capability level at the midpoint of the transition, and s controls how rapidly the transition occurs.
This is a natural model for the automation of a particular set of research activities. As models become more capable, they progressively automate those activities, increasing researcher productivity. But the productivity gains eventually approach a limit as the relevant activities become largely automated.
We can calibrate the function to reproduce the current 12× uplift and approximately match the observed relationship between productivity and ECI at the current capability level. We can then vary the maximum productivity multiplier to see what different assumptions imply.
For example, we might examine scenarios where researcher productivity eventually approaches 24×, 48×, or 100× its unaided level. We don’t have enough evidence to really support which of these is most plausible, but in the long run the effects are similar.
What does this imply for future acceleration?
We can now combine the equations and calculate how capability growth changes as ECI increases.
There are three fundamentally different possibilities.
First, researcher productivity could continue increasing exponentially with ECI. In this case, the recursive feedback continually strengthens. Whether this produces accelerating capability growth depends on whether the productivity feedback is strong enough to overcome increasing research recalcitrance, which we have no way currently of estimating.
If recalcitrance increases more rapidly than the effective research contribution from AI, capability growth will eventually settle toward a linear trajectory, albeit potentially faster than the historical one. If productivity feedback is sufficiently strong relative to recalcitrance, the model can instead produce continuing acceleration.
As researcher productivity grows, this scenario becomes effectively indistinguishable from recursive self-improvement (RSI), and we can use the model I developed in my earlier paper. This relates model capability to the effective research effort available to improve training efficiency, and then to the capability of the next generation of models:
Here, is the effective training input required to achieve a given capability, S is the cumulative research effort required to achieve a given training efficiency, H is accumulated research effort, N is research resources, P(I) is researcher productivity as a function of model capability, and Q converts that productivity into effective research output. C is training compute.
As we have seen, training input requirements grow exponentially with ECI. The historical relationship between growing research inputs and improvements in training efficiency also suggests increasing research recalcitrance, with cumulative research-effort requirements growing as a power law of training efficiency. Since training efficiency must itself improve exponentially with ECI to sustain linear capability growth at fixed compute, research-effort requirements also grow exponentially with ECI. Consequently, RSI can sustain acceleration only if the effective research contribution from more capable models grows sufficiently rapidly to overcome both sources of recalcitrance. Exponential growth in researcher productivity alone does not guarantee this.
Second, researcher productivity could increase sigmoidally, eventually approaching a ceiling. In this case, we should expect a period of acceleration while models are becoming increasingly useful for research. But as the productivity uplift approaches its ceiling, the acceleration diminishes.
This is true even if the maximum productivity multiplier is very large. A 100× improvement in researcher productivity would be enormously valuable, but once that improvement had been achieved, it would no longer be a continuing source of acceleration.
Meanwhile, research recalcitrance would continue increasing as the training pipeline improves. Assuming the underlying growth in research inputs continues at its historical rate, capability growth will eventually return toward its underlying linear trend, just at a higher level.
Third, AI might progressively overcome new bottlenecks in the research process. In this case, the productivity relationship might be better represented by a sequence of sigmoidal transitions.
Initially, AI automates coding and experiment implementation. Subsequently, it might become competent at designing experiments, interpreting their results, developing new training algorithms, or conducting increasingly autonomous research.
Each transition could produce another substantial increase in research productivity.
This would allow successive periods of acceleration without requiring researcher productivity to increase exponentially with ECI indefinitely. But it would also mean that the future trajectory depends on when new research activities become automatable and how large their contribution to research productivity can be.
On this account, AI-driven research acceleration is likely to involve a succession of productivity improvements, each subject to diminishing returns. Some may overlap, producing periods of rapid acceleration. Others may be separated by periods in which capability growth slows as research recalcitrance increases.
Summing Up
The main takeaway for me is this: the particular magnitude of the acceleration we see in the METR report is not especially important in the long run. In the short run, it matters just in that we can expect to see RSI sooner. We are, I believe, on the edge of a period of acceleration as human researchers are first uplifted and then replaced by superintelligence. But what matters in the longer run is the form that RSI takes. How RSI proceeds depends on key parameters that it is extremely important to estimate. That is why the extensive redactions in the METR report are so frustrating: good data here is absolutely essential.
We need to be able to understand better the recalcitrance of the research pipeline, which itself is composed of many distinct aspects. We need to know much more about how much training compute went into producing the models throughout their life cycle. There are interacting positive feedback cycles between available compute, learning by doing, improved data, and many other factors. But all of these aspects of training are subject to increasing recalcitrance and diminishing returns in their own way.
Overall I am optimistic, because nothing in this discussion suggests finite-time blowup, which is the only scenario I find particularly worrisome. If we see continued increases in acceleration, or accelerating increases in acceleration, I will become much more worried. My own view is that there are diminishing returns to intelligence in every area of inquiry, and eventually the ability to perform a task saturates. The smartest computer in the world cannot beat me at tic-tac-toe. The possibility space of LLM’s is hyper-astronomical, and good ones are sparse. This suggests that the recalcitrance of LLM improvement is difficult to overcome.
But my opinions don’t really matter much here. We need data. Either the frontier labs can give it to us, or the open LLM community can empirically investigate recalcitrance and how the ability to do research improves as a function of ECI. This will give us much better guidance about the shape of things to come.
Modelling based on the METR report.
By Eugene Earnshaw · November 11, 2026 · Originally published on Substack
AI uplift in action
I’ve seen a lot of discussion about the METR acceleration report. I think many of the models used to discuss it predict acceleration that is too fast. Here is a simple-ish model that is tied as closely as I can make it to relevant empirical data to predict how things will unfold given what we know of AI acceleration.
Suppose that Anthropic uses some amount of compute to train models using their training pipeline. The amount of compute and the efficiency of the pipeline are both growing rapidly over time. The difficulty of improving models also increases rapidly as a function of model capacity, leading to the roughly linear growth in model capability we have historically observed. The rate of improvement in the pipeline is a function of average researcher productivity. Historical estimates suggest that training compute used to train models at the frontier has been growing 5x per annum and training efficiency has been growing at 3x. Given the historical growth rate of model capabilities (measured via ECI or AECI) is about 15 points a year, it suggests that multiplying the effective training effort by 15X results in a new model with ECI 15 points higher than the old one, or about 20% more effective training input needed for a model with 1 more ECI.
Then in october 2025 we have a shift with 2 elements. Researcher productivity begins to be uplifted by AI assistance, and other unspecified factors also increase the rate of research efficiency increase. The net effect is that over the following year, capabilities improve almost twice as fast.
Modelling training efficiency and capability
We can express our first assumption with a simple equation:
Here, C is training compute, E is training efficiency, I is model capability measured in AECI, and K is a constant. The parameter κ represents the increasing difficulty of training more capable models.
Rearranging:
This gives us an important result. If training compute and training efficiency grow exponentially, model capability grows linearly. In particular, 5× annual compute growth and 3× annual training-efficiency growth give us 15× annual growth in effective training inputs.
We can therefore estimate κ from the historical rate of capability improvement:
Now we can ask what must have changed to produce the acceleration observed by METR.
If we assume training compute continues growing at the same rate, the additional capability improvements must come from faster improvements in training efficiency.
With 1.9× acceleration, the annualized rate of capability improvement rises from approximately 15 to 28.5 AECI points. Our equation implies that achieving this requires training efficiency to improve at an annualized rate equivalent to approximately 34×, rather than the historical 3×.
But not all of that improvement is attributable to AI. Taking METR’s tentative 80% estimate as a share of the additional acceleration, approximately 10.8 of the additional 13.5 AECI points per year are attributable to AI, while 2.7 are attributable to other factors.
Without AI assistance, the current rate of progress would therefore be approximately 17.7 AECI points per year. This corresponds to an annualized training-efficiency improvement of approximately 4.9×.
This gives us the quantities we need to model the contribution of researcher productivity.
Modelling the productivity of research
Let P represent the productivity of Anthropic’s researchers relative to working without AI assistance. We want to know how improvements in researcher productivity P translate into improvements in training efficiency E.
There is no reason to expect these to be proportional. Doubling the amount of useful work researchers can accomplish need not double the rate at which the training pipeline improves. Additional experiments may produce diminishing returns, and progress may be constrained by experimental compute, the availability of useful research ideas, or other bottlenecks.
We can initially represent this relationship with a simple function:
Here, g is the current rate of improvement in training efficiency without AI assistance, incorporating the contributions of existing research resources and the other factors identified by METR. Q(P) represents how improvements in researcher productivity translate into additional research output, measured by improvements in training efficiency.
We normalize the function so that Q(1)=1: without AI assistance, research proceeds at the baseline rate of about 17.7 ECI points per year after the 2025 transition.
Anthropic’s researchers have reported productivity improvements reaching approximately 12× with the latest models, which we will adopt while recognizing that the uncertainty is substantial.
Our earlier calculation suggests that AI assistance must currently be increasing the rate of improvement in log training efficiency by approximately 2.23× relative to the no-AI counterfactual. Thus, using these observations as an approximate calibration, we have:
This is an important distinction. A 12× improvement in individual researcher productivity produces a much smaller improvement in the rate at which the training pipeline improves.
The simplest way to model this is to assume diminishing returns to additional researcher productivity:
Using our two calibration points:
The resulting equation is:
That is an estimate of the diminishing returns of increased researcher productivity on research efficiency over the year for which we have data.
Projecting Ahead
Since we have AECI values for the research models used by Anthropic over the course of the year, we can estimate the relationship between AECI and researcher uplift. This allows us to create a full recursive model of AI development that incorporates the feedback of model improvement on research efficiency. But there are two remaining uncertainties that need to be addressed. One is the question of research recalcitrance: how hard is it to increase research efficiency? The other is how the impact of ECI on research efficiency might change over time: we can’t assume that just because increasing ECI from 148 to 178 uplifted researcher productivity by 12x, that when ECI is 208 productivity will be 144x.
Let’s consider research recalcitrance first. We can make a reasonable assumption from the historical pattern of progress. Research efficiency has been improving geometrically, at approximately 3× annually. Meanwhile, the resources devoted to AI research have almost certainly also been increasing geometrically. Since the rate of improvement in log training efficiency has remained approximately constant despite increasing research inputs, this suggests that the difficulty of achieving further improvements has also been increasing geometrically.
We can represent this quite simply. Let X represent the logarithm of training efficiency:
Suppose effective research inputs, including researchers and the other resources available to them, have historically been growing exponentially:
Here, n is the growth rate of effective research inputs. If research output is proportional to effective inputs at a given level of research difficulty, maintaining a constant rate of improvement in X requires research recalcitrance to grow at the same rate.
We can therefore represent research recalcitrance as:
Here, X₀ is the current level of log training efficiency, and ρ measures research recalcitrance. The further training efficiency advances, the more difficult it becomes to achieve another improvement of the same proportional size.
During the historical period, X increased at approximately ln(3) per year, giving us:
This is useful, although it doesn’t give us an independent numerical estimate of recalcitrance. We would need to know how quickly effective research inputs were growing to calculate ρ. Notice that these are research inputs, not the compute used to train frontier models; we shouldn’t assume that the two have been growing at the same rate.
We can now incorporate research recalcitrance into our previous equation for research output:
Here, g₀ is the current rate of improvement in log training efficiency without AI assistance, which we estimated at approximately ln(4.9). This includes the additional non-AI contributions to research progress identified by METR. We treat these as a fixed improvement in the effectiveness of research, rather than trying to estimate how their contribution changed over time.
The equation says that the rate of improvement in training efficiency depends on the growth in research inputs, the productivity of researchers, and increasing research recalcitrance.
We also already know how training efficiency and training compute produce model capability:
Here, t is measured in years from October 2026, and κ is our previously estimated training recalcitrance of approximately 0.181.
These two equations give us most of what we need. We only have to specify how researcher productivity responds to increasing model capability.
How does researcher productivity respond to increasing ECI?
Anthropic’s published productivity surveys, together with the more recent estimates reported by METR, provide some evidence for this relationship.
As we saw earlier, a rough fit to the available observations suggests that reported researcher productivity uplift has been increasing approximately exponentially with ECI, roughly doubling every 8–11 additional points. The observations are sparse and the surveys differ substantially in methodology, so this is very tentative. But it gives us a starting point.
The simplest assumption would be that researcher productivity continues increasing exponentially with ECI:
Here, I₀ is current model capability, the initial uplift is 12×, and β measures the responsiveness of researcher productivity to further capability improvements. Our provisional estimate is β ≈ 0.063.
This produces a recursive feedback loop. More capable models increase researcher productivity, more productive researchers improve the training pipeline, and improvements in the training pipeline produce more capable models.
But extrapolating the observed exponential relationship indefinitely is probably unrealistic.
Consider what happens as models become capable of performing particular research tasks. Initially, they may contribute almost nothing. As their capabilities improve, they become increasingly useful, eventually approaching or exceeding human performance. But once they can perform a task extremely well, further improvements may provide progressively smaller gains in productivity.
For example, an AI that can implement research experiments reliably may provide an enormous productivity improvement over one that cannot. But once most experimental code can be produced quickly and reliably, further improvements in coding capabilities will not necessarily make research much faster. The bottlenecks will increasingly lie elsewhere. Some comments by researchers quoted in the METR report suggest that this may be accurate.
If so, this suggests that the relationship between model capability and research productivity may be sigmoidal: initially slow, then rapidly increasing, and eventually approaching a ceiling.
We can represent this with a logistic function:
Here, is the maximum productivity multiplier, is the capability level at the midpoint of the transition, and s controls how rapidly the transition occurs.
This is a natural model for the automation of a particular set of research activities. As models become more capable, they progressively automate those activities, increasing researcher productivity. But the productivity gains eventually approach a limit as the relevant activities become largely automated.
We can calibrate the function to reproduce the current 12× uplift and approximately match the observed relationship between productivity and ECI at the current capability level. We can then vary the maximum productivity multiplier to see what different assumptions imply.
For example, we might examine scenarios where researcher productivity eventually approaches 24×, 48×, or 100× its unaided level. We don’t have enough evidence to really support which of these is most plausible, but in the long run the effects are similar.
What does this imply for future acceleration?
We can now combine the equations and calculate how capability growth changes as ECI increases.
There are three fundamentally different possibilities.
First, researcher productivity could continue increasing exponentially with ECI. In this case, the recursive feedback continually strengthens. Whether this produces accelerating capability growth depends on whether the productivity feedback is strong enough to overcome increasing research recalcitrance, which we have no way currently of estimating.
If recalcitrance increases more rapidly than the effective research contribution from AI, capability growth will eventually settle toward a linear trajectory, albeit potentially faster than the historical one. If productivity feedback is sufficiently strong relative to recalcitrance, the model can instead produce continuing acceleration.
As researcher productivity grows, this scenario becomes effectively indistinguishable from recursive self-improvement (RSI), and we can use the model I developed in my earlier paper. This relates model capability to the effective research effort available to improve training efficiency, and then to the capability of the next generation of models:
Here, is the effective training input required to achieve a given capability, S is the cumulative research effort required to achieve a given training efficiency, H is accumulated research effort, N is research resources, P(I) is researcher productivity as a function of model capability, and Q converts that productivity into effective research output. C is training compute.
As we have seen, training input requirements grow exponentially with ECI. The historical relationship between growing research inputs and improvements in training efficiency also suggests increasing research recalcitrance, with cumulative research-effort requirements growing as a power law of training efficiency. Since training efficiency must itself improve exponentially with ECI to sustain linear capability growth at fixed compute, research-effort requirements also grow exponentially with ECI. Consequently, RSI can sustain acceleration only if the effective research contribution from more capable models grows sufficiently rapidly to overcome both sources of recalcitrance. Exponential growth in researcher productivity alone does not guarantee this.
Second, researcher productivity could increase sigmoidally, eventually approaching a ceiling. In this case, we should expect a period of acceleration while models are becoming increasingly useful for research. But as the productivity uplift approaches its ceiling, the acceleration diminishes.
This is true even if the maximum productivity multiplier is very large. A 100× improvement in researcher productivity would be enormously valuable, but once that improvement had been achieved, it would no longer be a continuing source of acceleration.
Meanwhile, research recalcitrance would continue increasing as the training pipeline improves. Assuming the underlying growth in research inputs continues at its historical rate, capability growth will eventually return toward its underlying linear trend, just at a higher level.
Third, AI might progressively overcome new bottlenecks in the research process. In this case, the productivity relationship might be better represented by a sequence of sigmoidal transitions.
Initially, AI automates coding and experiment implementation. Subsequently, it might become competent at designing experiments, interpreting their results, developing new training algorithms, or conducting increasingly autonomous research.
Each transition could produce another substantial increase in research productivity.
This would allow successive periods of acceleration without requiring researcher productivity to increase exponentially with ECI indefinitely. But it would also mean that the future trajectory depends on when new research activities become automatable and how large their contribution to research productivity can be.
On this account, AI-driven research acceleration is likely to involve a succession of productivity improvements, each subject to diminishing returns. Some may overlap, producing periods of rapid acceleration. Others may be separated by periods in which capability growth slows as research recalcitrance increases.
Summing Up
The main takeaway for me is this: the particular magnitude of the acceleration we see in the METR report is not especially important in the long run. In the short run, it matters just in that we can expect to see RSI sooner. We are, I believe, on the edge of a period of acceleration as human researchers are first uplifted and then replaced by superintelligence. But what matters in the longer run is the form that RSI takes. How RSI proceeds depends on key parameters that it is extremely important to estimate. That is why the extensive redactions in the METR report are so frustrating: good data here is absolutely essential.
We need to be able to understand better the recalcitrance of the research pipeline, which itself is composed of many distinct aspects. We need to know much more about how much training compute went into producing the models throughout their life cycle. There are interacting positive feedback cycles between available compute, learning by doing, improved data, and many other factors. But all of these aspects of training are subject to increasing recalcitrance and diminishing returns in their own way.
Overall I am optimistic, because nothing in this discussion suggests finite-time blowup, which is the only scenario I find particularly worrisome. If we see continued increases in acceleration, or accelerating increases in acceleration, I will become much more worried. My own view is that there are diminishing returns to intelligence in every area of inquiry, and eventually the ability to perform a task saturates. The smartest computer in the world cannot beat me at tic-tac-toe. The possibility space of LLM’s is hyper-astronomical, and good ones are sparse. This suggests that the recalcitrance of LLM improvement is difficult to overcome.
But my opinions don’t really matter much here. We need data. Either the frontier labs can give it to us, or the open LLM community can empirically investigate recalcitrance and how the ability to do research improves as a function of ECI. This will give us much better guidance about the shape of things to come.